{"meta":{"query_hash":"ae1062cd4a31","filters":{"topic":"Software Reliability and Analysis Research"},"cohort_total":333,"direct_labels_cover":0,"predictions_cover":333,"exported":333,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/ae1062cd4a31","api":"https://metacan.xera.ac/api/v1/cohort?topic=Software+Reliability+and+Analysis+Research"},"results":[{"id":"W145001103","doi":"","title":"IEEE Humanitarian Projects: Open Hardware for the Benefit of the Poorest Nations","year":2010,"lang":"en","type":"article","venue":"˜The œopen source business resource","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Humanitarian aid; Computer science; Computer security; Business; Economic growth; Economics","score_opus":0.04076732998213547,"score_gpt":0.3002581313308094,"score_spread":0.25949080134867397,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W145001103","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.19895442,0.0005899894,0.63816655,0.11175374,0.0020041822,0.022567421,0.00023348567,0.0005036425,0.025226591],"genre_scores_gemma":[0.98265207,0.000010170564,0.0021923543,0.0010726437,0.000323657,0.0007631063,0.000012479776,0.000050013994,0.012923516],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9974501,0.000211038,0.00046842848,0.00062847947,0.00076052337,0.00048143318],"domain_scores_gemma":[0.99350846,0.0020483192,0.00032596855,0.0030679964,0.0009683056,0.0000809502],"candidate_categories":["sts","scholarly_communication","open_science"],"consensus_categories":[],"category_scores_codex":[0.0020420607,0.00024376388,0.00034467957,0.0001335665,0.0026060275,0.0021589957,0.017528033,0.00012815569,0.000057560825],"category_scores_gemma":[0.00094301277,0.00012176576,0.0002221599,0.0026629216,0.0005300976,0.0005330875,0.004676809,0.00046291473,0.000025424553],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0006728768,0.0032313585,0.03602843,0.00085809844,0.0017600517,0.000013204903,0.03790665,0.06955251,0.0071115983,0.27509263,0.08075842,0.48701417],"study_design_scores_gemma":[0.00079254794,0.000041925283,0.027447058,0.00008944799,0.00008234413,0.000020208246,0.0008041969,0.035212606,0.00048700906,0.0030196053,0.93168056,0.0003224693],"about_ca_topic_score_codex":0.0040545193,"about_ca_topic_score_gemma":0.008710031,"teacher_disagreement_score":0.85092217,"about_ca_system_score_codex":0.000044950648,"about_ca_system_score_gemma":0.0003847856,"threshold_uncertainty_score":0.99887687},"labels":[],"label_agreement":null},{"id":"W1488504497","doi":"10.1007/978-3-642-13556-9_2","title":"Architectural Design Decisions for Achieving Reliable Software Systems","year":2010,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Computer science; Architectural pattern; Software system; Software construction; Software engineering; Software architecture; Software development; Software sizing; Reliability engineering; Software quality; Resource-oriented architecture; Verification and validation; Software; Engineering; Operating system","score_opus":0.04001206392220912,"score_gpt":0.2873187684320934,"score_spread":0.2473067045098843,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1488504497","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006128901,0.0007893209,0.99501586,0.0007097973,0.001803085,0.00118491,0.000012524824,0.00032435593,0.000098847704],"genre_scores_gemma":[0.0121515235,0.000054763845,0.98608464,0.00040852142,0.0005686021,0.00007778868,0.000009285454,0.00005365244,0.00059122726],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99398327,0.00008504208,0.00081053056,0.0022494493,0.0016741483,0.0011975822],"domain_scores_gemma":[0.98878473,0.0072867866,0.00031128532,0.002508646,0.00073814555,0.00037040943],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science"],"consensus_categories":[],"category_scores_codex":[0.0033643052,0.00064464804,0.0008803011,0.0013578171,0.00086786225,0.0015751082,0.006011626,0.00061308045,0.000013951179],"category_scores_gemma":[0.0025890938,0.00053820247,0.0003741461,0.0010223354,0.00085892563,0.00058622693,0.0017760433,0.0018200693,0.000047494825],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016956981,0.000037335725,0.00005655036,0.000086691245,0.000027496908,0.00003546459,0.0003273666,0.25085735,0.0002946323,0.0039771027,0.000045821624,0.74423724],"study_design_scores_gemma":[0.0003297709,0.0002985767,0.000048841455,0.00077089,0.000020050564,0.00012352904,1.6381979e-7,0.8190363,0.0008764659,0.17261256,0.004994324,0.0008885312],"about_ca_topic_score_codex":0.00007342652,"about_ca_topic_score_gemma":0.000060702365,"teacher_disagreement_score":0.7433487,"about_ca_system_score_codex":0.0003298608,"about_ca_system_score_gemma":0.0012140477,"threshold_uncertainty_score":0.9997069},"labels":[],"label_agreement":null},{"id":"W1488742231","doi":"10.1109/issre.2004.8","title":"An Extended Operational Profile Model","year":2005,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":32,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; University of Waterloo","funders":"","keywords":"Computer science; Software quality; Verification and validation; Software; Software sizing; Software development; Reliability engineering; Software reliability testing; Software metric; Software construction; Software engineering; Systems engineering; Engineering; Programming language","score_opus":0.023675057857663117,"score_gpt":0.3208768887338602,"score_spread":0.2972018308761971,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1488742231","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009373423,0.000014983929,0.9805477,0.0026235646,0.000009217789,0.00005952679,9.533563e-7,0.00016564889,0.007204934],"genre_scores_gemma":[0.63592213,0.0000028947604,0.36124057,0.00043819757,0.000041666095,0.000011972725,0.0000036360996,0.0000018365271,0.0023370776],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99911505,0.00003746207,0.000108697364,0.0002623093,0.0003191416,0.00015731933],"domain_scores_gemma":[0.9993236,0.000029770732,0.000009026154,0.00042868475,0.00011229946,0.00009662882],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002753569,0.00005116496,0.00006250286,0.000056001834,0.000100164834,0.00017648836,0.0005919537,0.000028226745,0.00035575128],"category_scores_gemma":[0.00002897122,0.00003988172,0.000037696624,0.00020552102,0.00002235431,0.00086896296,0.00008302186,0.00006860727,0.0002843271],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000045041925,0.0006320516,0.00055799005,0.0000054285465,0.00001767656,0.000002034015,0.00038987407,0.1528516,0.0038926713,0.4984966,0.004896523,0.33825308],"study_design_scores_gemma":[0.0000673745,0.000020468218,0.00026434433,6.4799434e-7,6.4070395e-7,0.0000013803348,0.0000041627386,0.99304116,0.0025729404,0.0028882036,0.0010781988,0.000060460035],"about_ca_topic_score_codex":0.000011341099,"about_ca_topic_score_gemma":0.000009925674,"teacher_disagreement_score":0.8401896,"about_ca_system_score_codex":0.000027008124,"about_ca_system_score_gemma":0.00010856748,"threshold_uncertainty_score":0.3895227},"labels":[],"label_agreement":null},{"id":"W1497968089","doi":"10.1023/a:1024424811345","title":"Fault Prediction Modeling for Software Quality Estimation: Comparing Commonly Used Techniques","year":2003,"lang":"en","type":"article","venue":"Empirical Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":128,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"McGill University","keywords":"Computer science; Software quality; Software; Artificial neural network; Approximation error; Reliability engineering; Data mining; Artificial intelligence; Machine learning; Algorithm; Software development; Engineering","score_opus":0.07721747522635733,"score_gpt":0.35017777725256943,"score_spread":0.27296030202621213,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1497968089","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.053146422,0.00015347908,0.94354683,0.0001609792,0.00017580987,0.00040069985,0.000010654374,0.0023896769,0.000015474858],"genre_scores_gemma":[0.4940997,0.0000050260546,0.50558245,0.000055510933,0.00005706008,0.00013402355,0.000022531862,0.000024001802,0.000019716093],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.997311,0.00013075804,0.00066744105,0.0006625305,0.000638091,0.00059017184],"domain_scores_gemma":[0.99729365,0.0013593758,0.0000840328,0.00073612266,0.00028215235,0.00024469342],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0014008299,0.00029060853,0.0004805184,0.00022537248,0.00031160493,0.00027641992,0.0006873557,0.00020127918,0.000008864174],"category_scores_gemma":[0.005380118,0.00029285802,0.00026569795,0.00082408136,0.000032555752,0.00066360127,0.0001687635,0.00041583023,0.000011261964],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010647868,0.00016720335,0.094680086,0.00030849248,0.00008055412,0.0000038885787,0.0003951948,0.8871707,0.00007651346,0.0023423545,0.00036640986,0.014397975],"study_design_scores_gemma":[0.00030616747,0.00006204323,0.0018328445,0.000075264645,0.000014416909,0.000008300578,0.000010899327,0.99122363,0.0011579024,0.0022703018,0.002696077,0.00034214844],"about_ca_topic_score_codex":0.000029579869,"about_ca_topic_score_gemma":0.000004850769,"teacher_disagreement_score":0.44095328,"about_ca_system_score_codex":0.0002562927,"about_ca_system_score_gemma":0.000112179325,"threshold_uncertainty_score":0.9999524},"labels":[],"label_agreement":null},{"id":"W150194969","doi":"","title":"Software Accessibility: Where Are We Today?","year":2008,"lang":"en","type":"article","venue":"˜The œopen source business resource","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Software; Operating system","score_opus":0.04533278082877135,"score_gpt":0.2936830832399386,"score_spread":0.24835030241116723,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W150194969","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07495241,0.0041242554,0.88394976,0.030399926,0.00023012393,0.0014976235,0.00001582303,0.0013025742,0.0035275358],"genre_scores_gemma":[0.94516146,0.0011026633,0.010819917,0.0032214485,0.0006643167,0.0003263056,0.00002581711,0.00016075373,0.0385173],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99480534,0.00066149794,0.0006801207,0.0014423672,0.0014368014,0.0009739035],"domain_scores_gemma":[0.99393374,0.0009807316,0.00037254355,0.0036556276,0.0007320052,0.00032534398],"candidate_categories":["metaepi_narrow","sts","scholarly_communication","open_science"],"consensus_categories":[],"category_scores_codex":[0.0014532178,0.00047240063,0.0007137808,0.00023108984,0.00185038,0.0015006699,0.010082935,0.00023982958,0.00025605343],"category_scores_gemma":[0.00044628503,0.00033427394,0.00028279904,0.003958545,0.0006249165,0.0012109266,0.005462256,0.0005539484,0.00040805145],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032193528,0.0016184434,0.21849936,0.000660668,0.00043156798,0.00060084584,0.0086713135,0.014261297,0.0001982176,0.0007252999,0.16080791,0.5932031],"study_design_scores_gemma":[0.00074745214,0.000036645724,0.04956362,0.00019301687,0.000027900274,0.00018529719,0.0004582327,0.009953664,0.000077078716,0.0011441268,0.9369554,0.0006575735],"about_ca_topic_score_codex":0.00075766246,"about_ca_topic_score_gemma":0.00027002027,"teacher_disagreement_score":0.8731298,"about_ca_system_score_codex":0.000161094,"about_ca_system_score_gemma":0.00038141882,"threshold_uncertainty_score":0.99991095},"labels":[],"label_agreement":null},{"id":"W1504754385","doi":"10.1007/978-3-642-13824-9_8","title":"Formal Lifetime Reliability Analysis Using Continuous Random Variables","year":2010,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Computer science; Reliability (semiconductor); Random variable; Reliability engineering; Reliability theory; Failure rate; Process (computing); Hazard; Mathematics; Statistics","score_opus":0.013473977106000052,"score_gpt":0.2652839295469684,"score_spread":0.2518099524409683,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1504754385","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012633435,0.00025911073,0.99574095,0.00034708093,0.0007862937,0.0004885012,0.000012069231,0.00022376697,0.0008788964],"genre_scores_gemma":[0.19572042,0.000057125326,0.8028347,0.00048760496,0.00051481894,0.000012509036,0.000015020713,0.000038649672,0.00031917854],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9928208,0.0001582366,0.0010733935,0.0025318316,0.0021074829,0.001308267],"domain_scores_gemma":[0.9928131,0.0022081449,0.0004487512,0.0032456515,0.00088533823,0.00039902993],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0055103726,0.00070223917,0.0015369299,0.002256484,0.00068782165,0.0012240345,0.005264641,0.00077206607,0.00014414165],"category_scores_gemma":[0.0010470703,0.00060550886,0.0007717761,0.003399443,0.0016035367,0.0012733905,0.0026769973,0.0020780417,0.00003986871],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007276921,0.0002289969,0.0029574954,0.00014367848,0.00050885166,0.00017036489,0.0010773167,0.50279504,0.0008771657,0.007081044,0.0000075155804,0.48407978],"study_design_scores_gemma":[0.00048820375,0.000088453686,0.00020712799,0.000096013915,0.00017738507,0.000028837834,1.6372759e-7,0.922132,0.0009353956,0.074408524,0.0007205317,0.00071737665],"about_ca_topic_score_codex":0.00025226758,"about_ca_topic_score_gemma":0.00013509628,"teacher_disagreement_score":0.48336238,"about_ca_system_score_codex":0.0003990552,"about_ca_system_score_gemma":0.0011952822,"threshold_uncertainty_score":0.9998128},"labels":[],"label_agreement":null},{"id":"W1509712248","doi":"","title":"Deriving real-time monitors from system requirements documentation","year":2000,"lang":"en","type":"dissertation","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Correctness; System requirements specification; Documentation; Computer science; Software requirements specification; Process (computing); System requirements; Automaton; Equivalence (formal languages); Software engineering; Reliability engineering; Programming language; Engineering; Software; Operating system; Theoretical computer science; Software development; Software design","score_opus":0.013683638999239276,"score_gpt":0.30602625905638753,"score_spread":0.29234262005714823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1509712248","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.88802516,0.00030505355,0.02396925,0.00009415212,0.0012306499,0.0008134761,0.00002010995,0.0016399248,0.083902225],"genre_scores_gemma":[0.88201296,0.0010167084,0.035253722,0.000042357125,0.00056398375,0.00021937206,0.0035789004,0.0000910496,0.07722093],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99704653,0.000208722,0.00054191455,0.00080113957,0.0010535087,0.00034816997],"domain_scores_gemma":[0.9984533,0.00022607185,0.00019047847,0.00083177094,0.00016240274,0.00013596182],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00034055064,0.00026880408,0.000386457,0.00028599324,0.00022210802,0.00068369444,0.0011123224,0.0002471998,0.0008705499],"category_scores_gemma":[0.000040357307,0.00025116,0.00019614754,0.00056266866,0.000014909698,0.000806881,0.000071695125,0.00020407753,0.0014912231],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003764236,0.0007789258,0.009882587,0.0019937614,0.002409506,0.00023409796,0.011961774,0.00085680827,0.0414121,0.005369144,0.012683209,0.91204166],"study_design_scores_gemma":[0.0052348464,0.0009432188,0.19583061,0.010159226,0.0014832442,0.00002000936,0.01694943,0.5815573,0.15425394,0.01649029,0.0072883884,0.009789502],"about_ca_topic_score_codex":0.0042640315,"about_ca_topic_score_gemma":0.000108117645,"teacher_disagreement_score":0.90225214,"about_ca_system_score_codex":0.00036824495,"about_ca_system_score_gemma":0.00018492124,"threshold_uncertainty_score":0.99999404},"labels":[],"label_agreement":null},{"id":"W1527406032","doi":"10.5539/eer.v5n1p75","title":"Using Simulation to Test the Reliability of Regression Models","year":2015,"lang":"en","type":"article","venue":"Energy and Environment Research","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Division of Chemical, Bioengineering, Environmental, and Transport Systems; National Science Foundation","keywords":"Reliability (semiconductor); Computer science; Range (aeronautics); Regression analysis; Sample size determination; Linear regression; Statistics; Regression; Statistical model; Mathematics; Machine learning; Engineering","score_opus":0.16087222873893245,"score_gpt":0.376096487047994,"score_spread":0.21522425830906153,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1527406032","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2801248,0.00031615153,0.71650726,0.0022847883,0.000021113621,0.0001380664,0.0000012843664,0.00001856955,0.0005879618],"genre_scores_gemma":[0.9925084,0.00013897485,0.007053752,0.000027222495,0.000025849447,0.000011398161,9.126246e-7,0.000004175005,0.00022930076],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99771446,0.000433887,0.00018480275,0.00035272224,0.0010445465,0.00026958776],"domain_scores_gemma":[0.9979129,0.0010377804,0.000028664563,0.0007240384,0.0001030208,0.00019358737],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0031783143,0.00007474345,0.000118566655,0.00012385577,0.0002140099,0.000058559985,0.00051262893,0.00005725923,0.0000096253425],"category_scores_gemma":[0.0006141488,0.00004597283,0.000034951547,0.00045731507,0.00018595869,0.00026338684,0.0008316206,0.00014803096,0.0000058503474],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001664687,0.00013432998,0.0017589901,0.000006886958,0.0000041283038,0.0000013084099,0.00043938402,0.9757305,0.0006437095,0.0032450561,0.00010271484,0.017916381],"study_design_scores_gemma":[0.00010181051,0.00014355558,0.00054412603,0.000014302615,0.0000017447409,7.080847e-7,0.00006957879,0.96440285,0.0016880017,0.028816314,0.004160533,0.000056491222],"about_ca_topic_score_codex":0.0004609045,"about_ca_topic_score_gemma":0.000009345894,"teacher_disagreement_score":0.7123836,"about_ca_system_score_codex":0.00010710205,"about_ca_system_score_gemma":0.000061595885,"threshold_uncertainty_score":0.18747176},"labels":[],"label_agreement":null},{"id":"W1529782179","doi":"10.1007/978-3-540-45236-2_9","title":"Lessons Learned from a Successful Implementation of Formal Methods in an Industrial Project","year":2003,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":57,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Notation; Engineering management; Software engineering; Operations research; Engineering","score_opus":0.14197543767826332,"score_gpt":0.43721716113703923,"score_spread":0.2952417234587759,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1529782179","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0042192372,0.00012145019,0.99377304,0.0007673419,0.000398078,0.00049568457,0.000017045455,0.000048998158,0.00015910922],"genre_scores_gemma":[0.115650594,0.00008425414,0.88357025,0.0003334088,0.00023819268,0.000029718743,0.000031295378,0.00002871067,0.000033571687],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99537426,0.00042469136,0.00084265845,0.0015018054,0.0011478371,0.0007087644],"domain_scores_gemma":[0.9968338,0.0011026822,0.00036987118,0.0012795889,0.00028741994,0.00012666595],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0037818328,0.00037574474,0.0007050931,0.0015217679,0.00014872634,0.00042176613,0.0029678892,0.00040569413,0.00004814995],"category_scores_gemma":[0.0002731273,0.0003437534,0.00016121227,0.0016850797,0.0004870668,0.0013566666,0.0009870615,0.0010587373,0.0000030532103],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000100895695,0.000046664933,0.0020479206,0.000015689156,0.000012951254,0.00001354435,0.0011223033,0.0035678355,0.0002598361,0.0022385558,0.0000034271225,0.9906612],"study_design_scores_gemma":[0.0026591597,0.0011071694,0.002804451,0.000575366,0.000045344754,0.000017578683,0.000018917102,0.56729084,0.027528029,0.39422587,0.0020239104,0.0017033487],"about_ca_topic_score_codex":0.0015409777,"about_ca_topic_score_gemma":0.0017526681,"teacher_disagreement_score":0.9889578,"about_ca_system_score_codex":0.00032241782,"about_ca_system_score_gemma":0.001447011,"threshold_uncertainty_score":0.9999015},"labels":[],"label_agreement":null},{"id":"W1542123695","doi":"10.3968/j.pam.1925252820130602.1738","title":"Modelling Academic Risks of Students in a Polytechnic System with the Use of Discriminant Analysis","year":2013,"lang":"en","type":"article","venue":"Progress in applied mathematics","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Linear discriminant analysis; Identification (biology); Bayesian probability; Computer science; Mathematics education; Work (physics); Regression analysis; Academic achievement; Psychology; Statistics; Machine learning; Artificial intelligence; Mathematics; Engineering","score_opus":0.11895334231691275,"score_gpt":0.35725886025875336,"score_spread":0.23830551794184063,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1542123695","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.45975453,0.00012048314,0.539564,0.00004151498,0.0000024684944,0.00048213248,8.2041504e-7,0.000021891563,0.000012124881],"genre_scores_gemma":[0.86414313,0.000035397963,0.13554223,0.0000026458436,0.000002207943,0.00026308149,5.9013877e-7,0.0000069214457,0.0000037964987],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99796325,0.00005700968,0.00062605616,0.0002696404,0.0008196432,0.00026437454],"domain_scores_gemma":[0.99831724,0.00044603084,0.00031881206,0.00079167436,0.00008626383,0.000039983097],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00096799334,0.00013224255,0.00050687586,0.00044207388,0.000025817542,0.00007317885,0.0013584095,0.000084310865,0.0000016059302],"category_scores_gemma":[0.000021960537,0.000076774355,0.00008300772,0.0023560515,0.00018331727,0.00013718275,0.00040260787,0.00030035514,0.0000026261573],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000038267357,0.0018916726,0.26704836,0.0033740236,0.0008455644,0.000009387086,0.017165622,0.55827427,0.00019609902,0.120216504,0.000009425032,0.030930825],"study_design_scores_gemma":[0.00016952307,0.000020031599,0.0027329293,0.00021620293,0.00008594721,8.033808e-7,0.00068849284,0.99437976,0.0003637249,0.0012518992,0.0000014339761,0.00008927897],"about_ca_topic_score_codex":0.00020224268,"about_ca_topic_score_gemma":0.000022692106,"teacher_disagreement_score":0.4361055,"about_ca_system_score_codex":0.000057945967,"about_ca_system_score_gemma":0.00003193621,"threshold_uncertainty_score":0.31307673},"labels":[],"label_agreement":null},{"id":"W1567126008","doi":"10.1109/ias.1990.152363","title":"Managing motor maintenance","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"General Electric (Canada)","funders":"","keywords":"Spare part; Component (thermodynamics); Maintenance engineering; Computer science; Motor controller; Spare time; Reliability engineering; Engineering; Manufacturing engineering; Systems engineering; Operations management","score_opus":0.025201226518727037,"score_gpt":0.2527680208404613,"score_spread":0.22756679432173427,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1567126008","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0013002453,0.00013717254,0.9440015,0.008170923,0.00005074613,0.000048279515,1.7206983e-7,0.00025049664,0.046040438],"genre_scores_gemma":[0.91110635,0.00020969904,0.051500786,0.0008263588,0.00004096725,0.000010156568,1.6400905e-7,0.000004406639,0.036301143],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9991764,0.000028952687,0.00009364369,0.0002379391,0.00023168737,0.00023138753],"domain_scores_gemma":[0.9993222,0.000081906495,0.000012719124,0.000464975,0.00004751442,0.00007070304],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00018903067,0.00005095586,0.000072425595,0.00007406144,0.000072420386,0.00012149703,0.0006748059,0.00001851911,0.00049973],"category_scores_gemma":[0.00007635729,0.000039480292,0.000062423635,0.00040425136,0.000028254613,0.00025962116,0.00017658607,0.0000760691,0.0009458904],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000013549017,0.00013634958,0.0026291872,0.000023506418,0.000030497524,0.00005292215,0.0002864295,0.00013192768,0.00037329184,0.14950061,0.02771745,0.8191165],"study_design_scores_gemma":[0.00010077229,0.000025045267,0.00084037124,0.0000067480396,0.0000011910754,0.0000061874666,0.00001410893,0.95819235,0.0002596393,0.012143846,0.02830354,0.00010621728],"about_ca_topic_score_codex":0.000026228532,"about_ca_topic_score_gemma":0.0000025277577,"teacher_disagreement_score":0.9580604,"about_ca_system_score_codex":0.000023487717,"about_ca_system_score_gemma":0.00000410802,"threshold_uncertainty_score":0.999832},"labels":[],"label_agreement":null},{"id":"W1567356114","doi":"10.1007/978-3-642-00867-2_11","title":"DREP: A Requirements Engineering Process for Dependable Reactive Systems","year":2009,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":21,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Computer science; Interrupt; Process (computing); Context (archaeology); Requirements engineering; Reliability (semiconductor); Software engineering; Quality (philosophy); Risk analysis (engineering); Systems engineering; Engineering; Programming language; Embedded system","score_opus":0.030033471474984044,"score_gpt":0.2976352152486902,"score_spread":0.26760174377370616,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1567356114","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000120620665,0.000504981,0.9968081,0.00024895466,0.00066203164,0.00084956555,0.0000061038245,0.00020587147,0.0005937889],"genre_scores_gemma":[0.60158837,0.00008184862,0.39464375,0.00054545136,0.0012372936,0.00020165749,0.000024904026,0.00010559164,0.001571117],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99544287,0.000025173837,0.00055785856,0.0017118419,0.0014011866,0.0008610609],"domain_scores_gemma":[0.99689966,0.00073301373,0.0002387949,0.001275599,0.0006433709,0.00020953943],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015824086,0.00046358147,0.0006401848,0.0009161343,0.00025434926,0.0007888276,0.0032273578,0.00030285734,0.0000036709669],"category_scores_gemma":[0.0004187362,0.0004237226,0.00019027162,0.00084652536,0.00019712208,0.0008581238,0.0005252639,0.0006511977,0.000015178885],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000018355347,0.000090011425,0.00006058026,0.0004223795,0.00006519563,0.000088295994,0.0007272176,0.44675258,0.00040654672,0.0109858,0.000016927188,0.5403661],"study_design_scores_gemma":[0.0002394822,0.00021372529,0.000018097999,0.0005921543,0.00001201796,0.000034670527,2.5455563e-7,0.95368695,0.0018570252,0.041534703,0.0012812588,0.0005296605],"about_ca_topic_score_codex":0.000032668533,"about_ca_topic_score_gemma":0.000010989432,"teacher_disagreement_score":0.6021643,"about_ca_system_score_codex":0.0005382092,"about_ca_system_score_gemma":0.00069631863,"threshold_uncertainty_score":0.9998215},"labels":[],"label_agreement":null},{"id":"W1574130501","doi":"10.5539/mas.v9n8p213","title":"Mills Model Based Evaluation of Security of Software Systems","year":2015,"lang":"en","type":"article","venue":"Modern Applied Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Vulnerability (computing); Division (mathematics); Context (archaeology); Software security assurance; Reliability (semiconductor); Software; Process (computing); Software system; Secure coding; Computer security; Reliability engineering; Software engineering; Information security; Security service; Operating system; Engineering","score_opus":0.06696400011709529,"score_gpt":0.3407199670536613,"score_spread":0.273755966936566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1574130501","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11085388,0.00014394335,0.8876971,0.00003821623,0.000049970102,0.0002955061,0.000005522892,0.00005099339,0.00086486875],"genre_scores_gemma":[0.971571,0.000001536323,0.028354337,0.000017998585,0.000007390681,0.000035640787,0.0000013876114,0.000003993387,0.000006699514],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9950673,0.00010519644,0.00036764512,0.0005436586,0.0036095204,0.00030667748],"domain_scores_gemma":[0.99668396,0.00013818941,0.00017907654,0.0010266887,0.0017721447,0.00019995973],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008491243,0.00010822675,0.00026289647,0.0002974557,0.000096306205,0.00008517315,0.0017324601,0.00005699562,0.0000022477077],"category_scores_gemma":[0.00070961064,0.00009479795,0.000061636434,0.0015819238,0.00050957326,0.0003609654,0.00031900985,0.000104855004,0.0000068746945],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010938123,0.00017124026,0.0005235831,0.00006982396,0.000006774616,2.312117e-7,0.0018461983,0.94495267,0.028364126,0.0059744827,0.00003912407,0.018040838],"study_design_scores_gemma":[0.0003272237,0.000023777322,0.00007420751,0.000016233593,0.000010275656,3.9360017e-7,0.00005020528,0.9506083,0.014344384,0.034450468,0.000003604437,0.000090935115],"about_ca_topic_score_codex":0.00006898179,"about_ca_topic_score_gemma":0.000003884522,"teacher_disagreement_score":0.8607171,"about_ca_system_score_codex":0.00018130176,"about_ca_system_score_gemma":0.0020851067,"threshold_uncertainty_score":0.38657483},"labels":[],"label_agreement":null},{"id":"W1580192808","doi":"10.1007/978-3-642-05151-7_5","title":"A Synergy between Static and Dynamic Analysis for the Detection of Software Security Vulnerabilities","year":2009,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":11,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Computer science; Static analysis; Vulnerability (computing); Software security assurance; Vulnerability management; Vulnerability assessment; Control flow graph; Symbolic execution; Software; Computer security; Security analysis; Secure coding; Key (lock); Generator (circuit theory); Security testing; Information security; Programming language; Security information and event management; Security service; Operating system; Cloud computing; Cloud computing security","score_opus":0.014835437827803308,"score_gpt":0.2724506335372322,"score_spread":0.25761519570942887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1580192808","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011083334,0.0010329827,0.9967105,0.00053080096,0.000111328125,0.00040005372,0.000025922876,0.000067515895,0.000012582128],"genre_scores_gemma":[0.9180937,0.00012864385,0.08149955,0.000109162516,0.000073515774,0.0000172284,0.000007716993,0.000011455859,0.000059065598],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9968504,0.00008957865,0.00056362833,0.001125843,0.0009134267,0.00045716937],"domain_scores_gemma":[0.992296,0.00565268,0.00027042144,0.0012320875,0.00044080438,0.000108011234],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0019268725,0.00033326892,0.0007429797,0.001102639,0.00037050378,0.00034107507,0.0019109836,0.00021769752,0.000004131449],"category_scores_gemma":[0.0006692105,0.00024261021,0.000325778,0.0015442449,0.0010351435,0.0003186717,0.00058076065,0.00053932494,7.721522e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000005813377,0.00001685939,0.00050746894,0.0001055014,0.00017135414,0.0000020092916,0.0007388158,0.04459165,0.000023092085,0.00064467516,3.5765618e-7,0.9531924],"study_design_scores_gemma":[0.00011780244,0.00021063605,0.001711397,0.00007251679,0.00014918945,0.0000032327155,8.569801e-7,0.7517615,0.00039971442,0.24519859,0.000118181044,0.0002564202],"about_ca_topic_score_codex":0.00015647132,"about_ca_topic_score_gemma":0.00052463135,"teacher_disagreement_score":0.952936,"about_ca_system_score_codex":0.00020340718,"about_ca_system_score_gemma":0.00027834575,"threshold_uncertainty_score":0.9893357},"labels":[],"label_agreement":null},{"id":"W1589080523","doi":"10.1023/a:1023014713253","title":"An Investigation on the Occurrence of Service Requests in Commercial Software Applications","year":2003,"lang":"en","type":"article","venue":"Empirical Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":16,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science; Process (computing); Software quality; Service (business); Identification (biology); Software; Data mining; Reliability engineering; Software development; Engineering","score_opus":0.05197182144963392,"score_gpt":0.3105143093181651,"score_spread":0.2585424878685312,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1589080523","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26259613,0.0000639882,0.7357953,0.0009933586,0.000052160096,0.0002511404,0.0000074814466,0.00023096433,0.000009467744],"genre_scores_gemma":[0.9566641,0.0000076501865,0.04234078,0.0007805287,0.00003243275,0.00014679268,0.000013463335,0.000010975854,0.000003313556],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983722,0.0001996501,0.00033259403,0.0003778434,0.0004247532,0.00029295243],"domain_scores_gemma":[0.99736685,0.0015005817,0.000061006842,0.0007750814,0.00015504994,0.00014141829],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007843614,0.00015170129,0.00019987725,0.00016817312,0.00009778313,0.00006655514,0.00088919664,0.000093882365,0.000015654163],"category_scores_gemma":[0.0016468933,0.00012225869,0.00006225789,0.0020100342,0.0000473778,0.00028756904,0.000078975034,0.00038700964,0.000024974124],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000008271277,0.00049538026,0.84782034,0.00023138936,0.000033030836,0.000007555244,0.0021897992,0.10595814,0.00031401054,0.017958106,0.0004947129,0.024489282],"study_design_scores_gemma":[0.0011982288,0.0005419247,0.7858499,0.0005522298,0.000036456495,0.000021964166,0.00013730333,0.15706402,0.014028675,0.02164332,0.017156469,0.0017695006],"about_ca_topic_score_codex":0.000054132855,"about_ca_topic_score_gemma":0.00002955944,"teacher_disagreement_score":0.69406796,"about_ca_system_score_codex":0.000087955195,"about_ca_system_score_gemma":0.0001225398,"threshold_uncertainty_score":0.49855646},"labels":[],"label_agreement":null},{"id":"W1602556105","doi":"10.1002/0471028959.sof320","title":"Software Engineering Standards in the Nuclear Power Industry—Safety‐Related Applications","year":2002,"lang":"en","type":"other","venue":"Encyclopedia of Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ontario Power Generation","funders":"","keywords":"Nuclear power; Consistency (knowledge bases); Software deployment; Nuclear industry; Safety standards; Engineering; Business; Computer science; Risk analysis (engineering); International trade; Engineering management; Electrical engineering; Software engineering","score_opus":0.004874249613400017,"score_gpt":0.2211324582079217,"score_spread":0.2162582085945217,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1602556105","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000222997,0.006080589,0.9445338,0.00028695643,0.00075744936,0.0017695612,0.00033845368,0.003106081,0.042904112],"genre_scores_gemma":[0.027451882,0.01959347,0.7561476,0.00040395244,0.0023501408,0.0020223048,0.0003804038,0.004780622,0.18686962],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99641997,0.000080035956,0.0007507183,0.0007575249,0.0013060994,0.00068566063],"domain_scores_gemma":[0.9970933,0.00082698365,0.00021932642,0.0015530379,0.00013672425,0.00017058026],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00087842625,0.0005329684,0.0006918503,0.00097113923,0.00007620839,0.000110485205,0.0024057007,0.00091746973,0.0009759617],"category_scores_gemma":[0.0019020344,0.00047500897,0.00027506845,0.0022283792,0.00007297176,0.00023090969,0.0003917231,0.0019444484,0.00008336244],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000038378632,0.0023840903,0.0221171,0.007174349,0.0023638576,0.00082696043,0.015407883,0.25486836,0.000033803623,0.040261485,0.48332343,0.1712003],"study_design_scores_gemma":[0.00031230424,0.000043466913,0.0011952304,0.00051281607,0.00003021642,0.00001707182,0.000034177792,0.0052142614,0.0000021292387,0.000037082304,0.99201083,0.00059039646],"about_ca_topic_score_codex":0.00007846513,"about_ca_topic_score_gemma":0.000006376761,"teacher_disagreement_score":0.50868744,"about_ca_system_score_codex":0.00023143711,"about_ca_system_score_gemma":0.00015270842,"threshold_uncertainty_score":0.9999373},"labels":[],"label_agreement":null},{"id":"W1606514821","doi":"10.1007/978-3-540-85553-8_15","title":"Assessment of Real-Time Software Specifications Quality Using COSMIC-FFP","year":2008,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University; École de Technologie Supérieure; Université du Québec à Montréal","funders":"","keywords":"Computer science; Software requirements specification; Software; Software requirements; Software construction; Software system; Software development; Functional requirement; Verification and validation; Software engineering; Reliability engineering; Operating system; Engineering","score_opus":0.08012313928811668,"score_gpt":0.34996270264938456,"score_spread":0.2698395633612679,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1606514821","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0009839664,0.00018120394,0.99619055,0.0002656819,0.0002981874,0.0003568878,0.000019576632,0.00015775822,0.0015461834],"genre_scores_gemma":[0.055715565,0.00040555804,0.9432201,0.00013031307,0.00019782016,0.000006487006,0.000013483737,0.000028825714,0.00028186754],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99440664,0.00015840352,0.0010348087,0.0016681707,0.0020869083,0.0006450396],"domain_scores_gemma":[0.99410653,0.001717332,0.0005601969,0.0025467863,0.0008493593,0.00021978296],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0019969244,0.00044987779,0.00090019713,0.0010522212,0.00039422914,0.0002693383,0.003897607,0.0003120679,0.00006132079],"category_scores_gemma":[0.00036018906,0.00042503423,0.0003190251,0.0013937097,0.0013986668,0.00051316206,0.0014791533,0.0008061813,0.000028352924],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013049067,0.00054165703,0.0063150283,0.00038372813,0.00015444521,0.00014765504,0.0015903448,0.25012925,0.0061965976,0.028641794,0.00014334556,0.7057431],"study_design_scores_gemma":[0.0003366496,0.0001695202,0.008638914,0.0004888878,0.000022667045,0.00006621264,3.6985335e-7,0.9354538,0.0013793797,0.05162845,0.0007849898,0.001030194],"about_ca_topic_score_codex":0.00021725727,"about_ca_topic_score_gemma":0.00002959144,"teacher_disagreement_score":0.7047129,"about_ca_system_score_codex":0.0007035017,"about_ca_system_score_gemma":0.0022116236,"threshold_uncertainty_score":0.9998202},"labels":[],"label_agreement":null},{"id":"W1607036753","doi":"10.1007/978-3-642-24580-0_3","title":"Improved Usage Model for Web Application Reliability Testing","year":2011,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Markov chain; Component (thermodynamics); Reliability (semiconductor); Markov model; Data mining; Reliability engineering; Machine learning","score_opus":0.04299277726514847,"score_gpt":0.2840944800349134,"score_spread":0.24110170276976492,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1607036753","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000091193324,0.000114268645,0.9969671,0.0003290381,0.00025552965,0.0010258495,0.000012960967,0.00026691143,0.0009371598],"genre_scores_gemma":[0.2234953,0.000012899626,0.77554435,0.00032487515,0.00019768483,0.00010091011,0.000006575587,0.000031477106,0.0002859161],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9953156,0.000041312684,0.0006888968,0.002320417,0.00085638306,0.00077740074],"domain_scores_gemma":[0.9944331,0.0016751549,0.00031849975,0.0024228857,0.0009211355,0.00022922279],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0028639808,0.0004809229,0.00058844226,0.0006285865,0.00039150496,0.0004133415,0.004242203,0.00040179916,0.000005132977],"category_scores_gemma":[0.0011316998,0.00043441646,0.00024319725,0.0009305612,0.00074985885,0.0006260121,0.0014572173,0.0007784194,0.000018554254],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00000992815,0.00006602824,0.0001352172,0.00012763037,0.000010240784,0.0000027121469,0.00029111624,0.103411704,0.0011065485,0.008230639,0.000007655416,0.88660055],"study_design_scores_gemma":[0.00013082157,0.00008636893,0.000025053925,0.000065259206,0.0000072753733,0.000004130509,2.573425e-8,0.7104441,0.00039173523,0.28835973,0.0001509509,0.0003345356],"about_ca_topic_score_codex":0.00007259817,"about_ca_topic_score_gemma":0.00006188843,"teacher_disagreement_score":0.88626605,"about_ca_system_score_codex":0.00040855812,"about_ca_system_score_gemma":0.0012614462,"threshold_uncertainty_score":0.99981076},"labels":[],"label_agreement":null},{"id":"W1671257183","doi":"10.1109/rams.2002.981650","title":"Reliability estimation of hierarchical software systems","year":2003,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Military College of Canada","funders":"","keywords":"Reliability engineering; Computer science; Software quality; Software reliability testing; Reliability (semiconductor); Reliability block diagram; Criticality; Software; Process (computing); Software sizing; Software system; Software development; Software construction; Fault tree analysis; Engineering; Programming language","score_opus":0.016181020023914583,"score_gpt":0.27671932867208615,"score_spread":0.26053830864817157,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1671257183","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0374263,0.00007728878,0.960471,0.00011219816,0.000086799264,0.00012995774,9.568789e-7,0.00014669672,0.0015488386],"genre_scores_gemma":[0.8371229,0.0000065241425,0.16260262,0.000015901738,0.000005398058,0.000010655514,0.0000011592815,0.0000028327834,0.00023197712],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99832237,0.00031506363,0.0003318409,0.00032628042,0.0004990997,0.00020534382],"domain_scores_gemma":[0.9981922,0.0006479801,0.000054617005,0.00079280685,0.00021051643,0.00010190691],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012693538,0.00007862827,0.00020059227,0.000103136095,0.00006790494,0.00007174734,0.00048736818,0.00006674131,0.000043401225],"category_scores_gemma":[0.00269511,0.00006297011,0.00009800807,0.00065904355,0.0000939134,0.00029145565,0.00008744671,0.00013347028,0.000042187174],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009464188,0.00056641886,0.020916464,0.00040151717,0.00004996304,0.0000060171615,0.00033376666,0.102372184,0.00025580084,0.8046632,0.0008825218,0.06954271],"study_design_scores_gemma":[0.00033287733,0.00014207377,0.0071776076,0.000039694016,0.000009760587,0.000015230315,0.00003634835,0.92540634,0.0034198312,0.060935646,0.0022353458,0.00024927358],"about_ca_topic_score_codex":0.00008250612,"about_ca_topic_score_gemma":0.000001268015,"teacher_disagreement_score":0.8230341,"about_ca_system_score_codex":0.00004253709,"about_ca_system_score_gemma":0.00013319444,"threshold_uncertainty_score":0.3226493},"labels":[],"label_agreement":null},{"id":"W17529471","doi":"10.1007/978-3-642-38911-5_8","title":"Modeling Component Erroneous Behavior and Error Propagation for Dependability Analysis","year":2013,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"","keywords":"Dependability; Computer science; Component (thermodynamics); Unified Modeling Language; Protocol (science); Maintainability; Software; Propagation of uncertainty; Aspect-oriented programming; Distributed computing; Programming language; Algorithm; Software engineering","score_opus":0.03251929677582509,"score_gpt":0.29283129821347936,"score_spread":0.26031200143765426,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W17529471","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011393553,0.00027758235,0.9861775,0.0005419606,0.00022280608,0.0012032035,0.000009618454,0.00012745899,0.000046338904],"genre_scores_gemma":[0.7103912,0.000025046782,0.28915462,0.000128531,0.00009411548,0.00009636363,0.000018131217,0.000018063749,0.00007391943],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99538636,0.00006851605,0.00069677463,0.0020362034,0.0011484489,0.00066366745],"domain_scores_gemma":[0.9967306,0.00067514455,0.00019213067,0.0014600211,0.0006896649,0.00025241464],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0016813226,0.00045853265,0.0007931389,0.0012787007,0.00041226184,0.00077931624,0.0019759876,0.00031737948,0.000026120091],"category_scores_gemma":[0.0002357855,0.00039838368,0.00033184144,0.0010174989,0.00058711227,0.00060100463,0.0010921917,0.0005996985,0.000013437086],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009185687,0.00010076858,0.0007408473,0.00010944322,0.00009891416,0.000012516622,0.00042777258,0.4564984,0.00010966073,0.0028586506,0.0000017999616,0.53903204],"study_design_scores_gemma":[0.00016306905,0.000111984904,0.00033356607,0.00005390027,0.00012794489,0.000010292556,2.877557e-7,0.9633332,0.00015300304,0.035255056,0.000032487926,0.00042524136],"about_ca_topic_score_codex":0.00022349124,"about_ca_topic_score_gemma":0.00032436693,"teacher_disagreement_score":0.6989977,"about_ca_system_score_codex":0.00036713568,"about_ca_system_score_gemma":0.0003652183,"threshold_uncertainty_score":0.9998468},"labels":[],"label_agreement":null},{"id":"W1792129892","doi":"","title":"Towards the profiling of scientific software for accuracy","year":2011,"lang":"en","type":"article","venue":"Conference of the Centre for Advanced Studies on Collaborative Research","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Military College of Canada; Queen's University","funders":"","keywords":"Computer science; Software; Profiling (computer programming); Pareto principle; Search-based software engineering; Propagation of uncertainty; Software quality; Code (set theory); Pareto analysis; Software sizing; Data mining; Software construction; Software development; Reliability engineering; Set (abstract data type); Algorithm; Mathematical optimization; Engineering; Mathematics; Programming language","score_opus":0.255076314978505,"score_gpt":0.42636188734353153,"score_spread":0.1712855723650265,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1792129892","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.33753204,0.008849376,0.5745205,0.03830775,0.0043819794,0.0314865,0.0019822728,0.00030806256,0.0026315127],"genre_scores_gemma":[0.94564635,0.00030619028,0.052508313,0.00002022246,0.000031763364,0.000585403,0.000004538726,0.000014387109,0.000882864],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.99666375,0.00054174056,0.00047654958,0.0006080609,0.0011053898,0.00060449046],"domain_scores_gemma":[0.97356725,0.0057342784,0.00033763584,0.0013846545,0.018891774,0.00008441536],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0036914684,0.00017870468,0.00043502156,0.00019611436,0.0011874142,0.00012156958,0.002741841,0.00005578158,0.000006633163],"category_scores_gemma":[0.031012757,0.00009636798,0.00020038954,0.003283423,0.0018685971,0.00032242673,0.000997157,0.00027900378,0.0000026359144],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0014201627,0.0010369907,0.001977554,0.0016372134,0.0009767707,0.0000014588301,0.039160654,0.0007904778,0.0159494,0.71052665,0.0039905375,0.22253215],"study_design_scores_gemma":[0.001125366,0.0008710425,0.00034947737,0.00048134214,0.000029410137,2.7501565e-7,0.019295974,0.0051710545,0.8311,0.13575527,0.005578712,0.00024206757],"about_ca_topic_score_codex":0.0000074168356,"about_ca_topic_score_gemma":0.00006801778,"teacher_disagreement_score":0.8151506,"about_ca_system_score_codex":0.00012941471,"about_ca_system_score_gemma":0.0014311994,"threshold_uncertainty_score":0.9771494},"labels":[],"label_agreement":null},{"id":"W1797917470","doi":"","title":"Identification, specification and measurement, using international standards, of the system non functional requirements allocated to real-time embedded software","year":2011,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Software engineering; Software requirements specification; Software development; Computer science; Software requirements; Software construction; Verification and validation; Software project management; Systems engineering; Software system; Functional requirement; Non-functional requirement; Software development process; Software; Engineering; Operating system; Operations management","score_opus":0.08203347247129586,"score_gpt":0.291877684963005,"score_spread":0.20984421249170912,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1797917470","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06681912,0.000009270787,0.93117815,0.00018532515,0.00023154353,0.00030049033,0.000010073028,0.00008815196,0.0011778612],"genre_scores_gemma":[0.9789337,0.000007945648,0.02070742,0.000018307777,0.00002879594,0.000014720231,0.0000040642244,0.0000065297518,0.0002785004],"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99742776,0.00011602996,0.00043153373,0.00037880565,0.0014984546,0.00014738916],"domain_scores_gemma":[0.9977224,0.000033939486,0.00017000384,0.00060766767,0.0013879704,0.00007804727],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0018219482,0.0000964078,0.00013220472,0.00015902704,0.00016880943,0.00010041369,0.00069742004,0.00004188125,0.000072709794],"category_scores_gemma":[0.000270885,0.00007261324,0.00006177504,0.0005785845,0.000061085804,0.00035879062,0.00026316583,0.000057479054,0.000020140824],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028971475,0.00084896374,0.0810952,0.00034781932,0.0009258215,0.000003743119,0.005291891,0.0011680076,0.85641026,0.018358227,0.007338236,0.027922105],"study_design_scores_gemma":[0.0015154462,0.00015324108,0.38211876,0.0005637818,0.0001306923,0.000026728625,0.0010399523,0.38482964,0.22517951,0.0028030814,0.0008962875,0.00074286613],"about_ca_topic_score_codex":0.0002605519,"about_ca_topic_score_gemma":0.00001887492,"teacher_disagreement_score":0.9121146,"about_ca_system_score_codex":0.00035987352,"about_ca_system_score_gemma":0.00019668804,"threshold_uncertainty_score":0.29610822},"labels":[],"label_agreement":null},{"id":"W1822490568","doi":"","title":"IT and infrastructure's lost dependability","year":2008,"lang":"en","type":"article","venue":"International Conference on Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Dependability; Information system; Computer science; Information infrastructure; Core (optical fiber); Software; Quarter (Canadian coin); Computer security; Risk analysis (engineering); Engineering; Business; Software engineering; Telecommunications","score_opus":0.02887983941232447,"score_gpt":0.27492236476378823,"score_spread":0.24604252535146376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1822490568","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17392023,0.000031680964,0.82296944,0.0014024918,0.00033412353,0.00008737665,0.000009961849,0.00025476934,0.0009899597],"genre_scores_gemma":[0.9611889,0.000098825665,0.038307667,0.00017405467,0.000059643866,0.000008800715,0.000004990999,0.000007324406,0.0001498],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9986508,0.000023411758,0.00020218437,0.00040449004,0.00050808,0.00021101795],"domain_scores_gemma":[0.9990244,0.00024632638,0.000035671783,0.00033676904,0.00023096566,0.00012587776],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00017967503,0.0001472502,0.00015145529,0.00017609044,0.00008160486,0.00012700191,0.0007249033,0.000066831184,0.00019978861],"category_scores_gemma":[0.000943873,0.00013917292,0.00006026976,0.00019634108,0.00006374182,0.0003732124,0.00026120295,0.00027479895,0.000069538786],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001028189,0.0004892592,0.42471606,0.00021341418,0.0006094287,0.0005652315,0.005483791,0.041160204,0.0026251497,0.37390995,0.004396177,0.14572851],"study_design_scores_gemma":[0.0007169154,0.00016153473,0.20171432,0.00013617742,0.0000074512477,0.0002694994,0.00004521465,0.77317286,0.001652401,0.0045721494,0.016803123,0.0007483549],"about_ca_topic_score_codex":0.000016754528,"about_ca_topic_score_gemma":0.0000029570183,"teacher_disagreement_score":0.7872687,"about_ca_system_score_codex":0.00008466374,"about_ca_system_score_gemma":0.00008659322,"threshold_uncertainty_score":0.56753063},"labels":[],"label_agreement":null},{"id":"W185646288","doi":"","title":"Anisotropic Laplace trend to enhance software reliability growth modelling","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Systems, Applications & Products in Data Processing (Canada); Concordia University","funders":"","keywords":"Laplace transform; Reliability engineering; Reliability (semiconductor); Computer science; Software quality; Software; Statistic; Mathematical optimization; Software development; Mathematics; Engineering; Statistics; Mathematical analysis; Physics","score_opus":0.016243635427958907,"score_gpt":0.2988234154306988,"score_spread":0.28257978000273987,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W185646288","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09183045,0.00005542772,0.9048799,0.0013043492,0.00009449135,0.00015882951,0.0000015740217,0.00045108324,0.0012238622],"genre_scores_gemma":[0.61270505,0.000020597834,0.38559297,0.00038985303,0.00005629628,0.000008673322,0.0000017496088,0.000008726423,0.0012160726],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.997206,0.00009193583,0.00041241507,0.00088078843,0.00071375247,0.00069513917],"domain_scores_gemma":[0.99746734,0.0007392521,0.000046656816,0.0010966434,0.0002646305,0.0003854855],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016667059,0.00018770949,0.0002616776,0.00023104351,0.00021254292,0.00018146014,0.0012753619,0.00010202657,0.00008055372],"category_scores_gemma":[0.0005797849,0.00016226481,0.00015161639,0.0015780346,0.00006001909,0.00049155037,0.0004474339,0.00025393293,0.00027124683],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032404487,0.0022421794,0.10742151,0.00052989525,0.00021636808,0.00021209073,0.006112458,0.15508652,0.0031973538,0.13213587,0.011537731,0.580984],"study_design_scores_gemma":[0.0010131723,0.0011230918,0.014717988,0.00015443914,0.000047214675,0.000031619296,0.0003201125,0.706091,0.13468939,0.09981166,0.039273582,0.0027267383],"about_ca_topic_score_codex":0.0003623591,"about_ca_topic_score_gemma":0.00019390984,"teacher_disagreement_score":0.57825726,"about_ca_system_score_codex":0.00015873078,"about_ca_system_score_gemma":0.00007891328,"threshold_uncertainty_score":0.66169673},"labels":[],"label_agreement":null},{"id":"W1860683784","doi":"10.1109/issre.1991.145358","title":"Practical applications of software reliability models","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bell (Canada)","funders":"","keywords":"Reliability (semiconductor); Computer science; Software engineering; Software quality; Software; Focus (optics); Software development; Process (computing); Reliability engineering; Engineering; Programming language","score_opus":0.06322917223966136,"score_gpt":0.32569986309162824,"score_spread":0.2624706908519669,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1860683784","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00044728196,0.000054827353,0.9916143,0.0033389395,0.000011381584,0.0001672166,0.000001637833,0.00018224775,0.004182194],"genre_scores_gemma":[0.52707773,0.000054635762,0.47199577,0.00008051782,0.000014006741,0.000051926007,7.8376036e-7,0.000003328967,0.0007213156],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99863344,0.00008867101,0.00026872152,0.00036204702,0.0004507007,0.00019639442],"domain_scores_gemma":[0.9977938,0.0006769658,0.000051026687,0.0010758399,0.00029252123,0.00010987776],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00046987578,0.00007239546,0.00015579278,0.00008251896,0.000076104414,0.000046699828,0.000607662,0.000061441155,0.00031357136],"category_scores_gemma":[0.0004834858,0.000059326605,0.000111050875,0.0007922264,0.00011122776,0.0005626099,0.00024461272,0.00014742425,0.00014720584],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000071383192,0.0028216587,0.007553176,0.00020367412,0.00008058209,0.0000063699995,0.00065673364,0.007628625,0.00008969724,0.6175578,0.02176426,0.34163025],"study_design_scores_gemma":[0.00011894026,0.000044176486,0.00020624096,0.000004424845,0.0000073493643,0.0000058585288,0.00001805496,0.90127057,0.00061142293,0.09243541,0.00515264,0.0001248947],"about_ca_topic_score_codex":0.000046439163,"about_ca_topic_score_gemma":0.000002849935,"teacher_disagreement_score":0.89364195,"about_ca_system_score_codex":0.000029306306,"about_ca_system_score_gemma":0.00003997326,"threshold_uncertainty_score":0.34333864},"labels":[],"label_agreement":null},{"id":"W1944451760","doi":"10.5267/j.dsl.2015.7.002","title":"An effort allocation model considering different budgetary constraint on fault detection process and fault correction process","year":2015,"lang":"en","type":"article","venue":"Decision Science Letters","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Process (computing); Fault (geology); Constraint (computer-aided design); Computer science; Fault detection and isolation; Reliability engineering; Budget constraint; Mathematical optimization; Engineering; Mathematics; Economics; Artificial intelligence; Microeconomics","score_opus":0.03555906236138411,"score_gpt":0.33545725411165295,"score_spread":0.2998981917502688,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1944451760","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.523701,0.0000050542717,0.47509536,0.0007425542,0.00020117876,0.00014423675,4.391767e-7,0.00009018961,0.000019964362],"genre_scores_gemma":[0.99463516,0.0000057928078,0.0040567815,0.0012207378,0.00003303531,0.000035564506,0.000001781284,0.000007207618,0.0000039330544],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9963311,0.000064630076,0.00034795472,0.0009995204,0.0018448728,0.0004119165],"domain_scores_gemma":[0.9980201,0.00022466847,0.00011749725,0.00060862396,0.0005177891,0.00051128346],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017634064,0.00019635477,0.00021510788,0.0005552298,0.0004659887,0.00069141336,0.0008789065,0.000071925715,0.0000016135996],"category_scores_gemma":[0.0012300139,0.00015654115,0.000046118097,0.0013041913,0.0005419644,0.0018948577,0.00012429278,0.00027944983,0.000012282268],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008215612,0.0002265485,0.0064012725,0.000024438514,0.000009706931,0.000011595453,0.0030536382,0.4968428,0.04196791,0.0001512952,0.000114629365,0.451114],"study_design_scores_gemma":[0.00028911108,0.00015464728,0.0025868877,0.00004261818,0.000004765983,0.000034784323,0.00033438677,0.9582669,0.035434727,0.0026478458,0.000009723706,0.00019359098],"about_ca_topic_score_codex":0.00003072614,"about_ca_topic_score_gemma":0.000023558168,"teacher_disagreement_score":0.47103858,"about_ca_system_score_codex":0.00020797635,"about_ca_system_score_gemma":0.00027448297,"threshold_uncertainty_score":0.6667315},"labels":[],"label_agreement":null},{"id":"W1965086994","doi":"10.1109/sere-c.2013.30","title":"Reliability Models Applied to Mobile Applications","year":2013,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Computer science; Laptop; Reliability (semiconductor); Smart phone; Reliability engineering; Software; Embedded system; Mobile phone; Controller (irrigation); Mobile device; Operating system; Engineering; Telecommunications","score_opus":0.015365408840581818,"score_gpt":0.2686449609344572,"score_spread":0.2532795520938754,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1965086994","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0059634815,0.000012553804,0.97083634,0.00074247754,0.000013059251,0.001600356,7.457782e-7,0.00035082575,0.020480176],"genre_scores_gemma":[0.8452603,0.000005661298,0.14628659,0.0005181462,0.000025034093,0.006843609,0.0000017357921,0.0000051450147,0.0010537324],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9985334,0.00003437515,0.00022121437,0.00055064727,0.00035920079,0.0003011202],"domain_scores_gemma":[0.9979156,0.00019492183,0.000019967929,0.0013680689,0.0002447046,0.00025673088],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00036107397,0.00009654889,0.00014721425,0.00010291449,0.00012780783,0.00022068825,0.0011393708,0.000050544448,0.00028692698],"category_scores_gemma":[0.000030243962,0.00007721404,0.00007303219,0.0009679629,0.000041033025,0.00043585268,0.00044162443,0.000109638815,0.0032999904],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000031163186,0.0004685726,0.00024082304,0.000037455146,0.000022792197,2.9553703e-7,0.0004079345,0.06758579,0.0030223066,0.31950775,0.008318738,0.6003844],"study_design_scores_gemma":[0.00017773776,0.00008936287,0.0010730276,0.0000032337464,0.000004837434,0.0000012392437,0.00011559003,0.57260776,0.0024743823,0.3782768,0.044771485,0.00040454435],"about_ca_topic_score_codex":0.0002369503,"about_ca_topic_score_gemma":0.0000049270157,"teacher_disagreement_score":0.8392969,"about_ca_system_score_codex":0.000055231332,"about_ca_system_score_gemma":0.000053687694,"threshold_uncertainty_score":0.99747604},"labels":[],"label_agreement":null},{"id":"W1972568562","doi":"10.4271/2015-01-1488","title":"Front Underride Protection Devices (FUPDs): Multi-Objective Optimization","year":2015,"lang":"en","type":"article","venue":"SAE technical papers on CD-ROM/SAE technical paper series","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Ontario Tech University","funders":"","keywords":"Computer science; Front (military); Engineering; Mechanical engineering","score_opus":0.03552643987034871,"score_gpt":0.29042480640786467,"score_spread":0.25489836653751596,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1972568562","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.36822587,0.007731604,0.21855627,0.12031017,0.0052321074,0.029241147,0.00028251088,0.07141365,0.17900668],"genre_scores_gemma":[0.9221531,0.00014225791,0.075588,0.000958129,0.00012792529,0.0005203088,0.000024683322,0.00006344086,0.00042210615],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.99385166,0.000562134,0.0010255758,0.0017891467,0.001753569,0.0010178887],"domain_scores_gemma":[0.9959978,0.0004961665,0.00029501828,0.0019710464,0.00055130356,0.0006886786],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0017481188,0.00068267225,0.00086278206,0.00042182283,0.00054952677,0.0004469475,0.0022123647,0.00080696924,0.00017270625],"category_scores_gemma":[0.0028421255,0.0005750434,0.00047745235,0.0017102568,0.0007928557,0.001629012,0.001031172,0.0014349946,0.00034014098],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00094258494,0.0034298196,0.0012659146,0.0002088316,0.00044472454,0.00016040816,0.00053019746,0.024862118,0.8907371,0.019790662,0.004656443,0.05297121],"study_design_scores_gemma":[0.0017241499,0.0025105095,0.9682623,0.00025875206,0.00009624291,0.00011419787,0.0003269248,0.00026162094,0.00031427102,0.0061116214,0.018701397,0.0013179764],"about_ca_topic_score_codex":0.0003220557,"about_ca_topic_score_gemma":0.06524747,"teacher_disagreement_score":0.96699643,"about_ca_system_score_codex":0.00091727474,"about_ca_system_score_gemma":0.00040259262,"threshold_uncertainty_score":0.9996701},"labels":[],"label_agreement":null},{"id":"W1976430831","doi":"10.1145/2737182.2737189","title":"Modeling Fault Tolerance Tactics with Reusable Aspects","year":2015,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Albaha University","keywords":"Dependability; Computer science; Unified Modeling Language; Model transformation; Software engineering; Software fault tolerance; Applications of UML; Fault tolerance; UML tool; Context (archaeology); Process (computing); Software; Programming language; Reliability engineering; Distributed computing; Consistency (knowledge bases); Engineering; Artificial intelligence","score_opus":0.048661701088319485,"score_gpt":0.290114288761452,"score_spread":0.24145258767313252,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1976430831","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04473334,0.00008389556,0.93703294,0.0012310862,0.00004124932,0.00006351804,1.8369616e-7,0.00019905198,0.016614737],"genre_scores_gemma":[0.8878568,0.000011270747,0.11087499,0.00017639401,0.00003623956,0.0000053940557,5.445886e-7,0.0000048858974,0.0010334837],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99862206,0.000046153506,0.00013204537,0.00033153492,0.00058924535,0.00027898408],"domain_scores_gemma":[0.9986916,0.000050658564,0.000019135703,0.0006692436,0.00036942738,0.0001998917],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004431883,0.00008734807,0.00013802251,0.000064768356,0.00007386243,0.00022123639,0.0006766252,0.000038259186,0.000013621942],"category_scores_gemma":[0.00018618758,0.000061343686,0.000033604185,0.00064030546,0.000027959486,0.00058228825,0.00017245881,0.0001495854,0.00013185402],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019355577,0.00014024126,0.0011313562,0.00001700982,0.000033508324,0.000042272346,0.00047893747,0.95451456,0.00009655694,0.035062596,0.0011194304,0.007344145],"study_design_scores_gemma":[0.00019858612,0.000076062155,0.000007693602,0.000008992083,0.000002708329,0.000009952228,0.00006016189,0.9905764,0.00089813455,0.0069817654,0.0010737769,0.000105772815],"about_ca_topic_score_codex":0.0002912758,"about_ca_topic_score_gemma":0.00005663819,"teacher_disagreement_score":0.84312344,"about_ca_system_score_codex":0.00006069028,"about_ca_system_score_gemma":0.00022472156,"threshold_uncertainty_score":0.2501523},"labels":[],"label_agreement":null},{"id":"W1980089213","doi":"10.1109/compsac.2012.23","title":"Evaluating Reliability-Testing Usage Models","year":2012,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Reliability (semiconductor); Computer science; Reliability engineering; Goodness of fit; Markov chain; Markov model; Non-regression testing; Software quality; Markov process; Data mining; Machine learning; Statistics; Software; Engineering; Mathematics","score_opus":0.21979385986111616,"score_gpt":0.4020445771824858,"score_spread":0.18225071732136966,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980089213","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15155163,0.00013418819,0.83347017,0.00043973592,0.0000951341,0.00012052165,2.7577704e-7,0.00035847188,0.013829862],"genre_scores_gemma":[0.6650276,0.0000025322863,0.33444005,0.00009913299,0.00006664017,0.000012675433,4.3990246e-7,0.00000446527,0.0003464651],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9977809,0.00027532072,0.0002798967,0.0003685691,0.000699986,0.0005952823],"domain_scores_gemma":[0.99726063,0.001325012,0.000048542046,0.00087735563,0.00025834845,0.000230118],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004452098,0.00011132726,0.00015952828,0.00009508201,0.00021159119,0.00015995212,0.00079919474,0.00005483513,0.00007580212],"category_scores_gemma":[0.0022453303,0.000089568406,0.00009111459,0.0009712772,0.00004884851,0.0016928981,0.000514769,0.0001970298,0.00020329842],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006527797,0.00067813107,0.09663051,0.00011403362,0.000050096238,0.0000037403345,0.0022832972,0.09269369,0.0034716674,0.12356829,0.00063825777,0.6798618],"study_design_scores_gemma":[0.000083767525,0.000036970145,0.0015220921,0.0000089817295,0.000004494515,0.0000044962944,0.000019836556,0.9770011,0.00061605935,0.020382227,0.00019110758,0.00012889675],"about_ca_topic_score_codex":0.00016534123,"about_ca_topic_score_gemma":0.0000011487044,"teacher_disagreement_score":0.8843074,"about_ca_system_score_codex":0.00007253727,"about_ca_system_score_gemma":0.00008350309,"threshold_uncertainty_score":0.36524937},"labels":[],"label_agreement":null},{"id":"W1980099301","doi":"10.1145/1594156.1594165","title":"Selection of customers for operational and usage profiling","year":2009,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"IBM (Canada)","funders":"","keywords":"Profiling (computer programming); Computer science; Selection (genetic algorithm); Artificial intelligence; Operating system","score_opus":0.020222527349720812,"score_gpt":0.308693057931512,"score_spread":0.28847053058179123,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980099301","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08086588,0.000018670793,0.917453,0.000828868,0.0000110227575,0.00012192521,4.9487386e-7,0.000032557473,0.0006675859],"genre_scores_gemma":[0.8791604,0.000006481671,0.12055149,0.00007603988,0.000011362953,0.0000038337757,9.650252e-7,6.031745e-7,0.00018883118],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9995969,0.000014913038,0.00008069419,0.00012193271,0.000113522095,0.00007202574],"domain_scores_gemma":[0.9997145,0.00008164748,0.000012284368,0.0000624958,0.00010490359,0.000024147672],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00026755375,0.000026407333,0.000054192005,0.000056947436,0.00005448163,0.000037078382,0.000093153656,0.000017196588,0.0000065462277],"category_scores_gemma":[0.00007472917,0.000021224874,0.000023394321,0.00020862625,0.000013846546,0.00017878698,0.000014391713,0.000025753992,8.699417e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027669617,0.00024067826,0.022628305,0.00007581511,0.000043972763,4.290852e-7,0.00032942565,0.0019408532,0.08215049,0.50120944,0.0006133671,0.39073956],"study_design_scores_gemma":[0.0002351305,0.0002210304,0.006555408,0.0000050503804,0.0000036385266,0.000002664594,0.000017200575,0.8990156,0.08772811,0.005496093,0.0006439308,0.000076138494],"about_ca_topic_score_codex":0.000005676487,"about_ca_topic_score_gemma":0.0000014662746,"teacher_disagreement_score":0.89707476,"about_ca_system_score_codex":0.00000877715,"about_ca_system_score_gemma":0.00003611576,"threshold_uncertainty_score":0.08655252},"labels":[],"label_agreement":null},{"id":"W1982739237","doi":"10.1016/s0305-0548(03)00143-6","title":"Resource allocation during tests for optimally reliable software","year":2003,"lang":"en","type":"article","venue":"Computers & Operations Research","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Computer science; Reliability engineering; Test plan; Test Management Approach; System under test; Software reliability testing; Software; White-box testing; Reliability (semiconductor); Test strategy; Test case; Software system; Solver; Regression testing; Software construction; Statistics; Mathematics; Engineering; Operating system; Programming language","score_opus":0.06357624890948606,"score_gpt":0.36566602436698364,"score_spread":0.3020897754574976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1982739237","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03170052,0.00028673976,0.9637487,0.002489133,0.0001180683,0.0009159273,0.000004314669,0.00026867515,0.0004679385],"genre_scores_gemma":[0.35219437,0.00006812324,0.6437321,0.00013965489,0.0001315272,0.0004911553,0.000040430892,0.000034293964,0.0031683906],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9961318,0.0005660145,0.000426877,0.0009191122,0.0010570061,0.000899238],"domain_scores_gemma":[0.99558496,0.0011624387,0.000025707473,0.0013965687,0.0015429029,0.0002874436],"candidate_categories":["sts","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0032440417,0.0001876659,0.00024552163,0.00065253297,0.0017796396,0.0011970564,0.0016276252,0.00012690731,0.000031680884],"category_scores_gemma":[0.0026398012,0.00018554786,0.00014648263,0.002166821,0.00018192863,0.00080316357,0.00045153213,0.0004971089,0.00014012349],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006269268,0.0011065059,0.002735779,0.00037321346,0.00020492461,0.0000409032,0.002213771,0.7236557,0.008326646,0.19367693,0.02862751,0.038975433],"study_design_scores_gemma":[0.0015946161,0.0004895913,0.002117658,0.00015103933,0.000012188483,0.000051511965,0.00019980804,0.8981147,0.014346047,0.0035441893,0.07869383,0.00068481883],"about_ca_topic_score_codex":0.00006297958,"about_ca_topic_score_gemma":0.000041440748,"teacher_disagreement_score":0.32049385,"about_ca_system_score_codex":0.00029520007,"about_ca_system_score_gemma":0.0006366451,"threshold_uncertainty_score":0.9998398},"labels":[],"label_agreement":null},{"id":"W1986968726","doi":"10.4271/2013-01-0727","title":"Forward Collision Warning Timing in Near Term Applications","year":2013,"lang":"en","type":"article","venue":"SAE International Journal of Transportation Safety","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Chrysler (Canada)","funders":"","keywords":"Term (time); Collision; Warning system; Computer science; Computer security; Risk analysis (engineering); Business; Telecommunications; Physics","score_opus":0.01470087945175076,"score_gpt":0.3018674777838837,"score_spread":0.2871665983321329,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1986968726","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20465223,0.00008056564,0.79159117,0.0031229563,0.00018457687,0.0001758673,0.00000817675,0.000022328484,0.00016212698],"genre_scores_gemma":[0.95709145,0.0001524806,0.042477485,0.00011249527,0.000076797085,0.000017844897,0.000023165874,0.0000056916433,0.00004259062],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99808156,0.000057218866,0.0007547114,0.0001719504,0.00078714034,0.00014743107],"domain_scores_gemma":[0.99831367,0.00023386431,0.00030109414,0.00015460566,0.0009031669,0.00009360443],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005976895,0.00008685809,0.00017214479,0.0003177029,0.00007626888,0.0001824382,0.00091151305,0.000053019572,0.00011045415],"category_scores_gemma":[0.00006588551,0.000077644705,0.00015928612,0.00042850844,0.00004039449,0.0011543077,0.000018669201,0.0002500876,0.000043792432],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013495068,0.00048312487,0.25551802,0.000052300056,0.00031923602,0.00010731312,0.00562991,0.07929499,0.0036994934,0.010343039,0.00045111845,0.6439665],"study_design_scores_gemma":[0.0012009015,0.000069544905,0.9571001,0.00015155031,0.000015089687,0.000025700596,0.0002028706,0.020953765,0.0005995821,0.0063390303,0.013164706,0.00017717392],"about_ca_topic_score_codex":0.00007681863,"about_ca_topic_score_gemma":0.000058406553,"teacher_disagreement_score":0.7524392,"about_ca_system_score_codex":0.00013903213,"about_ca_system_score_gemma":0.00015769569,"threshold_uncertainty_score":0.31662592},"labels":[],"label_agreement":null},{"id":"W1987409374","doi":"10.1007/s10664-008-9084-6","title":"Another viewpoint on “evaluating web software reliability based on workload and failure data extracted from server logs”","year":2008,"lang":"en","type":"article","venue":"Empirical Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":22,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Workload; Reliability (semiconductor); Computer science; Replication (statistics); Reliability engineering; Software quality; Web application; Term (time); Data mining; Software; World Wide Web; Engineering; Statistics; Software development; Operating system","score_opus":0.08123406468878247,"score_gpt":0.3199451369208011,"score_spread":0.23871107223201865,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987409374","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4123276,0.00029956765,0.58386445,0.0019493002,0.00015044305,0.00028601897,0.000058160527,0.0010561008,0.000008349882],"genre_scores_gemma":[0.6820945,0.00005738727,0.315537,0.0017709819,0.00023850794,0.000049224425,0.000103502156,0.00007242258,0.000076496224],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.995521,0.00028980718,0.0005774604,0.0016340318,0.0012844811,0.0006932264],"domain_scores_gemma":[0.9922481,0.0043278993,0.00010631193,0.0028032626,0.00017868105,0.00033570617],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011863708,0.00046357463,0.00056984177,0.00022878601,0.00030892485,0.0001907357,0.0016807633,0.0002955817,0.00017193695],"category_scores_gemma":[0.0076682307,0.00040468964,0.00017101326,0.0010892368,0.00010921696,0.00059105863,0.00080516544,0.0010101192,0.00011904037],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024070432,0.0016884321,0.58530873,0.00041480656,0.00028075514,0.00042440457,0.0011868854,0.2951661,0.0006537586,0.00005798776,0.009971132,0.10460629],"study_design_scores_gemma":[0.0010132845,0.0003014305,0.1601306,0.0004338396,0.000028605331,0.000010990271,0.000008477783,0.8281066,0.00030684826,0.00023919152,0.008633146,0.00078700134],"about_ca_topic_score_codex":0.00007434332,"about_ca_topic_score_gemma":0.000012527767,"teacher_disagreement_score":0.53294045,"about_ca_system_score_codex":0.00021110011,"about_ca_system_score_gemma":0.0002269843,"threshold_uncertainty_score":0.9998405},"labels":[],"label_agreement":null},{"id":"W1988492505","doi":"10.1145/974043.974074","title":"Computing the performability of layered distributed systems with a management architecture","year":2004,"lang":"en","type":"article","venue":"ACM SIGSOFT Software Engineering Notes","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"","keywords":"Computer science; Distributed computing; Queueing theory; Fault tree analysis; Reliability (semiconductor); Fault management; Computation; Architecture; Software; Reliability engineering; Computer network; Operating system; Engineering; Algorithm","score_opus":0.0094764955577076,"score_gpt":0.22082208768235145,"score_spread":0.21134559212464385,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1988492505","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.26635143,0.00023003532,0.7323162,0.00036640305,0.00007080759,0.00028304596,0.000012553488,0.00036832385,0.0000012026808],"genre_scores_gemma":[0.91762906,0.000009537049,0.082247704,0.00001784837,0.000036717,0.000026212612,0.000013809825,0.000016201682,0.000002897533],"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","domain_scores_codex":[0.99799085,0.000057790327,0.00037784412,0.00045133245,0.0006524606,0.0004697505],"domain_scores_gemma":[0.98503196,0.012920124,0.000112412505,0.0016307797,0.0001907338,0.00011398481],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.000666784,0.00025041567,0.00033698403,0.00013894738,0.00018954156,0.0001548605,0.0018279143,0.000073853975,0.0000025359236],"category_scores_gemma":[0.012607849,0.00016273026,0.0001334924,0.0011224977,0.00010101947,0.00017985578,0.0006838117,0.00036666065,0.0000060302914],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012921099,0.00009004858,0.06878694,0.00053545553,0.00015247338,0.000017369004,0.00062726747,0.92045486,0.00006731641,0.001792705,0.000007526929,0.0074551366],"study_design_scores_gemma":[0.0041437447,0.0009870626,0.86039335,0.0023009726,0.00023299197,0.0002811905,0.0004302433,0.11878167,0.00681049,0.0014137763,0.0020291517,0.0021953364],"about_ca_topic_score_codex":0.00014078444,"about_ca_topic_score_gemma":0.0000048291504,"teacher_disagreement_score":0.8016732,"about_ca_system_score_codex":0.00011849814,"about_ca_system_score_gemma":0.00007285324,"threshold_uncertainty_score":0.99570936},"labels":[],"label_agreement":null},{"id":"W1999727136","doi":"10.1021/ci100204x","title":"Reducing Docking Score Variations Arising from Input Differences","year":2010,"lang":"en","type":"article","venue":"Journal of Chemical Information and Modeling","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University Health Network","funders":"","keywords":"Docking (animal); Computer science; Biological system; Chemistry; Biology; Medicine","score_opus":0.035163012605213456,"score_gpt":0.28686995113966757,"score_spread":0.25170693853445414,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1999727136","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.50770646,0.000018281722,0.49176267,0.00037452907,0.00007346825,0.00001194996,3.7913284e-7,0.000009420648,0.00004284616],"genre_scores_gemma":[0.9289541,0.000032699256,0.07080998,0.00009684044,0.00010233129,4.63304e-7,0.0000017802421,0.0000012649145,4.994124e-7],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.998986,0.000016270707,0.00050121895,0.000066333494,0.00031982086,0.000110336965],"domain_scores_gemma":[0.999092,0.00015821254,0.00020920378,0.00012310564,0.00031023944,0.000107249696],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00048377507,0.00006102544,0.00014144537,0.00014879637,0.00009239942,0.0004102787,0.00028342954,0.00006903201,0.000008812668],"category_scores_gemma":[0.00052125886,0.000047619324,0.000062058156,0.00016771002,0.000022380054,0.002141471,0.00009813446,0.0004427927,0.0000018133677],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000043897417,0.00014913791,0.0069432408,0.00009740863,0.0001903672,0.0000046016644,0.01686595,0.055593748,0.20001267,0.020065024,0.00013948901,0.6998945],"study_design_scores_gemma":[0.0001491395,0.0000075890152,0.00010339789,0.000048208105,0.000007764189,0.000014131383,0.000054469645,0.9858262,0.004180923,0.0094877,0.000062079904,0.00005837788],"about_ca_topic_score_codex":0.000041774205,"about_ca_topic_score_gemma":8.0514656e-7,"teacher_disagreement_score":0.93023247,"about_ca_system_score_codex":0.000017170209,"about_ca_system_score_gemma":0.000086684835,"threshold_uncertainty_score":0.39563268},"labels":[],"label_agreement":null},{"id":"W200607821","doi":"","title":"Source Code Verification Tools for Software Security Bugs","year":2006,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Defence Research and Development Canada","funders":"","keywords":"Computer science; Java; Static program analysis; Source code; Software security assurance; Context (archaeology); Secure coding; Software bug; Software engineering; Software; Programming language; Code (set theory); Static analysis; Software development; Computer security; Information security; Set (abstract data type)","score_opus":0.023129832370869866,"score_gpt":0.27891172607299575,"score_spread":0.2557818937021259,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W200607821","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016678445,0.00005849552,0.9808583,0.0011725605,0.000041775376,0.00028869553,0.000007142364,0.00037723398,0.0005173539],"genre_scores_gemma":[0.8462692,0.000011020163,0.15032832,0.00024380715,0.00014387353,0.00016793954,0.000049209622,0.000010635079,0.0027759739],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99879867,0.000047660295,0.00020129808,0.00039243896,0.0002946176,0.00026533572],"domain_scores_gemma":[0.9984678,0.00063330436,0.00004209094,0.00058878877,0.00020967005,0.000058338414],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00048097447,0.00008653228,0.00012729043,0.00006872358,0.0001701118,0.00038385403,0.00065568683,0.0000654598,0.000033385706],"category_scores_gemma":[0.0004457458,0.00007618353,0.00012233426,0.0003791192,0.000043028762,0.0005413701,0.00011005227,0.00008100753,0.00007248901],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003772744,0.00079087855,0.018332748,0.00021681427,0.00006739342,0.00000257617,0.0006134024,0.0036272947,0.001909674,0.2300088,0.046980884,0.69741184],"study_design_scores_gemma":[0.00081302505,0.00014061412,0.008965048,0.000018072576,0.000018743993,0.0000061890055,0.00007310451,0.37685323,0.017155912,0.11968822,0.47571534,0.00055250135],"about_ca_topic_score_codex":0.00016647833,"about_ca_topic_score_gemma":0.00005459482,"teacher_disagreement_score":0.83053,"about_ca_system_score_codex":0.00004782751,"about_ca_system_score_gemma":0.00006245274,"threshold_uncertainty_score":0.37015134},"labels":[],"label_agreement":null},{"id":"W2013542440","doi":"10.1109/compsacw.2010.20","title":"The Level of Decomposition Impact on Component Fault Tolerance","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Reliability engineering; Merge (version control); Software fault tolerance; Fault tolerance; Computer science; Software quality; Component (thermodynamics); Reliability (semiconductor); Component-based software engineering; Software; Software system; Distributed computing; Software development; Engineering; Parallel computing; Programming language","score_opus":0.041148960368516216,"score_gpt":0.35882143285384,"score_spread":0.31767247248532376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2013542440","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7978521,0.000015782953,0.19922684,0.0016438655,0.00012072251,0.000083391526,0.000003241395,0.000042282943,0.0010118],"genre_scores_gemma":[0.99372303,0.000012138281,0.006028273,0.000062912,0.000023883413,0.000004419129,0.0000010971,0.0000020903062,0.00014217595],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9990526,0.0000602361,0.00016380736,0.00017489739,0.00037426475,0.00017420907],"domain_scores_gemma":[0.9985759,0.00057465036,0.000043669897,0.00060601794,0.00013537334,0.000064359265],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005094734,0.00006883394,0.00010523891,0.00004623864,0.00016151613,0.00009955453,0.0007585097,0.000034442473,0.000032614545],"category_scores_gemma":[0.00007181501,0.00003587504,0.00012405767,0.00025521364,0.000081451915,0.000119101824,0.00010021674,0.00019915837,0.000050181497],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014333878,0.0010858781,0.014886059,0.000033217166,0.00016375646,0.0000075949906,0.0005611992,0.011980361,0.2256089,0.19575384,0.005126489,0.54464936],"study_design_scores_gemma":[0.0004220636,0.00031298643,0.23830694,0.00002150267,0.0000051558873,0.000010661103,0.00001832807,0.62806714,0.12168564,0.008389737,0.0025466178,0.00021323135],"about_ca_topic_score_codex":0.0002184589,"about_ca_topic_score_gemma":0.00007326815,"teacher_disagreement_score":0.6160868,"about_ca_system_score_codex":0.00001817304,"about_ca_system_score_gemma":0.00004353447,"threshold_uncertainty_score":0.14629418},"labels":[],"label_agreement":null},{"id":"W2016492784","doi":"10.1007/s10270-009-0128-1","title":"A dependability profile within MARTE","year":2009,"lang":"en","type":"article","venue":"Software & Systems Modeling","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":172,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Natural Sciences and Engineering Research Council of Canada; Ministerio de Ciencia y Tecnología","keywords":"Dependability; Computer science; Unified Modeling Language; Petri net; Stochastic Petri net; Applications of UML; Formal methods; Software engineering; Software; Reliability engineering; Distributed computing; Programming language","score_opus":0.02643716325503444,"score_gpt":0.2744401544358737,"score_spread":0.24800299118083927,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2016492784","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.078954086,0.0006449867,0.9180558,0.00032814936,0.00032754996,0.00045827715,0.0000057201273,0.0010157484,0.00020970644],"genre_scores_gemma":[0.94455844,0.000009050307,0.05488383,0.00011139048,0.00012788584,0.00004860728,0.000007737059,0.000013782712,0.00023926876],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9963632,0.0003199497,0.0007494416,0.0009535309,0.0009687091,0.00064512755],"domain_scores_gemma":[0.99749315,0.00020790109,0.00013506996,0.0014806679,0.00041143867,0.0002717516],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002154403,0.00026846523,0.0004647229,0.00020164599,0.0003478533,0.00052838185,0.0013202094,0.00016559694,0.000012955389],"category_scores_gemma":[0.0007311554,0.00023271456,0.0002266049,0.00084291585,0.00003872625,0.00079263066,0.00021550247,0.00041134687,0.00014805801],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000059023452,0.0006374104,0.014148507,0.0005645726,0.00014805533,0.000106563195,0.0032696868,0.8334596,0.0007032095,0.019901304,0.00076038373,0.1262417],"study_design_scores_gemma":[0.00020296934,0.00008812821,0.00011864308,0.000106517495,0.000009439408,0.00002699506,0.000085377374,0.99210995,0.00007933588,0.0067208884,0.00016345369,0.00028829733],"about_ca_topic_score_codex":0.00033781445,"about_ca_topic_score_gemma":0.000011858889,"teacher_disagreement_score":0.86560434,"about_ca_system_score_codex":0.00018664889,"about_ca_system_score_gemma":0.00027851362,"threshold_uncertainty_score":0.9489825},"labels":[],"label_agreement":null},{"id":"W2017813642","doi":"10.1109/eit.2006.252187","title":"Analysis of Service-Oriented Architectures with Sensitivity Analysis","year":2006,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Windsor","funders":"","keywords":"Computer science; Software engineering; Software construction; Software development; Software system; Software; Software sizing; Service (business); Software design; Sensitivity (control systems); Verification and validation; Software reliability testing; Systems engineering; Reliability engineering; Engineering; Operating system","score_opus":0.006828420345467531,"score_gpt":0.24188692030318112,"score_spread":0.2350584999577136,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2017813642","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.44256365,0.000012315533,0.5562301,0.00044808324,0.0000038336034,0.00004132483,0.000006153675,0.000079422694,0.0006150877],"genre_scores_gemma":[0.97898805,0.0000011193134,0.02067676,0.00012913869,0.000011255993,0.0000040579266,0.000029819012,0.0000036219446,0.00015615582],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99782616,0.00024685226,0.0003144941,0.00054755626,0.0007671748,0.000297762],"domain_scores_gemma":[0.99778044,0.0004719385,0.00011485574,0.001051421,0.00049520365,0.00008614915],"candidate_categories":["bibliometrics"],"consensus_categories":[],"category_scores_codex":[0.00066748797,0.00014839885,0.0005940981,0.001759512,0.00008829321,0.00007115775,0.0004034375,0.000052070878,0.000079317564],"category_scores_gemma":[0.00004397008,0.000102709935,0.00046552165,0.02373402,0.000085832326,0.00010669157,0.00017814928,0.000119404955,0.0000059726076],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017391194,0.0002221796,0.52917147,0.000015249651,0.0068010846,0.000017224284,0.00014372931,0.4564432,0.00043046672,0.004241735,0.000016269074,0.0024799982],"study_design_scores_gemma":[0.00009237938,0.000024535115,0.39493835,0.0000018912943,0.0018967148,9.74315e-7,0.000016458353,0.6012182,0.0014754736,0.00018595568,0.000040114024,0.00010893992],"about_ca_topic_score_codex":0.011545853,"about_ca_topic_score_gemma":0.029265603,"teacher_disagreement_score":0.53642446,"about_ca_system_score_codex":0.000029700008,"about_ca_system_score_gemma":0.000060943094,"threshold_uncertainty_score":0.99701697},"labels":[],"label_agreement":null},{"id":"W2019243921","doi":"10.1002/nav.20194","title":"On optimality of one‐bug‐look‐ahead policies for a software testing model","year":2007,"lang":"en","type":"article","venue":"Naval Research Logistics (NRL)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of New Brunswick","funders":"","keywords":"Counterexample; Monotonic function; Software; Computer science; Sample (material); Sense (electronics); Mathematical optimization; Mathematical economics; Mathematics; Discrete mathematics; Programming language","score_opus":0.36816490180597294,"score_gpt":0.4710107124231524,"score_spread":0.10284581061717946,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019243921","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03127173,0.000115350675,0.9647072,0.0005220106,0.000053164455,0.00060246146,0.000070822534,0.00017404363,0.0024831994],"genre_scores_gemma":[0.71848434,0.000023610115,0.2808486,0.00007270461,0.000101278834,0.000040988038,0.000009149521,0.000021122876,0.00039819043],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9937728,0.00030949924,0.0007272652,0.0008534282,0.0027727615,0.0015642459],"domain_scores_gemma":[0.97656965,0.017908363,0.0001504629,0.0014500136,0.0034708625,0.00045064779],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.012903073,0.00023839175,0.0004990114,0.0007513315,0.0005636523,0.00024348048,0.0022069553,0.0002435421,0.000009711425],"category_scores_gemma":[0.06462536,0.00021905218,0.00023136594,0.0021396372,0.0008984037,0.00021430873,0.0009833343,0.0009646769,0.00003926756],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0008660515,0.003078399,0.012865019,0.0015139162,0.00025753397,0.000075942124,0.001422727,0.13596028,0.008689629,0.64567375,0.003430839,0.18616594],"study_design_scores_gemma":[0.00063960947,0.0012457012,0.0025918693,0.00016232055,0.000015548143,0.0000036875133,0.00008561203,0.7177476,0.008019355,0.26892677,0.00022948817,0.0003324308],"about_ca_topic_score_codex":0.000812068,"about_ca_topic_score_gemma":0.000089252266,"teacher_disagreement_score":0.68721265,"about_ca_system_score_codex":0.00036617467,"about_ca_system_score_gemma":0.00092662446,"threshold_uncertainty_score":0.9432537},"labels":[],"label_agreement":null},{"id":"W2019817087","doi":"10.1177/01466216045280142","title":"SIMCAT 1.0: A SAS Computer Program for Simulating Computer Adaptive Testing","year":2006,"lang":"en","type":"article","venue":"Applied Psychological Measurement","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; Université du Québec à Montréal","funders":"","keywords":"Computerized adaptive testing; Computer science; Computer program; Programming language; Statistics; Psychometrics; Mathematics","score_opus":0.13975375466909867,"score_gpt":0.3452510035012429,"score_spread":0.2054972488321442,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2019817087","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013610403,0.00004292583,0.98130655,0.00029482672,0.00016117349,0.001588768,0.0000017954849,0.00072425016,0.0022692985],"genre_scores_gemma":[0.54592633,3.8834847e-7,0.45310724,0.0002808103,0.00033609295,0.0003301705,0.0000036490073,0.000009504896,0.0000058099877],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99633354,0.00012463718,0.00056951575,0.0011281433,0.0010945606,0.0007496279],"domain_scores_gemma":[0.997682,0.0007764028,0.00015757441,0.0006643397,0.00056333654,0.00015637297],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017032363,0.00029835204,0.00039808362,0.000107365064,0.0003276842,0.00030422083,0.00094036676,0.00015279495,0.0000104249575],"category_scores_gemma":[0.00008879098,0.00023440535,0.00020321355,0.0007968162,0.00011826622,0.000104488056,0.0002649583,0.00029008716,0.000056438512],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000042322525,0.0013087069,0.0008178631,0.000026920443,0.000051608462,0.0000056573144,0.000041427902,0.0138803655,0.0007395647,0.013170947,0.0013898385,0.96852475],"study_design_scores_gemma":[0.0018368769,0.0021524641,0.020721387,0.000062103754,0.000028292738,0.0000070773804,0.000010182962,0.9399897,0.00038301543,0.02971858,0.004445454,0.0006448523],"about_ca_topic_score_codex":0.000023848468,"about_ca_topic_score_gemma":0.000005192936,"teacher_disagreement_score":0.96787995,"about_ca_system_score_codex":0.00015392368,"about_ca_system_score_gemma":0.000033254943,"threshold_uncertainty_score":0.95587724},"labels":[],"label_agreement":null},{"id":"W2022974228","doi":"10.4271/2013-01-2258","title":"An Approach of Developing System Redundancy Management Requirements","year":2013,"lang":"en","type":"article","venue":"SAE technical papers on CD-ROM/SAE technical paper series","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bombardier (Canada)","funders":"","keywords":"Redundancy (engineering); Computer science; Reliability engineering; Engineering; Operating system","score_opus":0.025053764561846433,"score_gpt":0.2836842820569973,"score_spread":0.25863051749515087,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2022974228","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.676445,0.0011839096,0.006859552,0.010750972,0.0008484208,0.00918613,0.000050093215,0.015445226,0.27923068],"genre_scores_gemma":[0.87303275,0.00016171289,0.12548606,0.0003951099,0.000052553958,0.0005741973,0.000021209611,0.000050645955,0.00022576979],"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","domain_scores_codex":[0.99338055,0.00039759476,0.0014619014,0.0017351265,0.0019164792,0.0011083415],"domain_scores_gemma":[0.9954184,0.0002935563,0.00031693256,0.0031823567,0.00032565306,0.00046306892],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0015198368,0.00062525406,0.0010080906,0.00043318313,0.000441,0.00033845188,0.003966308,0.0005152217,0.00010907021],"category_scores_gemma":[0.00028866026,0.00052048557,0.00045095652,0.002007288,0.0007695009,0.0014892442,0.0013574773,0.00078767835,0.00017102993],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005580012,0.00082156714,0.00049433176,0.0006044312,0.00015177317,0.000033518307,0.000055956312,0.000048788064,0.71131605,0.25370076,0.0009015871,0.03181543],"study_design_scores_gemma":[0.0005250438,0.0007131947,0.9882105,0.00046752195,0.000051708892,0.000043055355,0.00027644433,0.000008682512,0.00022086655,0.0049105347,0.0038816573,0.0006907877],"about_ca_topic_score_codex":0.00010653625,"about_ca_topic_score_gemma":0.0014086313,"teacher_disagreement_score":0.9877162,"about_ca_system_score_codex":0.0005537772,"about_ca_system_score_gemma":0.00013492876,"threshold_uncertainty_score":0.9997247},"labels":[],"label_agreement":null},{"id":"W2023156601","doi":"10.5555/2820126.2820139","title":"A methodology for the simplification of tabular designs in model-based development","year":2015,"lang":"en","type":"article","venue":"MacSphere (McMaster University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Automotive industry; Code refactoring; Software engineering; Traceability; Readability; Programming language; Software; Engineering","score_opus":0.18123033687229742,"score_gpt":0.30929783022177126,"score_spread":0.12806749334947384,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023156601","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019451771,0.00004449872,0.9956762,0.0005630508,0.000019976345,0.00022598806,0.0000010919338,0.000023616516,0.0015004094],"genre_scores_gemma":[0.38291138,0.0000047694025,0.60927063,0.00009744047,0.000008264525,0.0000075569073,0.000003173877,0.0000067455553,0.0076900194],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99898666,0.00024330347,0.0001508893,0.00024631317,0.00018767672,0.00018518555],"domain_scores_gemma":[0.9987089,0.0006499196,0.000064801716,0.00032538394,0.00018590079,0.0000651126],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010259972,0.00007544962,0.00014900554,0.00014666546,0.00006188657,0.000021836318,0.0008273411,0.00005761763,0.00015153353],"category_scores_gemma":[0.0001556107,0.000062443476,0.00005875523,0.0007564307,0.00006568475,0.00014365793,0.00014834899,0.00007729162,0.000004222383],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013927069,0.00014597333,0.0042636385,0.000049783874,0.000052289906,0.000005308768,0.0017044605,0.25902554,0.00055227324,0.01699414,0.00014139428,0.7169259],"study_design_scores_gemma":[0.0007969938,0.000056908175,0.00039785163,0.000010789061,0.000013823317,2.9465673e-7,0.00046598827,0.9525619,0.0042088735,0.0022809268,0.039101902,0.000103711915],"about_ca_topic_score_codex":0.000051023017,"about_ca_topic_score_gemma":0.00013673255,"teacher_disagreement_score":0.7168222,"about_ca_system_score_codex":0.00013066175,"about_ca_system_score_gemma":0.00040305094,"threshold_uncertainty_score":0.2546371},"labels":[],"label_agreement":null},{"id":"W2023398194","doi":"10.3166/jesa.44.229-253","title":"Modélisation d'un système par automate stochastique hybride pour l'évaluation de la fiabilité dynamique","year":2010,"lang":"fr","type":"article","venue":"Journal Européen des Systèmes Automatisés","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Humanities; Physics; Philosophy","score_opus":0.021512256082259634,"score_gpt":0.29345530394523334,"score_spread":0.2719430478629737,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2023398194","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3922377,0.0016990012,0.60029346,0.0029594682,0.0012033253,0.0004298771,0.000019236006,0.0004674096,0.0006905602],"genre_scores_gemma":[0.85285264,0.00048192585,0.1438127,0.00015585704,0.0009348696,0.00006077172,0.000010346114,0.000101786456,0.0015891066],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9867488,0.0071711056,0.0018359824,0.000985956,0.0017808612,0.001477299],"domain_scores_gemma":[0.9915791,0.0034997927,0.0010813033,0.001304775,0.0015620376,0.0009729559],"candidate_categories":["metaresearch","metaepi_narrow","sts","scholarly_communication","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.015182645,0.00070710405,0.0009389117,0.0007154267,0.0013467902,0.0033005436,0.0018983859,0.0006697714,0.00067446753],"category_scores_gemma":[0.008897348,0.000688574,0.0007033186,0.0014524155,0.0009900765,0.0023839164,0.00040125105,0.0028204843,0.0005144477],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029563631,0.00084227946,0.011705346,0.0010628318,0.0005035811,0.00080114475,0.005214798,0.052472945,0.017704895,0.042293474,0.0010127921,0.8663564],"study_design_scores_gemma":[0.0005285401,0.00015992552,0.262521,0.0006114712,0.000118252065,0.0050334786,0.000065676,0.6248815,0.0013539637,0.10393295,0.00036138287,0.0004318816],"about_ca_topic_score_codex":0.0005155189,"about_ca_topic_score_gemma":0.00008127734,"teacher_disagreement_score":0.8659245,"about_ca_system_score_codex":0.0013849703,"about_ca_system_score_gemma":0.0030037586,"threshold_uncertainty_score":0.9999533},"labels":[],"label_agreement":null},{"id":"W2024062765","doi":"10.1109/prdc.2013.33","title":"Derivation of Stochastic Reward Net for Compatibility and Conformance Verification of Component Erroneous Behavior Model","year":2013,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Center for Evolutionary Biology and Medicine, University of Pittsburgh; Albaha University","keywords":"Compatibility (geochemistry); Computer science; Component (thermodynamics); Unified Modeling Language; Conformance testing; Model checking; Behavioral modeling; Reliability engineering; Distributed computing; Theoretical computer science; Programming language; Artificial intelligence; Engineering; Software","score_opus":0.029971621044234415,"score_gpt":0.2823924052743948,"score_spread":0.2524207842301604,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2024062765","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.46505633,0.000012669994,0.5345069,0.000078594436,0.0000073103797,0.00031143837,0.0000030721776,0.0000142047,0.0000094575635],"genre_scores_gemma":[0.9442963,0.000004078239,0.055581078,0.000012900408,0.000002548906,0.00007291721,0.000009943266,0.000002173389,0.000018082055],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9991249,0.000025390505,0.00032736443,0.00019659086,0.000213073,0.00011268821],"domain_scores_gemma":[0.9988652,0.00018633611,0.0001240911,0.00038075433,0.0003966002,0.00004704468],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002871446,0.00006209587,0.00017835925,0.00006575336,0.0000413727,0.0000242793,0.00024828338,0.000035519504,0.000007904658],"category_scores_gemma":[0.000092092494,0.000053587748,0.000041551546,0.00014001664,0.00010945199,0.0002665002,0.00007546475,0.000039343307,0.0000018034318],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020845745,0.003096373,0.05236676,0.0019984827,0.00018240094,3.4960294e-7,0.005746029,0.3115979,0.18319821,0.09819972,0.0008601604,0.34254515],"study_design_scores_gemma":[0.00018309648,0.00008177335,0.033934876,0.0000086053205,0.000008064792,4.997943e-7,0.000023581688,0.96025354,0.0034257476,0.0020257302,0.0000020779646,0.00005241808],"about_ca_topic_score_codex":0.0002878375,"about_ca_topic_score_gemma":0.0000103163065,"teacher_disagreement_score":0.6486556,"about_ca_system_score_codex":0.00002037724,"about_ca_system_score_gemma":0.000045222034,"threshold_uncertainty_score":0.2185245},"labels":[],"label_agreement":null},{"id":"W2024720842","doi":"10.1016/j.orl.2012.05.008","title":"Simultaneous testing of multicharacteristic components","year":2012,"lang":"en","type":"article","venue":"Operations Research Letters","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Component (thermodynamics); Computer science; Construct (python library); Simple (philosophy); Heuristic; Test strategy; Sequential analysis; Algorithm; Mathematical optimization; Mathematics; Statistics; Artificial intelligence","score_opus":0.10618553926644932,"score_gpt":0.37003103829978545,"score_spread":0.26384549903333615,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2024720842","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.85505885,0.00006129416,0.14238709,0.0019532803,0.000073917814,0.00022660148,0.0000053697363,0.000055966368,0.000177614],"genre_scores_gemma":[0.9553502,0.000006421039,0.044317357,0.00012521159,0.000082282284,0.000030963198,0.000008758767,0.000007324094,0.000071487455],"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9973523,0.0005479119,0.00028965482,0.00025152095,0.00093669526,0.00062188605],"domain_scores_gemma":[0.9962461,0.002298225,0.000020080846,0.00064227625,0.00058210146,0.00021118754],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0019901562,0.00008902284,0.00015967782,0.00034524073,0.00040708945,0.00019766267,0.00083721575,0.00003842771,0.000043187596],"category_scores_gemma":[0.0048187887,0.00007946531,0.000053675896,0.0013281931,0.0002167371,0.00062979705,0.00037204518,0.00035086225,0.0002066102],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011436659,0.0008351174,0.03553327,0.000094891366,0.000079425066,0.000030332007,0.002110082,0.045182824,0.87228197,0.0031051077,0.00052580796,0.040209714],"study_design_scores_gemma":[0.00020059673,0.000057778416,0.0123884175,0.000038104223,0.00000427443,0.000010430554,0.000050508075,0.97715837,0.008598399,0.00002650124,0.0013147515,0.00015189104],"about_ca_topic_score_codex":0.00042359674,"about_ca_topic_score_gemma":0.000008586074,"teacher_disagreement_score":0.93197554,"about_ca_system_score_codex":0.00009568837,"about_ca_system_score_gemma":0.00007325906,"threshold_uncertainty_score":0.5768888},"labels":[],"label_agreement":null},{"id":"W2027958489","doi":"10.1109/ares.2014.54","title":"Analysis of Algorithms for Computation of Direct Partial Logic Derivatives in Multiple-Valued Decision Diagrams","year":2014,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Ministère de la Santé et des Services sociaux","keywords":"Computer science; Reliability (semiconductor); Algorithm; Dimension (graph theory); Representation (politics); Component (thermodynamics); Binary decision diagram; Theoretical computer science; Influence diagram; Mathematics; Data mining; Decision tree","score_opus":0.055423174324658174,"score_gpt":0.3523291786181512,"score_spread":0.29690600429349306,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2027958489","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.25973105,0.000021062133,0.740001,0.00004349117,0.00002017792,0.00013141969,0.000002353778,0.000021516014,0.000027892554],"genre_scores_gemma":[0.83081454,0.000007493489,0.16912502,0.000013512884,0.0000074736126,0.000015785447,0.000008925334,0.0000023613225,0.00000490436],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.998373,0.00022891963,0.0004874303,0.00034515484,0.00038460246,0.00018094578],"domain_scores_gemma":[0.996189,0.0030198526,0.00014286874,0.00031734526,0.00028226152,0.000048681173],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013813685,0.000088294255,0.00046996016,0.0005466112,0.000038527538,0.00002877807,0.00041007414,0.00005467004,0.0000066731013],"category_scores_gemma":[0.0020513055,0.00006821745,0.0002591803,0.0023360648,0.00008678614,0.00016388069,0.00011939563,0.00004767935,0.0000010857331],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000065553424,0.000576576,0.21993083,0.000057523048,0.00041176,4.6607917e-7,0.00087700615,0.2088422,0.00119904,0.0053159366,0.0000138796095,0.5627092],"study_design_scores_gemma":[0.0003922769,0.00012690002,0.08041564,0.000013043382,0.00004361119,4.147226e-8,0.000026692232,0.91108847,0.004649326,0.0031659538,0.000009913377,0.00006815],"about_ca_topic_score_codex":0.00023654518,"about_ca_topic_score_gemma":0.00016482297,"teacher_disagreement_score":0.70224625,"about_ca_system_score_codex":0.000022784037,"about_ca_system_score_gemma":0.000026655685,"threshold_uncertainty_score":0.2781827},"labels":[],"label_agreement":null},{"id":"W2028198237","doi":"10.1007/s11633-007-0388-6","title":"Software operational profile based test case allocation using fuzzy logic","year":2007,"lang":"en","type":"article","venue":"International Journal of Automation and Computing","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":13,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Indian Institute of Technology Kharagpur; University of Ottawa","keywords":"Computer science; Software; Software reliability testing; Fuzzy logic; Software quality; Reliability engineering; Software sizing; Test case; Software construction; Data mining; Software system; Software development; Artificial intelligence; Machine learning; Engineering; Operating system; Regression analysis","score_opus":0.027966603596039166,"score_gpt":0.3372344780745888,"score_spread":0.30926787447854964,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2028198237","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21077159,0.000055315053,0.78790534,0.0008776662,0.00023432571,0.000049545026,0.000001437295,0.000042853375,0.00006194888],"genre_scores_gemma":[0.7448903,0.0000025696725,0.2546413,0.0002773804,0.00017326105,2.5779892e-7,0.0000037948141,0.000003444142,0.000007722151],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9983686,0.00006759505,0.0005684296,0.00016593332,0.0006835718,0.00014582036],"domain_scores_gemma":[0.99742067,0.0006912476,0.00035553097,0.00010564164,0.0013318041,0.00009511745],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017128568,0.00009178618,0.00012992758,0.00039620567,0.00017824974,0.00033825738,0.00039948634,0.000051414572,0.000022799853],"category_scores_gemma":[0.00088127365,0.000082038554,0.00007888665,0.00029109244,0.000040925002,0.0005997702,0.00011167638,0.00016617391,0.0000045004704],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000055290417,0.0007337183,0.048148066,0.00007159104,0.00024659708,0.0023427147,0.0015512367,0.08949168,0.0063049486,0.018913953,0.00026114564,0.8318791],"study_design_scores_gemma":[0.00044182208,0.00006788807,0.008355486,0.000072466086,0.000007937476,0.0028855845,0.00008277823,0.98556393,0.0012290053,0.0010765449,0.00011677757,0.00009975778],"about_ca_topic_score_codex":0.000024504627,"about_ca_topic_score_gemma":0.000004292184,"teacher_disagreement_score":0.89607227,"about_ca_system_score_codex":0.00014326302,"about_ca_system_score_gemma":0.00020543666,"threshold_uncertainty_score":0.33454353},"labels":[],"label_agreement":null},{"id":"W2031392029","doi":"10.1109/issre.2008.22","title":"A Hardware-Assisted Tool for Fast, Full Code Coverage Analysis","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta; University of Calgary","funders":"","keywords":"Computer science; Instrumentation (computer programming); Debugging; Code (set theory); Code coverage; Process (computing); Unreachable code; Embedded system; Software; Dead code; Redundant code; Source code; Operating system; Computer hardware; Code generation; Programming language; Key (lock)","score_opus":0.03624551820477032,"score_gpt":0.2944457366308323,"score_spread":0.25820021842606194,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2031392029","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.056656405,0.000045809615,0.94105107,0.00080075715,0.000039548682,0.00020399458,0.000021250418,0.00025205652,0.0009290975],"genre_scores_gemma":[0.9078276,0.00006454882,0.08426723,0.00041000242,0.000048555543,0.00008735738,0.000030286612,0.000009208492,0.00725517],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99796927,0.00009659404,0.00033691726,0.0006166947,0.0005570419,0.00042351178],"domain_scores_gemma":[0.99796134,0.00055691454,0.00006365166,0.00095297047,0.0003231291,0.00014197879],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005050823,0.00014987269,0.00038795706,0.00037678686,0.000327973,0.00017052058,0.00097133423,0.00007848624,0.00018584233],"category_scores_gemma":[0.00030167092,0.0001242818,0.0006106183,0.002449887,0.00007511302,0.00037330092,0.00022579146,0.000116025716,0.00009348968],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0004531108,0.0036001983,0.31190512,0.00034067797,0.011690083,0.0004365406,0.0028567282,0.02872086,0.006062968,0.052897595,0.077480346,0.5035558],"study_design_scores_gemma":[0.0012152588,0.0002870407,0.07016136,0.000008074656,0.00029647475,0.000041123294,0.00002407211,0.88643503,0.0014497514,0.001579705,0.037867837,0.00063426094],"about_ca_topic_score_codex":0.00007755318,"about_ca_topic_score_gemma":0.00010452802,"teacher_disagreement_score":0.8577142,"about_ca_system_score_codex":0.00007326899,"about_ca_system_score_gemma":0.00014638882,"threshold_uncertainty_score":0.50680643},"labels":[],"label_agreement":null},{"id":"W2032494193","doi":"10.1145/1142958.1142961","title":"How should software reliability engineering (SRE) be taught?","year":2006,"lang":"en","type":"article","venue":"ACM SIGSOFT Software Engineering Notes","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Nortel (Canada)","funders":"","keywords":"Reliability (semiconductor); Software quality; Software engineering; Computer science; Software; Engineering management; Software peer review; Work (physics); Software development; Software Engineering Process Group; Medical education; Engineering; Software construction; Medicine; Mechanical engineering","score_opus":0.022359914709258053,"score_gpt":0.24394602014958858,"score_spread":0.22158610544033053,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2032494193","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1340432,0.0011299514,0.8549733,0.0040153335,0.00062686286,0.00043532535,0.000046324905,0.004726654,0.0000030401245],"genre_scores_gemma":[0.6758006,0.000032408647,0.32312948,0.00013194385,0.0003852605,0.00010367284,0.000063990185,0.00009640908,0.0002562452],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9947837,0.00008744395,0.00072213204,0.0015273872,0.0014092348,0.0014700905],"domain_scores_gemma":[0.94780105,0.047039833,0.00018172746,0.0039359275,0.00055308297,0.00048839144],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0010036989,0.00079647446,0.0008157827,0.00057472347,0.00030215198,0.0010089158,0.0033628382,0.0004441614,0.000039604158],"category_scores_gemma":[0.17929663,0.0007909635,0.0005546021,0.0020205777,0.00010714431,0.0015395948,0.0013581833,0.0010231177,0.000052383715],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032138018,0.0007818666,0.7420036,0.0015967463,0.00030391762,0.00031550732,0.0005443093,0.19984312,0.007069831,0.006570197,0.006170664,0.034768075],"study_design_scores_gemma":[0.0035920092,0.0007319038,0.57406294,0.0010240587,0.0003227244,0.0002768882,0.000055350407,0.11457287,0.070698604,0.0038259525,0.22187935,0.008957349],"about_ca_topic_score_codex":0.00029327345,"about_ca_topic_score_gemma":0.00001745902,"teacher_disagreement_score":0.5417574,"about_ca_system_score_codex":0.0003429945,"about_ca_system_score_gemma":0.00017164105,"threshold_uncertainty_score":0.99945414},"labels":[],"label_agreement":null},{"id":"W2033787847","doi":"10.1109/ccece.2010.5575155","title":"Condition-based reliability modeling for systems with partial and standby redundancy","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; University of Alberta","funders":"","keywords":"Reliability engineering; Redundancy (engineering); Reliability (semiconductor); Computer science; Reliability theory; Maintenance engineering; Failure rate; Engineering","score_opus":0.015950701620490604,"score_gpt":0.28214020702261566,"score_spread":0.26618950540212505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2033787847","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20080404,0.000023570521,0.79789835,0.00066803244,0.00008064156,0.00029531174,0.000007327504,0.00012117983,0.000101553174],"genre_scores_gemma":[0.9489054,0.0000022231939,0.05082616,0.000041151146,0.000047025405,0.00009627257,0.000006060281,0.0000056662925,0.0000700198],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99870926,0.00005258937,0.00021156261,0.00045820556,0.0003142641,0.00025412955],"domain_scores_gemma":[0.99845964,0.00045134843,0.00003703931,0.0005688504,0.00034025917,0.00014283592],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00090842484,0.00010060338,0.0001765767,0.00006378398,0.00021723083,0.0002956462,0.00029353137,0.00006890144,0.000014519668],"category_scores_gemma":[0.00026732273,0.000071092094,0.00004980851,0.00021453254,0.00010782544,0.00028707206,0.00004758228,0.00016770062,0.0000031544625],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00059275853,0.0012476274,0.08145517,0.0017235737,0.00024102796,0.000020653075,0.00089504884,0.5090075,0.005804117,0.3710015,0.001359769,0.026651222],"study_design_scores_gemma":[0.00043367202,0.00011218728,0.00007467703,0.000014455278,0.00000740786,0.0000023486045,0.000022414779,0.9955158,0.0006051878,0.0023503671,0.0007481811,0.00011326757],"about_ca_topic_score_codex":0.00018714528,"about_ca_topic_score_gemma":0.000053599084,"teacher_disagreement_score":0.7481014,"about_ca_system_score_codex":0.000020219924,"about_ca_system_score_gemma":0.00019093377,"threshold_uncertainty_score":0.28990516},"labels":[],"label_agreement":null},{"id":"W2033944127","doi":"10.1023/a:1004101402897","title":"Start-Up Demonstration Tests with Rejection of Units upon Observing d Failures","year":2000,"lang":"en","type":"article","venue":"Annals of the Institute of Statistical Mathematics","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":54,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Mathematics; Test (biology); Statistics; Function (biology); Statistical hypothesis testing; Distribution (mathematics); Calculus (dental); Applied mathematics; Mathematical analysis; Medicine","score_opus":0.0802316870046062,"score_gpt":0.3248804992497101,"score_spread":0.24464881224510387,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2033944127","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.43600342,0.00004045276,0.56218517,0.0008606757,0.00004893303,0.00019621127,0.00003599914,0.000025207417,0.00060391397],"genre_scores_gemma":[0.86338276,0.000051114435,0.1364639,0.0000195312,0.000008146267,0.000003234399,0.000003123375,0.0000045756988,0.000063605454],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9985033,0.000076022196,0.00049504405,0.00015919631,0.0006092179,0.00015720172],"domain_scores_gemma":[0.99811053,0.0005212724,0.00021940198,0.0005481826,0.00054563856,0.000054995064],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00049757597,0.00010286549,0.0002833352,0.0000690681,0.00007977197,0.000028827319,0.000601722,0.000050121518,0.00003269575],"category_scores_gemma":[0.0010359322,0.000066811015,0.000054678854,0.00078632106,0.00031074486,0.00031854265,0.00008694849,0.000112462774,0.0000032503135],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010418825,0.00093458954,0.0014421251,0.0020353503,0.00027355278,0.000007921571,0.002105382,0.017071214,0.0037302524,0.84529626,0.00060287095,0.12639628],"study_design_scores_gemma":[0.0010158722,0.0014223154,0.017319765,0.0026677107,0.00019900406,0.000050491202,0.00042039505,0.3254997,0.113669254,0.53553474,0.0015755754,0.0006251725],"about_ca_topic_score_codex":0.00018409158,"about_ca_topic_score_gemma":0.00011446866,"teacher_disagreement_score":0.42737937,"about_ca_system_score_codex":0.000008816094,"about_ca_system_score_gemma":0.00017317443,"threshold_uncertainty_score":0.2724474},"labels":[],"label_agreement":null},{"id":"W2037087526","doi":"10.1109/iecon.2012.6389397","title":"Software quality engineering in service of IT consumer","year":2012,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Software quality; Social software engineering; Computer science; Software engineering; Software quality control; Quality (philosophy); Software quality analyst; Software development; Software construction; Software; Operating system","score_opus":0.046317996402359715,"score_gpt":0.3261213490762795,"score_spread":0.27980335267391976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2037087526","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.45821196,0.00022969049,0.53993237,0.00091807183,0.000059413236,0.0000830449,0.0000010000302,0.00012047756,0.0004439714],"genre_scores_gemma":[0.96354246,0.000010417411,0.036164887,0.0001879518,0.000011305244,0.0000050734584,7.0195546e-7,0.0000027253993,0.00007446182],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9990261,0.00006684177,0.0002416304,0.00014136045,0.00026467533,0.00025940192],"domain_scores_gemma":[0.9988926,0.00048806609,0.000028650902,0.00040383378,0.00010109073,0.00008574241],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010244884,0.00006141827,0.00015607747,0.00010977245,0.00001395283,0.000018658615,0.00043806952,0.000042725635,0.000081199614],"category_scores_gemma":[0.0004386282,0.00005334331,0.000046905974,0.0007910317,0.000013907711,0.00037259056,0.00020580206,0.00010275685,0.000059560894],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000046390023,0.00030320793,0.9571558,0.00029173077,0.000035778972,0.0000012723771,0.0016011249,0.0012195868,0.00078639894,0.01766145,0.00029440608,0.02064456],"study_design_scores_gemma":[0.0011490568,0.000049303395,0.8650161,0.00011911222,0.000011733767,0.00000705336,0.00025463515,0.104623795,0.016959246,0.0010163333,0.010112192,0.0006814467],"about_ca_topic_score_codex":0.00064830953,"about_ca_topic_score_gemma":0.000077487624,"teacher_disagreement_score":0.5053305,"about_ca_system_score_codex":0.000026155187,"about_ca_system_score_gemma":0.000034992947,"threshold_uncertainty_score":0.2175277},"labels":[],"label_agreement":null},{"id":"W2041489089","doi":"10.1145/1723028.1723047","title":"A lightweight and flexible process for designing intuitive error handling and effective error messages","year":2009,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"IBM (Canada)","funders":"","keywords":"Computer science; Process (computing); Error detection and correction; Algorithm; Programming language","score_opus":0.029033726844238945,"score_gpt":0.342182483863833,"score_spread":0.31314875701959405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2041489089","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.099851094,0.0005700873,0.8969282,0.0013626427,0.00001962491,0.00061879837,0.0000012596062,0.00017717903,0.0004711502],"genre_scores_gemma":[0.96555287,0.00003246018,0.03384898,0.00018841155,0.00003119436,0.000090135014,0.0000010772416,0.0000055416717,0.00024931267],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99864024,0.00008623089,0.00016649853,0.00055825175,0.00022339121,0.00032539153],"domain_scores_gemma":[0.9985725,0.0008341872,0.00004709334,0.00019370626,0.0002156865,0.00013680047],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007197726,0.00015016485,0.00026150895,0.00017022877,0.0003138324,0.00027797543,0.00026769293,0.000071641705,0.000004409365],"category_scores_gemma":[0.0003947594,0.00011383109,0.000054486412,0.00042706294,0.00008334488,0.00058849144,0.00009808646,0.00012798104,0.0000021009464],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030096,0.0005229445,0.009006461,0.00070889323,0.0003447657,0.00003143199,0.016050411,0.000445813,0.012581212,0.04918161,0.0004173935,0.9104081],"study_design_scores_gemma":[0.002922329,0.0025266604,0.018163366,0.00043155818,0.00010681942,0.000044009426,0.0011951434,0.41494018,0.37390804,0.18405801,0.000642675,0.0010612004],"about_ca_topic_score_codex":0.000010237618,"about_ca_topic_score_gemma":0.0000046387113,"teacher_disagreement_score":0.9093469,"about_ca_system_score_codex":0.000026105217,"about_ca_system_score_gemma":0.000055531735,"threshold_uncertainty_score":0.46418974},"labels":[],"label_agreement":null},{"id":"W2041953547","doi":"10.1142/s0218539305001665","title":"k-OUT-OF-n SYSTEM WITH SELF-LOOP UNITS","year":2005,"lang":"en","type":"article","venue":"International Journal of Reliability Quality and Safety Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Reliability (semiconductor); Loop (graph theory); Standard deviation; Moment-generating function; Reliability engineering; Computer science; Graph; Function (biology); Mathematics; Control theory (sociology); Engineering; Probability density function; Statistics; Combinatorics; Physics; Theoretical computer science; Control (management); Artificial intelligence","score_opus":0.020844364645484083,"score_gpt":0.287848612627805,"score_spread":0.2670042479823209,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2041953547","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.31922558,0.00022532803,0.67783546,0.0021172082,0.00031539187,0.000073387935,0.0000072305875,0.00007056764,0.00012986832],"genre_scores_gemma":[0.9526193,0.00017024305,0.04700228,0.00003083345,0.00015049265,0.0000013351396,0.0000013567047,0.00000613361,0.000018035169],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975781,0.00012812401,0.0008949521,0.00021268224,0.0010143917,0.00017177616],"domain_scores_gemma":[0.9966684,0.0008915275,0.0003148941,0.00031167976,0.0016685049,0.0001450063],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030379635,0.00013215758,0.00035509624,0.00021254206,0.000042755924,0.00008330273,0.00086845143,0.00007732198,0.000004476236],"category_scores_gemma":[0.000658705,0.00010377342,0.00011372031,0.00033039926,0.000056507808,0.0006480165,0.00016203705,0.00034635002,0.0000024494814],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0010013332,0.0011301982,0.049488273,0.0022620733,0.0023217397,0.00013197777,0.008077545,0.54542196,0.0028424729,0.25778165,0.00007647496,0.12946431],"study_design_scores_gemma":[0.005920022,0.0010017,0.08900124,0.0026885045,0.00015126253,0.00075843907,0.0011477561,0.85947114,0.008677947,0.0011825486,0.028705055,0.0012943606],"about_ca_topic_score_codex":0.00003273037,"about_ca_topic_score_gemma":0.000008180116,"teacher_disagreement_score":0.6333937,"about_ca_system_score_codex":0.000230834,"about_ca_system_score_gemma":0.0001886264,"threshold_uncertainty_score":0.42317572},"labels":[],"label_agreement":null},{"id":"W2044138490","doi":"10.1007/s10270-008-0084-1","title":"Model-driven assessment of system dependability","year":2008,"lang":"en","type":"article","venue":"Software & Systems Modeling","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":41,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Dependability; Computer science; Formalism (music); Probabilistic logic; Reliability engineering; Software engineering; Embedded system; Artificial intelligence; Engineering","score_opus":0.062635797815128,"score_gpt":0.30598090796932975,"score_spread":0.24334511015420174,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2044138490","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2056046,0.00027922328,0.79282194,0.000018423734,0.00019330364,0.0003532407,0.000014138355,0.0005495264,0.00016558709],"genre_scores_gemma":[0.8658468,0.00002471641,0.13392866,0.000007726485,0.000051423503,0.00006647012,0.000007847338,0.000019880617,0.00004642935],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99574196,0.0003632184,0.0010646319,0.000866385,0.0014066861,0.00055714307],"domain_scores_gemma":[0.9967582,0.00030035383,0.00024539232,0.0016550582,0.00080360915,0.00023737282],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012820361,0.00027603572,0.0007462045,0.00025868567,0.00038118128,0.00009975679,0.0014608244,0.00017793357,0.0000025646293],"category_scores_gemma":[0.00017893793,0.0002458294,0.00033876975,0.0007154551,0.000096079675,0.0006004891,0.0004803161,0.00034921616,0.000014339081],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000032905928,0.000087511864,0.031640135,0.0005864183,0.000052859996,0.000015430038,0.00030353857,0.96161926,0.00011468809,0.0049396544,0.000019717476,0.00061747554],"study_design_scores_gemma":[0.00023658205,0.00004012741,0.00012576192,0.00022842057,0.00001509776,0.00005441625,0.00018028423,0.99861753,0.000029496818,0.0002241941,0.000012609524,0.00023549233],"about_ca_topic_score_codex":0.000839375,"about_ca_topic_score_gemma":0.000015105086,"teacher_disagreement_score":0.66024226,"about_ca_system_score_codex":0.0004342671,"about_ca_system_score_gemma":0.0007103908,"threshold_uncertainty_score":0.9999994},"labels":[],"label_agreement":null},{"id":"W2047729406","doi":"10.4271/2015-01-0156","title":"Automated Decomposition and Allocation of Automotive Safety Integrity Levels Using Exact Solvers","year":2015,"lang":"en","type":"article","venue":"SAE International journal of passenger cars. Electronic and electrical systems","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Decomposition; Automotive industry; Structural integrity; Computer science; Reliability engineering; Automotive engineering; Engineering; Chemistry; Structural engineering","score_opus":0.03325067339439326,"score_gpt":0.33691542249888395,"score_spread":0.3036647491044907,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2047729406","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.68027943,0.0030416592,0.315205,0.00088272255,0.00032214448,0.00014719793,0.0000048178968,0.000056723908,0.000060262508],"genre_scores_gemma":[0.9982832,0.00028021226,0.0012702206,0.000025961437,0.000119262426,0.00000156839,0.0000031045656,0.000006444842,0.000009998626],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975554,0.00037246407,0.0006521522,0.00023174005,0.00088871736,0.0002995347],"domain_scores_gemma":[0.99719363,0.00034278666,0.00043396474,0.00012395873,0.0017171996,0.00018843084],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017492512,0.00013802272,0.00036803383,0.00036484233,0.00006102421,0.00015296144,0.00044666708,0.00011537783,0.000002030239],"category_scores_gemma":[0.00040867017,0.00011424918,0.000095181116,0.00043551828,0.00006959377,0.0005585649,0.000088869,0.00045900964,8.931084e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.002692914,0.003586921,0.13611874,0.0004936851,0.013235666,0.000526564,0.011098272,0.05608759,0.2773551,0.19119649,0.004327223,0.30328083],"study_design_scores_gemma":[0.0016618475,0.0010146027,0.029198624,0.00018707076,0.000091092195,0.0011101621,0.00017195052,0.95967007,0.0029463922,0.003136185,0.0005671729,0.00024483178],"about_ca_topic_score_codex":0.00031660506,"about_ca_topic_score_gemma":0.000014188415,"teacher_disagreement_score":0.90358245,"about_ca_system_score_codex":0.0008337511,"about_ca_system_score_gemma":0.00074600946,"threshold_uncertainty_score":0.46589464},"labels":[],"label_agreement":null},{"id":"W2049857333","doi":"10.1100/2012/614635","title":"Dependability Modeling and Assessment in UML-Based Software Development","year":2012,"lang":"en","type":"article","venue":"The Scientific World JOURNAL","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Dependability; Unified Modeling Language; Software engineering; Software development; Maintainability; Applications of UML; Verification and validation; Software quality; Software; Programming language; Engineering","score_opus":0.04711936695158537,"score_gpt":0.32193878975602536,"score_spread":0.27481942280444,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2049857333","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3772023,0.0004965848,0.6197918,0.0018345512,0.00045126487,0.00011112986,5.571855e-7,0.000035045243,0.00007678007],"genre_scores_gemma":[0.9073247,0.0000056037834,0.09232368,0.00006899497,0.000051022645,0.000008025511,0.0000010753632,0.0000043734435,0.00021257525],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9972901,0.00040132436,0.00042780774,0.00032959966,0.00094634754,0.0006048445],"domain_scores_gemma":[0.99865764,0.0002818964,0.00008610584,0.00055701,0.00017434047,0.00024300262],"candidate_categories":["scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0118987365,0.00012350257,0.00016796729,0.00043747906,0.0010501957,0.0010854417,0.0009815772,0.000027360244,0.000056522953],"category_scores_gemma":[0.0002481239,0.00007958281,0.00007655254,0.0013563657,0.00017830881,0.0006922512,0.0003639997,0.00070143305,0.000018932527],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016585773,0.00066648854,0.31727645,0.000059397535,0.000046163397,0.000013325331,0.003727566,0.030802725,0.0002466613,0.0018048238,0.00039232583,0.64494747],"study_design_scores_gemma":[0.00045535524,0.000013099076,0.049893077,0.0000828923,0.000009363482,0.00003875477,0.000110935434,0.9396406,0.0005264322,0.0049795164,0.003995324,0.00025466725],"about_ca_topic_score_codex":0.000021721311,"about_ca_topic_score_gemma":0.00019159068,"teacher_disagreement_score":0.90883785,"about_ca_system_score_codex":0.00025153306,"about_ca_system_score_gemma":0.00064606103,"threshold_uncertainty_score":0.99995154},"labels":[],"label_agreement":null},{"id":"W2055587761","doi":"10.1109/prdc.2013.24","title":"Reliability Prediction of Smartphone Applications through Failure Data Analysis","year":2013,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Computer science; Reliability (semiconductor); Reliability engineering; Software quality; Software; Embedded system; Engineering; Operating system; Software development","score_opus":0.035593678275928964,"score_gpt":0.2930579575104379,"score_spread":0.25746427923450893,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2055587761","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0066002775,0.00004528802,0.98997295,0.001993617,0.00001459055,0.00032334664,0.000042466232,0.00016311876,0.0008443715],"genre_scores_gemma":[0.78748214,0.00004419883,0.21182741,0.000071245966,0.000029052388,0.000109923974,0.0001493588,0.0000034160355,0.00028326054],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99808395,0.00011613519,0.00040849802,0.0006923629,0.00049543695,0.00020362138],"domain_scores_gemma":[0.99518955,0.00030460145,0.00008932957,0.0038962984,0.00043770124,0.00008252597],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00067662634,0.000094121206,0.00026000297,0.00015862685,0.0001032518,0.00010761374,0.0020540457,0.000072777984,0.0004923985],"category_scores_gemma":[0.00021435827,0.000073768715,0.0001422896,0.0035074372,0.00012414533,0.001443988,0.0008377524,0.00012931316,0.0001634164],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001444614,0.0040763738,0.543204,0.00051994773,0.0047938274,0.0000017724988,0.0017675578,0.01150372,0.00974046,0.08251674,0.11021302,0.23164812],"study_design_scores_gemma":[0.0002355039,0.00005492595,0.18420531,0.0000057779293,0.00027719917,0.0000012435844,0.00011667919,0.760598,0.0015378137,0.030333387,0.022427266,0.0002068703],"about_ca_topic_score_codex":0.0021824145,"about_ca_topic_score_gemma":0.00009957191,"teacher_disagreement_score":0.7808818,"about_ca_system_score_codex":0.00003123165,"about_ca_system_score_gemma":0.00007032696,"threshold_uncertainty_score":0.53914183},"labels":[],"label_agreement":null},{"id":"W2056331670","doi":"10.1007/s11155-006-9020-7","title":"Requirements Analysis for Engineering Computation: A Systematic Approach for Improving Reliability","year":2006,"lang":"en","type":"article","venue":"Reliable Computing","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":23,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Documentation; Computer science; Software requirements specification; Software engineering; Requirements engineering; System requirements specification; Requirements analysis; Reliability (semiconductor); Computation; Reliability engineering; Functional specification; Argument (complex analysis); Software; Software system; Systems engineering; Programming language; Software construction; Engineering","score_opus":0.024227704412229722,"score_gpt":0.278808969988708,"score_spread":0.2545812655764783,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056331670","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007064924,0.00029529465,0.9901294,0.00008725932,0.0001434843,0.0017379954,0.0000063360676,0.00044498764,0.00009033264],"genre_scores_gemma":[0.532809,7.0963324e-7,0.46681276,0.000014950379,0.00010565382,0.00014673495,0.000046423364,0.000015454065,0.000048271933],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99650174,0.00010087451,0.0010434025,0.0010184972,0.0005771489,0.00075832655],"domain_scores_gemma":[0.9963966,0.001610435,0.00035070145,0.00083463243,0.00069083995,0.000116804105],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003632932,0.00028407568,0.0008128477,0.0004952922,0.00048785267,0.00050915906,0.0009254255,0.00012290619,0.0000010168671],"category_scores_gemma":[0.00094165705,0.0002717338,0.00063990016,0.0022734243,0.00003302353,0.00038177974,0.000301343,0.00015871818,0.0000028814968],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007367206,0.00015061982,0.002940253,0.012490695,0.00023984231,7.0012914e-7,0.00007309271,0.97937095,0.00018674023,0.0035251821,0.000091261696,0.0009232976],"study_design_scores_gemma":[0.00047328707,0.00005034989,0.00052591314,0.00018771835,0.00029214594,0.0000021559079,0.000023846584,0.99689186,0.00021509865,0.0010059746,0.000040880976,0.000290785],"about_ca_topic_score_codex":0.00016926906,"about_ca_topic_score_gemma":0.0000020802308,"teacher_disagreement_score":0.52574414,"about_ca_system_score_codex":0.0002878961,"about_ca_system_score_gemma":0.00010532568,"threshold_uncertainty_score":0.9999735},"labels":[],"label_agreement":null},{"id":"W2056902367","doi":"10.1145/1294948.1294961","title":"Modification analysis support at the requirements level","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Computer science; Change impact analysis; Telephony; Source code; Decision support system; Regression testing; Software engineering; Reliability engineering; Risk analysis (engineering); Systems engineering; Data mining; Software development; Engineering; Software; Programming language; Software construction","score_opus":0.1317201055420356,"score_gpt":0.37993560252250397,"score_spread":0.24821549698046838,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2056902367","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.039507456,0.000017651135,0.9519952,0.0018439492,0.000029875398,0.00007509156,0.0000012169245,0.00007403668,0.006455527],"genre_scores_gemma":[0.97963375,0.000011595427,0.006887103,0.000302307,0.000018718372,0.0000052985247,0.000011673786,0.0000022653467,0.013127277],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9984599,0.00005853488,0.00024366016,0.00032727869,0.000635823,0.00027484665],"domain_scores_gemma":[0.9985226,0.00023375206,0.000050668463,0.0009635135,0.00014598556,0.00008351558],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0024219092,0.00006803532,0.000109353794,0.00020438158,0.00023687255,0.00010761982,0.00088864117,0.000035956815,0.00047346868],"category_scores_gemma":[0.00010482517,0.000041642415,0.00017138587,0.0021767018,0.00006213468,0.0002240762,0.0003072125,0.000071703886,0.00038184156],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000035458525,0.0005519555,0.3841432,0.000022291179,0.0021035585,0.00002720716,0.0018599482,0.0030388003,0.0061850953,0.07219185,0.016124973,0.5137157],"study_design_scores_gemma":[0.0004009862,0.00008944027,0.7069877,0.0000028806414,0.00029830253,0.000006096705,0.00017601866,0.22946106,0.02882624,0.00453793,0.028774893,0.0004384728],"about_ca_topic_score_codex":0.00019522435,"about_ca_topic_score_gemma":0.00043975725,"teacher_disagreement_score":0.9451081,"about_ca_system_score_codex":0.00011121399,"about_ca_system_score_gemma":0.000035530305,"threshold_uncertainty_score":0.518415},"labels":[],"label_agreement":null},{"id":"W2067391523","doi":"10.1504/ijseam.2014.063879","title":"Availability and performability analysis for a service degradation process with condition-based preventive maintenance I - formulation and optimisation","year":2014,"lang":"en","type":"article","venue":"International Journal of Strategic Engineering Asset Management","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Reliability engineering; Preventive maintenance; Degradation (telecommunications); Computer science; Process (computing); Service (business); Engineering; Business; Operating system; Telecommunications","score_opus":0.016084254271671757,"score_gpt":0.274501457626562,"score_spread":0.2584172033548902,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2067391523","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.40069294,0.0000068615245,0.5986351,0.00047525193,0.000023935909,0.00012026842,0.0000030672124,0.000012150092,0.000030470532],"genre_scores_gemma":[0.9519714,0.000010662161,0.04791373,0.000028426743,0.000021144579,0.000021945249,0.000022121128,0.0000035066205,0.000007045216],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9988958,0.00003627598,0.00031724435,0.00020809818,0.00043110372,0.000111430214],"domain_scores_gemma":[0.9985131,0.00022155883,0.00021594886,0.00013130094,0.0008591455,0.000058945592],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010723978,0.00009955512,0.00016202725,0.00032076705,0.00004892744,0.00021135288,0.00026346577,0.000028102288,0.0000034300222],"category_scores_gemma":[0.00007582704,0.00008127089,0.000058672198,0.00036870773,0.00001840526,0.0006006302,0.00003279541,0.000085981825,2.0995e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012928898,0.00015033997,0.032453023,0.00045746137,0.0011859492,0.0000033038077,0.00020516397,0.93260235,0.00007891478,0.025505967,0.000003477146,0.0072247465],"study_design_scores_gemma":[0.00065097713,0.00012632659,0.07967446,0.000060816903,0.000120301665,0.0000054111124,0.00007485431,0.9150874,0.000121353056,0.003967602,0.00002427987,0.00008621896],"about_ca_topic_score_codex":0.000008935916,"about_ca_topic_score_gemma":0.000009454751,"teacher_disagreement_score":0.5512785,"about_ca_system_score_codex":0.00009736807,"about_ca_system_score_gemma":0.00003600303,"threshold_uncertainty_score":0.3314131},"labels":[],"label_agreement":null},{"id":"W2068147352","doi":"10.1115/detc2010-28438","title":"APU FMEA Validation and Its Application to Fault Identification","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"National Research Council Canada","funders":"","keywords":"Failure mode and effects analysis; Reliability engineering; Computer science; Identification (biology); Process (computing); Fault (geology); Data mining; Root cause; Engineering","score_opus":0.014144928557254641,"score_gpt":0.30415704492975226,"score_spread":0.29001211637249763,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2068147352","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3022798,0.0000062103964,0.6947526,0.0023503504,0.000043427877,0.00016561936,4.5918645e-7,0.000095507894,0.00030604607],"genre_scores_gemma":[0.9910533,0.0000071401846,0.008229726,0.000095949625,0.000033438675,0.00005077862,0.0000054178718,0.0000025019467,0.0005217355],"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9991396,0.000028599285,0.00014340688,0.00032425803,0.0002524827,0.00011165736],"domain_scores_gemma":[0.999184,0.0000770178,0.000029397132,0.00041456317,0.00019785522,0.0000971498],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00054610305,0.00005020643,0.000058549107,0.00011020139,0.00010028615,0.00020487055,0.00036876413,0.000045453715,0.000020732821],"category_scores_gemma":[0.00023625846,0.000044775155,0.000020376669,0.0004966521,0.000012281261,0.0003911411,0.000121474135,0.00009271473,0.00036989446],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000021941771,0.00006956815,0.0020834415,0.0000180846,0.0000080108375,1.8029762e-7,0.00030158993,0.00013941797,0.5291606,0.12367496,0.00030106143,0.34424087],"study_design_scores_gemma":[0.00016597012,0.00003520263,0.02813957,0.0000039829524,0.0000075367593,0.000005410337,0.00002675201,0.4933875,0.45379466,0.012124217,0.012060665,0.0002485533],"about_ca_topic_score_codex":0.000033940145,"about_ca_topic_score_gemma":0.000026667669,"teacher_disagreement_score":0.6887735,"about_ca_system_score_codex":0.000010866067,"about_ca_system_score_gemma":0.00001863607,"threshold_uncertainty_score":0.4754368},"labels":[],"label_agreement":null},{"id":"W2073803340","doi":"10.1007/s11219-007-9029-0","title":"In search for a widely applicable and accepted software quality model for software quality engineering","year":2007,"lang":"en","type":"article","venue":"Software Quality Journal","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":46,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure; Ubisoft (Canada)","funders":"","keywords":"Software quality control; Software quality; Software quality analyst; Software quality management; Quality (philosophy); Software engineering; Computer science; Software development; Social software engineering; Software construction; Verification and validation; Software; Engineering; Operations management","score_opus":0.08279210101335747,"score_gpt":0.39928907863032326,"score_spread":0.31649697761696577,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2073803340","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12419852,0.00054313353,0.87214404,0.0010971193,0.00020902189,0.0011854316,0.00016711948,0.0004502744,0.0000053418003],"genre_scores_gemma":[0.23379922,0.000102431055,0.7642924,0.0007604843,0.00039832015,0.0002292554,0.000047712398,0.00009046024,0.00027970213],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99117464,0.00067201944,0.002689131,0.001516405,0.0018096265,0.0021381828],"domain_scores_gemma":[0.9848168,0.010456649,0.0006047134,0.0014100822,0.0016839411,0.0010278013],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.028898586,0.0006457652,0.001398985,0.0007865005,0.00092064915,0.00088412734,0.0020719669,0.0005408289,0.000034353994],"category_scores_gemma":[0.012217147,0.0006378393,0.000773237,0.0014736602,0.0002436185,0.0018137443,0.0007145758,0.0014830015,0.000009560914],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0025125348,0.0024665887,0.28799257,0.006681212,0.0009092254,0.00008316275,0.011748579,0.08919263,0.002410642,0.029105661,0.0028974984,0.5639997],"study_design_scores_gemma":[0.021671552,0.0013914393,0.21004494,0.0009882029,0.00022278092,0.00038320024,0.0026164881,0.52846545,0.006903578,0.21275382,0.007416884,0.0071416916],"about_ca_topic_score_codex":0.0004532177,"about_ca_topic_score_gemma":0.0002916127,"teacher_disagreement_score":0.556858,"about_ca_system_score_codex":0.00070289534,"about_ca_system_score_gemma":0.0008714026,"threshold_uncertainty_score":0.99995327},"labels":[],"label_agreement":null},{"id":"W2075425374","doi":"10.3166/jesa.40.915-935","title":"Modeling dynamic reliability using dynamic Bayesian networks","year":2006,"lang":"fr","type":"article","venue":"Journal Européen des Systèmes Automatisés","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Dynamic Bayesian network; Computer science; Influence diagram; Bayesian network; Inference; Process (computing); Representation (politics); Reliability (semiconductor); Bayesian inference; Graphical model; Bayesian probability; State (computer science); Component (thermodynamics); Machine learning; Artificial intelligence; Algorithm; Decision tree","score_opus":0.023260057009938936,"score_gpt":0.2895239469287378,"score_spread":0.26626388991879885,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2075425374","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15386927,0.025492845,0.8162596,0.0014558653,0.0018857537,0.0003378722,0.000015087575,0.0003549365,0.0003287172],"genre_scores_gemma":[0.8772229,0.0013978096,0.11728071,0.00007646957,0.00059759454,0.00000677478,0.000009568086,0.00011120725,0.0032969688],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9897654,0.002679197,0.0024382435,0.0012468913,0.0017898857,0.002080355],"domain_scores_gemma":[0.9950866,0.00062167196,0.00071732106,0.0015212281,0.0013524897,0.00070068863],"candidate_categories":["metaepi_narrow","sts","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0047648153,0.00078852626,0.0011870622,0.0006461775,0.0018725756,0.0031669196,0.002115646,0.0005199562,0.0002503352],"category_scores_gemma":[0.000805529,0.0007703281,0.00100338,0.0025761009,0.000733977,0.0025976452,0.0007849305,0.0021024484,0.00013794654],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013558402,0.00028650646,0.0014048675,0.0003381738,0.00011310535,0.00040722842,0.00018070893,0.7470021,0.000088787376,0.0008862549,0.00013628881,0.24914238],"study_design_scores_gemma":[0.0004861977,0.00014991566,0.019411778,0.0013055616,0.00018740841,0.0023459482,0.000051212934,0.9305054,0.0000036057227,0.044566423,0.00028349832,0.00070305215],"about_ca_topic_score_codex":0.002209567,"about_ca_topic_score_gemma":0.00046868096,"teacher_disagreement_score":0.7233536,"about_ca_system_score_codex":0.0027789932,"about_ca_system_score_gemma":0.0009871437,"threshold_uncertainty_score":0.99947476},"labels":[],"label_agreement":null},{"id":"W2077138769","doi":"10.1016/j.jss.2003.05.006","title":"Towards automatic monitoring of component-based software systems","year":2004,"lang":"en","type":"article","venue":"Journal of Systems and Software","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo; Bell (Canada); Queen's University","funders":"","keywords":"Correctness; Component (thermodynamics); Computer science; Asynchronous communication; Component-based software engineering; Distributed computing; Software; Concurrency; Software system; Real-time computing; Software architecture; Software quality; Embedded system; Reliability engineering; Software development; Programming language; Engineering; Computer network","score_opus":0.02532862100587766,"score_gpt":0.27847469964679067,"score_spread":0.253146078640913,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2077138769","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.38134286,0.0060407636,0.61149925,0.000115754694,0.0007567929,0.00015656851,0.0000059440094,0.000077614175,0.0000044437243],"genre_scores_gemma":[0.95270836,0.00010493915,0.04695735,0.0000080349955,0.00018302123,0.000006892381,7.5950237e-7,0.000013837697,0.00001682821],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99693084,0.00019495035,0.0011215904,0.00024934343,0.0011829421,0.00032035445],"domain_scores_gemma":[0.9972283,0.00040773526,0.00075243664,0.000488212,0.00084932975,0.00027398672],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014992272,0.00020351488,0.00079346483,0.00041311226,0.00015475362,0.00031499675,0.0008170159,0.00013276706,0.0000020761772],"category_scores_gemma":[0.00059559254,0.00015563649,0.00025687215,0.0005365726,0.00009455603,0.0005174805,0.00012806269,0.00030904866,0.0000037988339],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010880426,0.0015860681,0.5100981,0.014593545,0.0014652183,0.00094914075,0.0049460707,0.33688173,0.0020561896,0.0057370565,0.00040642274,0.121171646],"study_design_scores_gemma":[0.033882692,0.010655296,0.4850023,0.071815655,0.0013324108,0.007747556,0.013009642,0.33734345,0.014597539,0.010578302,0.007896498,0.0061386526],"about_ca_topic_score_codex":0.00041784218,"about_ca_topic_score_gemma":0.0000011100691,"teacher_disagreement_score":0.5713655,"about_ca_system_score_codex":0.00018703278,"about_ca_system_score_gemma":0.00047125688,"threshold_uncertainty_score":0.63466716},"labels":[],"label_agreement":null},{"id":"W2079518339","doi":"10.1109/ccece.2010.5575168","title":"Comparative study of heuristics for reliability optimization of complex systems","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Heuristics; Tabu search; Simulated annealing; Computer science; Mathematical optimization; Reliability (semiconductor); Metaheuristic; Artificial intelligence; Algorithm; Mathematics","score_opus":0.06703314413928539,"score_gpt":0.35788983063887253,"score_spread":0.2908566864995871,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2079518339","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2408726,0.0000042224437,0.75812936,0.000036103727,0.00008485967,0.0005815981,0.0000067792985,0.00003086591,0.0002536132],"genre_scores_gemma":[0.89889884,0.0000010928491,0.101006255,0.000002502421,0.0000124337075,0.00002575493,0.0000041113954,0.0000023320229,0.000046704943],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99866444,0.00013128718,0.00045797628,0.00026601207,0.00035562803,0.00012463359],"domain_scores_gemma":[0.9971697,0.00087233126,0.00015690461,0.0006902986,0.0010592752,0.0000515228],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00088539644,0.000076012046,0.00036660512,0.00009001691,0.000061450606,0.00003589494,0.00059376674,0.00004285216,0.000017581966],"category_scores_gemma":[0.00042755742,0.000060285205,0.00006704443,0.00044234068,0.00010229272,0.00012363833,0.00013660653,0.00009708442,0.0000011418524],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006424093,0.004627031,0.033567198,0.00040393398,0.00014427766,4.671432e-7,0.0034805334,0.92188406,0.0026605397,0.031507112,0.00077793276,0.0008826839],"study_design_scores_gemma":[0.00037568706,0.00039649563,0.00342526,0.000003995573,0.000011309475,3.4669446e-7,0.00052289595,0.99415153,0.0006344525,0.00032738532,0.00008869224,0.0000619427],"about_ca_topic_score_codex":0.0002831109,"about_ca_topic_score_gemma":0.00004846549,"teacher_disagreement_score":0.6580262,"about_ca_system_score_codex":0.000011700544,"about_ca_system_score_gemma":0.000058800408,"threshold_uncertainty_score":0.24583593},"labels":[],"label_agreement":null},{"id":"W2079764904","doi":"10.1109/iecon.2012.6389399","title":"From requirements to software trustworthiness using scenarios and finite state machine","year":2012,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Computer science; Software construction; Software engineering; Finite-state machine; Verification and validation; Software quality; Software; Software development; Software verification; Trustworthiness; Software metric; Programming language; Computer security; Engineering","score_opus":0.045804192474070686,"score_gpt":0.3205734196787983,"score_spread":0.27476922720472763,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2079764904","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3550764,0.0001883346,0.6441166,0.0002839099,0.000099819816,0.00009567121,0.0000078604535,0.0001043445,0.000027091493],"genre_scores_gemma":[0.81028736,0.000018335266,0.18879801,0.0005062927,0.000083228,0.000005881201,0.0000060839743,0.00000931054,0.00028550305],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9983964,0.0000968928,0.00022735608,0.00037699533,0.0004392191,0.00046311688],"domain_scores_gemma":[0.99865866,0.0003010238,0.00003946299,0.0005431175,0.00009588308,0.0003618627],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00057720416,0.00013912632,0.00019793799,0.00014190486,0.00016374863,0.00021582824,0.00048546106,0.000041029067,0.00011708185],"category_scores_gemma":[0.00034272086,0.00011229176,0.000049046266,0.000588479,0.00003460368,0.00084525946,0.0007019248,0.00012453593,0.00007388725],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003301095,0.0003039725,0.6596872,0.00004838212,0.00014641297,0.000026635307,0.0046743965,0.0047937487,0.0007840642,0.0003583461,0.00038564485,0.32875815],"study_design_scores_gemma":[0.0011878968,0.00013379556,0.07619214,0.00012964597,0.000052770043,0.000017268296,0.00011835961,0.90230685,0.002375487,0.008005321,0.008370941,0.001109502],"about_ca_topic_score_codex":0.0018894742,"about_ca_topic_score_gemma":0.00006336242,"teacher_disagreement_score":0.89751315,"about_ca_system_score_codex":0.00005086555,"about_ca_system_score_gemma":0.000041741176,"threshold_uncertainty_score":0.4579125},"labels":[],"label_agreement":null},{"id":"W2080002273","doi":"10.1287/isre.1050.0059","title":"Optimal Software Development: A Control Theoretic Approach","year":2005,"lang":"en","type":"article","venue":"Information Systems Research","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":40,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Debugging; Computer science; Product (mathematics); Construct (python library); Key (lock); Software; Constraint (computer-aided design); Control (management); Reliability engineering; Programming language; Engineering; Computer security; Mathematics","score_opus":0.03645303545881635,"score_gpt":0.3123647392672856,"score_spread":0.27591170380846924,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2080002273","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006050813,0.00023700898,0.98417217,0.0003775271,0.000063427455,0.00069189514,0.000004779657,0.00029198077,0.008110406],"genre_scores_gemma":[0.94870186,0.000016525813,0.050218992,0.00007588715,0.000110205656,0.00034379587,0.000021291888,0.000007504362,0.0005039428],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9951898,0.00061307044,0.00072997465,0.00029278244,0.0024053983,0.000768974],"domain_scores_gemma":[0.9968238,0.000655659,0.00010353084,0.00080967374,0.0013516028,0.000255723],"candidate_categories":["scholarly_communication","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007076903,0.00015156469,0.0002572886,0.00086272706,0.0005805615,0.0014220285,0.0016616086,0.0001414991,0.000030265273],"category_scores_gemma":[0.00091472885,0.00012416457,0.00008402747,0.0015386685,0.00016780522,0.0031716772,0.0003731664,0.000542256,0.00259252],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010574541,0.0004478207,0.0032420144,0.001257225,0.00029333765,0.00000848715,0.026166499,0.16454409,0.00003956432,0.20906553,0.009565866,0.58526385],"study_design_scores_gemma":[0.0008334946,0.000053723877,0.0005160706,0.000050257473,0.0000023308005,0.000037274385,0.00074552745,0.74113053,0.0001894291,0.000097456585,0.25609577,0.00024815058],"about_ca_topic_score_codex":0.00004512965,"about_ca_topic_score_gemma":0.0000010855968,"teacher_disagreement_score":0.94265103,"about_ca_system_score_codex":0.0003346986,"about_ca_system_score_gemma":0.00056142086,"threshold_uncertainty_score":0.9996146},"labels":[],"label_agreement":null},{"id":"W2084598998","doi":"10.1145/2379776.2379778","title":"Dependability modeling and analysis of software systems specified with UML","year":2012,"lang":"en","type":"review","venue":"ACM Computing Surveys","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":90,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Seventh Framework Programme; Ministerio de Economía y Competitividad","keywords":"Dependability; Maintainability; Computer science; Unified Modeling Language; Software engineering; Software quality; Applications of UML; Reliability (semiconductor); Reliability engineering; Software development; Software; Programming language; Engineering","score_opus":0.12444972692334078,"score_gpt":0.3504899269273082,"score_spread":0.2260402000039674,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2084598998","genre_codex":"methods","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00027064988,0.49107945,0.508167,0.0000033997485,0.000077846504,0.0002561466,0.000026368125,0.00011052997,0.000008596814],"genre_scores_gemma":[0.029710988,0.916719,0.05309396,0.0000045309025,0.00014084768,0.000021048414,0.00022520454,0.000059067286,0.000025341018],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9917734,0.003930701,0.0013352045,0.0012641472,0.000989202,0.0007073041],"domain_scores_gemma":[0.99028474,0.0050014723,0.00069060817,0.003190245,0.0005561018,0.0002768271],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.011469002,0.0005644425,0.003641138,0.0010487469,0.00022726784,0.00032671483,0.0026352594,0.00033185547,0.000007344486],"category_scores_gemma":[0.0013432894,0.00041542813,0.0007528994,0.00484588,0.00015895221,0.00027176016,0.0019159827,0.00065630145,0.000008749039],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000015577358,0.00008515311,0.0068824594,0.007083232,0.0022999097,0.0000052974538,0.00011070115,0.009658035,1.8531155e-8,0.00011540594,0.0000021494986,0.9737561],"study_design_scores_gemma":[0.00020925117,0.00009215638,0.0021173046,0.0043052584,0.006368074,0.000036204976,0.00002292543,0.97782,2.9647128e-7,0.00008674244,0.007826239,0.0011155499],"about_ca_topic_score_codex":0.0010654271,"about_ca_topic_score_gemma":0.00008526264,"teacher_disagreement_score":0.9726405,"about_ca_system_score_codex":0.00014962992,"about_ca_system_score_gemma":0.00032985382,"threshold_uncertainty_score":0.99982977},"labels":[],"label_agreement":null},{"id":"W2086340664","doi":"10.5555/2755753.2757191","title":"Towards an accurate reliability, availability and maintainability analysis approach for satellite systems based on probabilistic model checking","year":2015,"lang":"en","type":"article","venue":"Design, Automation, and Test in Europe","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal; Concordia University","funders":"","keywords":"Maintainability; Computer science; Markov chain; Erlang distribution; Avionics; Reliability engineering; Probabilistic logic; Geostationary orbit; Model checking; Reliability (semiconductor); Satellite system; Exponential distribution; Real-time computing; Satellite; Markov model; Geostationary Operational Environmental Satellite; Algorithm; Global Positioning System; GNSS applications; Engineering; Telecommunications; Aerospace engineering; Statistics","score_opus":0.07319186676392403,"score_gpt":0.30701048343312365,"score_spread":0.23381861666919962,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2086340664","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10338151,0.00010585683,0.8943848,0.00022907532,0.000027229975,0.0013973799,0.000017363522,0.0002342839,0.00022248339],"genre_scores_gemma":[0.8873451,0.000021822278,0.11224902,0.00005846313,0.000020434847,0.00020834696,0.0000546142,0.00001567686,0.000026530195],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.995643,0.001238064,0.0007592071,0.001296146,0.0006039947,0.0004595968],"domain_scores_gemma":[0.9949901,0.0019447093,0.0001813169,0.0012608567,0.0012712722,0.00035173638],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.009508323,0.00030417516,0.0005790611,0.00046962049,0.0002207897,0.00066011143,0.00061467214,0.00013152412,0.0000016023682],"category_scores_gemma":[0.010073066,0.00025854094,0.00009033774,0.002593793,0.00025203358,0.0007622109,0.00015692113,0.00020800364,0.0000015743992],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006369017,0.0008659528,0.09317772,0.000490054,0.00003590683,0.000002617728,0.0010460437,0.8942063,0.000018068806,0.004676106,0.000017341885,0.0054002325],"study_design_scores_gemma":[0.0004960543,0.00033299025,0.059023727,0.000018269595,0.00006891048,0.0000020067741,0.000052958894,0.93265533,0.000024173472,0.0070173037,0.00004331708,0.00026496337],"about_ca_topic_score_codex":0.00018339515,"about_ca_topic_score_gemma":0.00001765179,"teacher_disagreement_score":0.78396356,"about_ca_system_score_codex":0.0002559914,"about_ca_system_score_gemma":0.0005451654,"threshold_uncertainty_score":0.9999867},"labels":[],"label_agreement":null},{"id":"W2088900036","doi":"10.1145/1833335.1833342","title":"A taxonomy of software architecture-based reliability efforts","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Software quality; Computer science; Software architecture; Component (thermodynamics); Viewpoints; Architecture; Reliability (semiconductor); Reliability engineering; Software engineering; Granularity; Reference architecture; Software; Software development; Engineering; Programming language","score_opus":0.01451024800755707,"score_gpt":0.2503730997503336,"score_spread":0.23586285174277655,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2088900036","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12939627,0.000008877063,0.868527,0.00069617527,0.000102815655,0.00023811721,0.000002364155,0.00021297792,0.0008154218],"genre_scores_gemma":[0.62925357,5.438385e-7,0.37049434,0.00009971817,0.000020737662,0.00004030737,0.0000018570147,0.0000041439916,0.00008480187],"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9982958,0.00008546759,0.00032772863,0.000489137,0.0004918687,0.00031000737],"domain_scores_gemma":[0.99721676,0.00076302915,0.000078929916,0.0015153285,0.00026012823,0.00016581104],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00096436666,0.00013298768,0.00025461602,0.0001776902,0.0000826259,0.000066851266,0.0011533472,0.00010371022,0.00026315686],"category_scores_gemma":[0.0011547414,0.00010135543,0.00022931666,0.00068864966,0.00022221942,0.00016364951,0.00024733914,0.00040521738,0.00003251991],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006396487,0.0014361438,0.23351832,0.00045155958,0.00008451692,0.000020744497,0.00022455216,0.009970598,0.008136824,0.023827465,0.0016239897,0.7206413],"study_design_scores_gemma":[0.003603073,0.001091989,0.0935602,0.00012730458,0.000076717035,0.00004377812,0.00003277512,0.21978615,0.30512738,0.1509646,0.2232745,0.002311547],"about_ca_topic_score_codex":0.00020051269,"about_ca_topic_score_gemma":0.00009786228,"teacher_disagreement_score":0.7183298,"about_ca_system_score_codex":0.000017581318,"about_ca_system_score_gemma":0.00027111743,"threshold_uncertainty_score":0.41331545},"labels":[],"label_agreement":null},{"id":"W2090082685","doi":"10.1109/mper.2001.4311432","title":"Control/Maintenance Strategy for Fault Tolerant Mode and Reliability Analysis of Hydro Power Stations","year":2001,"lang":"en","type":"article","venue":"IEEE Power Engineering Review","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Reliability (semiconductor); Reliability engineering; Power (physics); Computer science; Fault tolerance; Fault (geology); Engineering","score_opus":0.015751753624769785,"score_gpt":0.302623471574089,"score_spread":0.2868717179493192,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2090082685","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.020438835,0.009381463,0.96881115,0.0006970299,0.00006728158,0.00043718825,0.000042305153,0.00008160316,0.0000431512],"genre_scores_gemma":[0.9843872,0.0069097374,0.008406988,0.00014651696,0.000007541337,0.00008483368,0.000006089133,0.0000100305,0.000041036034],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99826944,0.000061897874,0.00056835986,0.00045674684,0.0003109423,0.00033259706],"domain_scores_gemma":[0.9980151,0.00060487975,0.00010889376,0.0007511412,0.00038650358,0.0001334594],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00097409904,0.00017640961,0.0007329283,0.00021432314,0.000048743917,0.000051536565,0.00043915326,0.000053497395,0.000029233883],"category_scores_gemma":[0.0006202967,0.00014774689,0.00035565594,0.0015535988,0.000046842433,0.00025072787,0.000037190453,0.00013795643,0.0000029075081],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016584678,0.00019912067,0.0011620385,0.0016892718,0.0008505754,0.000008461703,0.00018270573,0.9820171,0.0005186071,0.003983625,0.0005345787,0.008837305],"study_design_scores_gemma":[0.00030460706,0.00009445462,0.0025494006,0.0004643912,0.0003681202,0.0000041905028,0.000003853339,0.98666054,0.00006534745,0.0002969655,0.00895766,0.00023044698],"about_ca_topic_score_codex":0.000047853435,"about_ca_topic_score_gemma":0.0000049244804,"teacher_disagreement_score":0.96394837,"about_ca_system_score_codex":0.0000491808,"about_ca_system_score_gemma":0.0000546291,"threshold_uncertainty_score":0.6024943},"labels":[],"label_agreement":null},{"id":"W2092239706","doi":"10.1007/s00521-012-1251-4","title":"A neural network approach for optimal software testing and maintenance","year":2012,"lang":"en","type":"article","venue":"Neural Computing and Applications","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Reliability engineering; Software; Artificial neural network; Software reliability testing; Imperfect; Reliability (semiconductor); Limit (mathematics); Set (abstract data type); Process (computing); Software bug; Software quality; Software development; Machine learning; Mathematics","score_opus":0.041796702308936416,"score_gpt":0.2999298468616547,"score_spread":0.2581331445527183,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2092239706","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06969657,0.00057142996,0.9288067,0.00025897278,0.000025808551,0.0003666324,0.0000025943916,0.0002094848,0.0000617748],"genre_scores_gemma":[0.61699206,0.0000044002445,0.3825195,0.0001226517,0.00025969514,0.00007229229,0.0000037566565,0.0000059347713,0.000019692909],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9987555,0.000049453804,0.0001874489,0.00038654587,0.00012708695,0.0004939263],"domain_scores_gemma":[0.9983398,0.0010286703,0.000068644375,0.00028245812,0.00010754844,0.00017293036],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00047683605,0.0001222771,0.00016183426,0.000038872404,0.00065417646,0.00018441818,0.00030607666,0.00004461836,3.151053e-7],"category_scores_gemma":[0.00020935459,0.0001073861,0.000046545323,0.0004790921,0.000094698364,0.00019930615,0.00032349114,0.00017110251,0.0000010337243],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006536319,0.00013223075,0.091143206,0.0001503626,0.000024333096,3.6009632e-7,0.00028677148,0.049853966,0.00011330669,0.012301818,0.0005592485,0.8454279],"study_design_scores_gemma":[0.0001482245,0.000033588225,0.0123162735,0.000008580997,0.000009694342,0.00003740084,0.000022677761,0.9856122,0.000010212465,0.0005959538,0.0010663761,0.00013880442],"about_ca_topic_score_codex":0.0000132831165,"about_ca_topic_score_gemma":2.5970647e-7,"teacher_disagreement_score":0.93575823,"about_ca_system_score_codex":0.00001213954,"about_ca_system_score_gemma":0.000013925794,"threshold_uncertainty_score":0.50314623},"labels":[],"label_agreement":null},{"id":"W2096150300","doi":"10.1109/cmpsac.2002.1045036","title":"Reliability assessment of web applications","year":2003,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Computer science; Reliability engineering; Reliability (semiconductor); Markov chain; Formal verification; Markov model; Formal specification; Formal methods; Model checking; Markov process; Software engineering; Programming language; Engineering; Machine learning","score_opus":0.01744962278465697,"score_gpt":0.33177493687908954,"score_spread":0.3143253140944326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2096150300","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006193807,0.000023109886,0.94981444,0.0002572738,0.000015298243,0.00014344185,7.5648285e-7,0.00007187901,0.043479986],"genre_scores_gemma":[0.8181036,0.000017399185,0.18145849,0.00003157032,0.000003630564,0.000042359745,4.6656265e-7,0.000001449653,0.00034103016],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9989703,0.00012526542,0.00020235647,0.0002582307,0.00030948548,0.00013433697],"domain_scores_gemma":[0.99859095,0.00025875366,0.00003586587,0.0008549284,0.0001945571,0.000064929576],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009050805,0.000049828977,0.00011500221,0.00006365142,0.00005553436,0.00003119355,0.0004820052,0.000030126066,0.00014871976],"category_scores_gemma":[0.0001571821,0.00003951577,0.00007947433,0.0007061063,0.00006042094,0.00014085716,0.00008162941,0.000081791004,0.000024549123],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[2.6763763e-7,0.0003335568,0.036369402,0.00003238906,0.000013959707,2.7339198e-7,0.000019466383,0.00034301818,0.0014949153,0.94872165,0.00030215003,0.012368957],"study_design_scores_gemma":[0.0009094292,0.00022750707,0.10915916,0.00001806564,0.00002752204,0.000008220724,0.00012147873,0.37933666,0.019166859,0.21348982,0.27691746,0.00061781594],"about_ca_topic_score_codex":0.00001618577,"about_ca_topic_score_gemma":0.0000035778498,"teacher_disagreement_score":0.8119098,"about_ca_system_score_codex":0.00003588494,"about_ca_system_score_gemma":0.00023064033,"threshold_uncertainty_score":0.1628377},"labels":[],"label_agreement":null},{"id":"W2098703785","doi":"10.1109/ccece.2001.933661","title":"Quantitative assessment of extreme programming practices","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Extreme programming; Software quality; Computer science; Software engineering; Software development; Software quality management; Software; Extreme programming practices; Software quality analyst; Quality (philosophy); Software quality control; Software release life cycle; Software development process; Risk analysis (engineering); Programming language","score_opus":0.20886726772720965,"score_gpt":0.4167547937113875,"score_spread":0.20788752598417787,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2098703785","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016767535,0.00021665133,0.9681687,0.0019296265,0.00003153887,0.00012660367,2.754169e-7,0.0001022321,0.012656834],"genre_scores_gemma":[0.58056337,0.000041056708,0.41889912,0.000019582518,0.0000051863244,0.00000810468,2.1921558e-7,0.0000015981974,0.00046178725],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9988489,0.0001248625,0.00018691545,0.0002333529,0.000441406,0.00016456883],"domain_scores_gemma":[0.99862576,0.0005931562,0.00018386192,0.0003633751,0.00018216667,0.000051690222],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00060856686,0.000056657445,0.00012407659,0.00009537675,0.000056593777,0.00010497242,0.00049319526,0.000023751778,0.00023461232],"category_scores_gemma":[0.0005422846,0.00004332441,0.00006689543,0.0006181547,0.000054999244,0.0005520854,0.00014271728,0.00009650532,0.000031493975],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000033560193,0.001300144,0.050643433,0.000113180366,0.00016209243,0.000017463659,0.001299002,0.00022942788,0.0019199778,0.38169283,0.0009423237,0.5616768],"study_design_scores_gemma":[0.0001484027,0.00030468337,0.0048221867,0.00001616428,0.00000870854,0.0000028252077,0.00026653998,0.9845485,0.00069388456,0.0011912995,0.007877571,0.00011924451],"about_ca_topic_score_codex":0.00009967537,"about_ca_topic_score_gemma":0.000016281623,"teacher_disagreement_score":0.9843191,"about_ca_system_score_codex":0.00002065005,"about_ca_system_score_gemma":0.000028921651,"threshold_uncertainty_score":0.25688404},"labels":[],"label_agreement":null},{"id":"W2101423475","doi":"10.1109/csmr.2001.914977","title":"Modeling Web maintenance centers through queue models","year":2001,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Computer science; Queue; Computer network","score_opus":0.048908072421653494,"score_gpt":0.2925033662589266,"score_spread":0.2435952938372731,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2101423475","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017690498,0.000120841505,0.9616906,0.003411613,0.00007015407,0.00008433014,6.352842e-7,0.00028756697,0.016643772],"genre_scores_gemma":[0.94222754,0.00057324563,0.05501924,0.00073143956,0.000032730382,0.000010358387,0.00000113447,0.0000068972317,0.0013974018],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99830216,0.000065055974,0.00023245865,0.00048582314,0.00043929907,0.0004751838],"domain_scores_gemma":[0.99892706,0.00006761963,0.000020419482,0.00071659865,0.00016334174,0.000104972074],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00035906603,0.000121181314,0.00016874383,0.00007479834,0.00013112735,0.00018303106,0.0010483384,0.00005475602,0.00005224017],"category_scores_gemma":[0.000057030444,0.000096070435,0.00013755039,0.00065584475,0.00004105914,0.0011625661,0.0003079248,0.00016739764,0.00011809773],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011989579,0.00015551933,0.00079274573,0.000015895925,0.00004721754,0.000045261986,0.00052513287,0.76892656,0.00018059294,0.21144089,0.0016419255,0.01621625],"study_design_scores_gemma":[0.00015906035,0.000014541601,0.0000044073045,0.000014806323,0.0000022322022,0.00001071789,0.000052998876,0.9505951,0.000049402588,0.04737399,0.0016025058,0.00012027861],"about_ca_topic_score_codex":0.0006191068,"about_ca_topic_score_gemma":0.00008041796,"teacher_disagreement_score":0.92453706,"about_ca_system_score_codex":0.00007288851,"about_ca_system_score_gemma":0.00008982554,"threshold_uncertainty_score":0.39176387},"labels":[],"label_agreement":null},{"id":"W2102830428","doi":"10.1109/sere.2012.33","title":"A Control Flow Representation for Component-Based Software Reliability Analysis","year":2012,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Computer science; Control flow; Control flow graph; Component (thermodynamics); Data-flow analysis; Control flow analysis; Theoretical computer science; Data flow diagram; Component-based software engineering; Distributed computing; Software; Data structure; Programming language; Software system; Database; Procedural programming","score_opus":0.027166407200500984,"score_gpt":0.31339049501251337,"score_spread":0.28622408781201236,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2102830428","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016051508,0.000058458078,0.98137456,0.0015570518,0.00009908584,0.0004538461,0.000019581508,0.00033044867,0.000055460347],"genre_scores_gemma":[0.7073714,0.0000013114187,0.2920138,0.00030012033,0.000054323133,0.0001249943,0.000042430052,0.0000054910324,0.00008614518],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976241,0.00030078183,0.00039760128,0.0005465733,0.00057713396,0.0005537989],"domain_scores_gemma":[0.9957972,0.002185,0.0000934121,0.0012313935,0.00042644798,0.0002665011],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0018760009,0.00015284042,0.00040290898,0.00034400416,0.0002168029,0.00015849767,0.00068971026,0.00008879876,0.00012899964],"category_scores_gemma":[0.0015300465,0.00012673013,0.0006375534,0.001956037,0.00007398136,0.00065292337,0.00009289824,0.00011531843,0.000049335235],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011751877,0.0011042852,0.8371278,0.000093561786,0.0009025242,0.0000012075079,0.00025461242,0.110696524,0.00024245857,0.0036611075,0.0018850823,0.04391329],"study_design_scores_gemma":[0.00077409967,0.00004403484,0.050311424,0.0000027253827,0.00022476823,3.7095256e-7,0.000012730497,0.94543856,0.0009590526,0.0008346599,0.0012213166,0.00017626052],"about_ca_topic_score_codex":0.00022269224,"about_ca_topic_score_gemma":0.000018116009,"teacher_disagreement_score":0.834742,"about_ca_system_score_codex":0.00010435548,"about_ca_system_score_gemma":0.00007754918,"threshold_uncertainty_score":0.51679045},"labels":[],"label_agreement":null},{"id":"W2103473627","doi":"10.1109/icc.1991.162297","title":"The RQMS as an evolving system of software quality measurements","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Nortel (Canada)","funders":"","keywords":"CLARITY; Statement (logic); Quality (philosophy); Computer science; Set (abstract data type); Software quality; Reliability (semiconductor); Field (mathematics); Software; Focus (optics); Software engineering; Software development; Telecommunications; Engineering management; Engineering; Operating system; Programming language; Mathematics","score_opus":0.10205692072864785,"score_gpt":0.33515330142099425,"score_spread":0.2330963806923464,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2103473627","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1308699,0.0017025835,0.8488084,0.0012147796,0.0003150296,0.00045049447,0.0000024349736,0.0008537266,0.015782664],"genre_scores_gemma":[0.990148,0.000018669194,0.009124435,0.000035475055,0.000020773725,0.000009209232,3.5450194e-7,0.000004124992,0.0006389346],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9974871,0.00046323295,0.0003689672,0.0003200377,0.001081342,0.00027931182],"domain_scores_gemma":[0.9976281,0.00052153715,0.00009316218,0.0011363938,0.00050203357,0.000118808835],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025342184,0.000089652516,0.00017581483,0.00004302714,0.00032136773,0.00024258856,0.0014047812,0.00004567485,0.000080814265],"category_scores_gemma":[0.0011463498,0.000055593424,0.000114881856,0.00051740115,0.00007907807,0.00045229497,0.00022264704,0.00010691521,0.00012566843],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002393861,0.001161585,0.24365726,0.00065764785,0.0004693291,0.000013815535,0.0034270473,0.0005160508,0.00187263,0.124876514,0.003952228,0.61937195],"study_design_scores_gemma":[0.0018348637,0.0008519281,0.08487313,0.00035394038,0.00006719467,0.000037373386,0.0031626613,0.87080777,0.019282315,0.012387149,0.004996205,0.001345445],"about_ca_topic_score_codex":0.00048285956,"about_ca_topic_score_gemma":0.00006372642,"teacher_disagreement_score":0.87029177,"about_ca_system_score_codex":0.000082613245,"about_ca_system_score_gemma":0.00003717353,"threshold_uncertainty_score":0.26104563},"labels":[],"label_agreement":null},{"id":"W2104268806","doi":"10.1007/978-3-642-35795-4_5","title":"Software Testing is Necessary But Not Sufficient for Software Trustworthiness","year":2013,"lang":"en","type":"book-chapter","venue":"Communications in computer and information science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Computer science; Trustworthiness; Software reliability testing; Software construction; Software quality; Verification and validation; Software engineering; Software; Software quality analyst; Software development; Software verification; Software quality assurance; Software metric; Reliability engineering; Programming language; Computer security; Engineering","score_opus":0.0683685272162676,"score_gpt":0.3153826507900473,"score_spread":0.2470141235737797,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2104268806","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0002720613,0.00033306214,0.9893803,0.0013098099,0.00022860814,0.0008163945,0.000056025503,0.00023961665,0.007364135],"genre_scores_gemma":[0.012985924,0.0009898236,0.9799307,0.002816602,0.00007810497,0.00021224593,0.00011414669,0.000022072876,0.0028503768],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9970732,0.000041423064,0.0009864161,0.0005632811,0.00087730924,0.00045840853],"domain_scores_gemma":[0.9926351,0.0018325034,0.00043847272,0.0030161748,0.0018824822,0.00019527861],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science"],"consensus_categories":[],"category_scores_codex":[0.0018617655,0.00032170516,0.00042266923,0.0011663671,0.0010735332,0.0014804386,0.005527319,0.00020528796,0.000022051452],"category_scores_gemma":[0.0008454543,0.0003058142,0.00012068191,0.0010075682,0.0010150834,0.0068043587,0.0038361456,0.00058898545,0.00010357172],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000036988074,0.000047431655,0.0005241602,0.00019214874,0.000016322972,4.2850507e-7,0.0017217216,0.0011607469,0.0000023019675,0.09262095,0.001241531,0.90246856],"study_design_scores_gemma":[0.00036359555,0.000096441436,0.002274848,0.0003459679,0.00001226736,0.000016677686,0.000035194436,0.9229171,0.00002644484,0.012161697,0.061177075,0.0005726479],"about_ca_topic_score_codex":0.00005656018,"about_ca_topic_score_gemma":0.000005488881,"teacher_disagreement_score":0.9217564,"about_ca_system_score_codex":0.00020144093,"about_ca_system_score_gemma":0.00068794994,"threshold_uncertainty_score":0.9999394},"labels":[],"label_agreement":null},{"id":"W2105709232","doi":"10.1145/2372251.2372271","title":"A hybrid release planning method and its empirical justification","year":2012,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Constraint (computer-aided design); Mathematical optimization; Constraint programming; Mathematics; Stochastic programming","score_opus":0.09567323129377242,"score_gpt":0.41152391015189926,"score_spread":0.31585067885812684,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2105709232","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08088164,0.00053209404,0.91618764,0.001601913,0.000035031047,0.000052437277,3.6368334e-7,0.00010690166,0.00060197146],"genre_scores_gemma":[0.92644584,0.000014552981,0.07288739,0.00023362733,0.000051931227,0.0000077646655,0.0000012839216,0.000002643096,0.00035497916],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99903685,0.00015142666,0.00012151956,0.00022486078,0.00022635444,0.00023900376],"domain_scores_gemma":[0.99905723,0.0004055931,0.000022511154,0.00025941845,0.00006217387,0.00019307417],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009557536,0.000059501657,0.0000926125,0.00008333015,0.000100326884,0.00008567705,0.0002391551,0.00002702205,0.000028398692],"category_scores_gemma":[0.00042935478,0.000047007692,0.000032705942,0.00024147292,0.0000144991745,0.00047245793,0.0001623929,0.000107763706,0.00008181861],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029274664,0.0006796494,0.1270393,0.00017999145,0.000111807334,0.000025822197,0.00466315,0.00083981635,0.0047525903,0.0960841,0.014027752,0.75156677],"study_design_scores_gemma":[0.00020174327,0.000041144744,0.036993165,0.000014116971,0.000019017263,0.00006160657,0.000055926248,0.93013084,0.014375739,0.0035941994,0.014266989,0.00024550798],"about_ca_topic_score_codex":0.000008455995,"about_ca_topic_score_gemma":1.17520976e-7,"teacher_disagreement_score":0.929291,"about_ca_system_score_codex":0.000018699664,"about_ca_system_score_gemma":0.000026183672,"threshold_uncertainty_score":0.1916918},"labels":[],"label_agreement":null},{"id":"W2106961328","doi":"10.1109/iri.2011.6009601","title":"Model based analysis of internet filtering systems","year":2011,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Correctness; Computer science; The Internet; Variety (cybernetics); Perspective (graphical); Set (abstract data type); Vulnerability (computing); Computer security; Attack patterns; Data mining; Software engineering; Intrusion detection system; Artificial intelligence; World Wide Web; Algorithm","score_opus":0.08747088800441173,"score_gpt":0.2857716920213697,"score_spread":0.198300804016958,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2106961328","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.035151247,0.000021856145,0.96099824,0.000015805737,0.0000205837,0.00003810728,0.000001593047,0.00006915638,0.003683402],"genre_scores_gemma":[0.94784546,0.0000019336599,0.05164788,0.000023264549,0.0000024556923,0.000006334484,0.0000015304903,0.0000021384772,0.00046898355],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99904156,0.000055690874,0.0002375564,0.00024333637,0.00027302303,0.000148856],"domain_scores_gemma":[0.9990693,0.0000856562,0.000050740015,0.00060638896,0.00012352131,0.00006439115],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004223338,0.00006360186,0.00023259198,0.00045716477,0.000015948899,0.00004459998,0.0007771459,0.000032719006,0.000121761856],"category_scores_gemma":[0.00004404498,0.000050160495,0.00018943923,0.0013271536,0.0000313371,0.00016474986,0.00017323621,0.000052293493,0.000012243703],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000031797248,0.0007266309,0.10422535,0.0002768415,0.0033281427,0.00001729797,0.0034644639,0.72868013,0.0026951653,0.14068791,0.0006067477,0.015259495],"study_design_scores_gemma":[0.000043749456,0.000018741104,0.0015256254,0.0000067071655,0.000054580138,1.3328182e-7,0.000012671458,0.9959131,0.0022074033,0.00014827387,0.00001243814,0.000056588182],"about_ca_topic_score_codex":0.00096778496,"about_ca_topic_score_gemma":0.000035006426,"teacher_disagreement_score":0.9126942,"about_ca_system_score_codex":0.000018951847,"about_ca_system_score_gemma":0.000036176425,"threshold_uncertainty_score":0.20454857},"labels":[],"label_agreement":null},{"id":"W2109245735","doi":"10.1109/issre.1996.558895","title":"Data partition based reliability modeling","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"IBM (Canada)","funders":"","keywords":"Software quality; Computer science; Reliability (semiconductor); Partition (number theory); Reliability engineering; Data mining; Software; Data modeling; Measure (data warehouse); Software development; Mathematics; Engineering; Database","score_opus":0.17373549791487738,"score_gpt":0.32173469011956085,"score_spread":0.14799919220468347,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2109245735","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0058795256,0.000056222903,0.98715,0.0043471972,0.000038739618,0.00006902935,0.000005307591,0.00027153167,0.0021824497],"genre_scores_gemma":[0.92413133,0.000021530052,0.07530747,0.00028901067,0.000027623839,0.000006237299,0.00001510882,0.0000031333007,0.00019855847],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9984665,0.00011391782,0.00020355638,0.0005677836,0.00041460464,0.0002336355],"domain_scores_gemma":[0.9971688,0.00018309831,0.000018944727,0.0024165057,0.00011314141,0.00009955401],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009425252,0.00007224817,0.000105625186,0.00006539032,0.00011817796,0.00015900248,0.0015846591,0.000043001808,0.0005558252],"category_scores_gemma":[0.0005061763,0.0000595825,0.000050185645,0.0005037322,0.00003222705,0.0008112753,0.0004621481,0.000121882986,0.00035366346],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000154044,0.0023669852,0.014811829,0.00019483751,0.0000705971,0.000033628952,0.00032701629,0.5396134,0.00036484236,0.027455896,0.04188241,0.3728631],"study_design_scores_gemma":[0.00008912574,0.000013754049,0.000049012153,0.000004199568,0.000002592317,5.110186e-7,0.000003441259,0.99451107,0.00007741052,0.0026539348,0.002516467,0.00007845492],"about_ca_topic_score_codex":0.000076114804,"about_ca_topic_score_gemma":0.00001638129,"teacher_disagreement_score":0.9182518,"about_ca_system_score_codex":0.000031325875,"about_ca_system_score_gemma":0.000025067631,"threshold_uncertainty_score":0.6085896},"labels":[],"label_agreement":null},{"id":"W2112920243","doi":"10.1109/icc.1991.162535","title":"A technique for improving software robustness to failure","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bell (Canada)","funders":"","keywords":"Computer science; Robustness (evolution); Software engineering; Software reliability testing; Software; Reliability engineering; Software quality; Software construction; Software sizing; Software development; Engineering; Programming language","score_opus":0.0248302318602509,"score_gpt":0.2689345462851377,"score_spread":0.24410431442488678,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2112920243","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00052856555,0.000028131599,0.9947424,0.003597847,0.000030645086,0.0005628889,0.0000019093934,0.00038448282,0.00012312172],"genre_scores_gemma":[0.1192808,0.0000015955417,0.8781261,0.00031258113,0.00004637299,0.00043919208,8.823216e-7,0.000009339571,0.0017831571],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99870074,0.000033274267,0.0001703859,0.0004582727,0.00026841104,0.00036892298],"domain_scores_gemma":[0.9986905,0.00026180994,0.000023964138,0.000648771,0.00021496463,0.0001600226],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004188876,0.0001070763,0.00015304271,0.0001760779,0.0001596071,0.00020693018,0.000990938,0.00007672503,0.0001243337],"category_scores_gemma":[0.00064008427,0.00008891874,0.00012106562,0.00078904454,0.000020394951,0.00030623924,0.000313413,0.000107983054,0.00006157321],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011044023,0.0004337218,0.0015299913,0.00039535566,0.000057554906,0.00001983729,0.00067846954,0.00712771,0.013596304,0.014396559,0.040880356,0.9208731],"study_design_scores_gemma":[0.0005665352,0.0004660388,0.00015652449,0.00005055861,0.000016401937,0.000032927644,0.00008358692,0.9082773,0.049943354,0.0037921462,0.03575466,0.0008600171],"about_ca_topic_score_codex":0.000031560725,"about_ca_topic_score_gemma":0.000029131948,"teacher_disagreement_score":0.92001307,"about_ca_system_score_codex":0.000050667517,"about_ca_system_score_gemma":0.000026036701,"threshold_uncertainty_score":0.36260012},"labels":[],"label_agreement":null},{"id":"W2114890558","doi":"10.1109/ase.2009.24","title":"State-Space Coverage Estimation","year":2009,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Reachability; State space; Graph; Software; Model checking; Pathfinder; Java; State (computer science); Theoretical computer science; Abstraction; Algorithm; Programming language; Statistics; Mathematics","score_opus":0.012293870489719567,"score_gpt":0.28470271768185934,"score_spread":0.27240884719213976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2114890558","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014642176,0.000022424181,0.97556514,0.0042516035,0.000018784154,0.000045737692,2.0507301e-7,0.00018707327,0.005266849],"genre_scores_gemma":[0.9389504,0.000019891691,0.058924392,0.00039439357,0.000007335768,0.0000012120039,0.0000010825646,0.0000010952929,0.0017001622],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9992674,0.00003591429,0.00009102583,0.00018559031,0.0002667399,0.00015333811],"domain_scores_gemma":[0.9994348,0.00009001853,0.000016387065,0.00033899723,0.000056244276,0.00006354586],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00024311358,0.000048399914,0.00006936212,0.000064061125,0.00006087962,0.00017180428,0.00035136667,0.000018759969,0.000047637488],"category_scores_gemma":[0.00011180333,0.000039005692,0.0000398645,0.00040788142,0.000011427181,0.00040742403,0.00004509368,0.000064407526,0.00021932412],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000029968728,0.0001028404,0.00055122294,0.0000046473156,0.000007253918,0.000009357241,0.0002201714,0.013506901,0.0003717354,0.07398293,0.0025477405,0.9086922],"study_design_scores_gemma":[0.00013013846,0.00008711066,0.008443893,0.0000039898546,0.0000012213455,0.0000023227688,0.00000296069,0.88386935,0.003152107,0.1019619,0.0022420802,0.00010294224],"about_ca_topic_score_codex":0.000024187759,"about_ca_topic_score_gemma":0.000002511074,"teacher_disagreement_score":0.92430824,"about_ca_system_score_codex":0.000026936974,"about_ca_system_score_gemma":0.000035632536,"threshold_uncertainty_score":0.2819041},"labels":[],"label_agreement":null},{"id":"W2114894079","doi":"10.1109/imcsit.2010.5680042","title":"Selecting the best strategy in a software certification process","year":2010,"lang":"en","type":"article","venue":"Proceedings of the International Multiconference on Computer Science and Information Technology","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Laurentian University; McMaster University","funders":"","keywords":"Certification; Computer science; Pairwise comparison; Software; Process (computing); Software engineering; Software quality; Reliability engineering; Software development; Engineering; Artificial intelligence; Operating system","score_opus":0.017528647768537117,"score_gpt":0.2822376936644438,"score_spread":0.26470904589590666,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2114894079","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9655617,0.0000027013298,0.021359583,0.011979777,0.00021808351,0.00024056203,0.000001104286,0.000073247305,0.0005632629],"genre_scores_gemma":[0.9932516,0.000009163193,0.006468336,0.00021739848,0.000016595643,0.00002993982,3.0068045e-7,0.0000014138561,0.0000052125247],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9985612,0.0000039971433,0.00032575266,0.00022581269,0.00069081277,0.00019242708],"domain_scores_gemma":[0.99771625,0.00009616922,0.00022921093,0.0002135705,0.0017135089,0.000031303138],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009668163,0.00008896144,0.00009464935,0.000606792,0.0002568955,0.00049873587,0.0031213365,0.000073734904,0.0000023965335],"category_scores_gemma":[0.001459902,0.000054274737,0.000023334027,0.0017354552,0.00059184164,0.0028038772,0.00054214493,0.00044428147,0.0000071324903],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009019936,0.00010952635,0.04542668,0.000035909245,0.000009533394,8.964564e-8,0.0012982725,0.00041402245,0.009959253,0.33931527,0.000027993347,0.60339445],"study_design_scores_gemma":[0.0002065078,0.00008701621,0.013993537,0.000067516936,0.0000019230883,0.000020618309,0.00046198585,0.9005072,0.06519561,0.019059375,0.00028501317,0.000113697206],"about_ca_topic_score_codex":0.000029850613,"about_ca_topic_score_gemma":0.000006823458,"teacher_disagreement_score":0.9000932,"about_ca_system_score_codex":0.000045321947,"about_ca_system_score_gemma":0.0002205497,"threshold_uncertainty_score":0.58002716},"labels":[],"label_agreement":null},{"id":"W2116398478","doi":"10.1109/aero.1996.499667","title":"Software reliability assurance in early development phases: a case study in an industrial setting","year":2002,"lang":"en","type":"article","venue":"1996 IEEE Aerospace Applications Conference. Proceedings","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Software quality assurance; Software quality analyst; Software quality; Computer science; Quality assurance; Software engineering; Task (project management); Reliability (semiconductor); Software development; Software; Process (computing); Product (mathematics); Work (physics); Scale (ratio); Quality (philosophy); Systems engineering; Process management; Engineering; Operations management; Operating system","score_opus":0.07034565830570041,"score_gpt":0.3118235505199298,"score_spread":0.2414778922142294,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2116398478","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96014315,0.000035383076,0.036894057,0.00046062501,0.00003596963,0.002065253,0.000003957569,0.0002809007,0.000080687976],"genre_scores_gemma":[0.97244436,0.000014564243,0.025017256,0.00004084325,0.000058478934,0.002323577,0.000003034177,0.00001795649,0.00007990762],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9961091,0.0001029276,0.0008597538,0.0014618307,0.00074142526,0.0007249285],"domain_scores_gemma":[0.9977823,0.00031228358,0.00023265828,0.00077441067,0.00062577013,0.00027260516],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018277157,0.0003434049,0.0004750814,0.00046262285,0.00038082604,0.0006968246,0.0014273588,0.00022074807,0.000032885848],"category_scores_gemma":[0.0004558871,0.00035467822,0.00006698471,0.00325229,0.00014075423,0.0015687868,0.00025574325,0.00081846345,0.000051857383],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000111575455,0.0020843684,0.91871303,0.000043942415,0.000015735875,0.00007810211,0.018506674,0.00008165002,0.00014916211,0.00019251792,0.000121518606,0.06000211],"study_design_scores_gemma":[0.025912771,0.0039064595,0.4781951,0.0013947627,0.00018045494,0.0009519804,0.074279234,0.38306805,0.009381519,0.0064950953,0.007360403,0.008874175],"about_ca_topic_score_codex":0.0014075506,"about_ca_topic_score_gemma":0.0022564027,"teacher_disagreement_score":0.44051796,"about_ca_system_score_codex":0.00037884744,"about_ca_system_score_gemma":0.00036469285,"threshold_uncertainty_score":0.9998905},"labels":[],"label_agreement":null},{"id":"W2118006046","doi":"10.1109/hase.1998.731597","title":"Software component independence","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Computer science; Component (thermodynamics); Independence (probability theory); Markov chain; Markov model; Markov process; Software; Software quality; Continuation; Reliability (semiconductor); Theoretical computer science; Programming language; Reliability engineering; Software development; Engineering; Mathematics; Machine learning; Statistics","score_opus":0.03278174749652785,"score_gpt":0.2596420986004136,"score_spread":0.22686035110388575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2118006046","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.007511181,0.00020645751,0.98527056,0.0017405263,0.00006828274,0.00006536654,4.3221493e-7,0.00042652187,0.004710691],"genre_scores_gemma":[0.9236283,0.00006096442,0.07293744,0.00040751975,0.000027069758,0.000008524125,5.4798005e-7,0.0000038454173,0.0029258088],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99861693,0.000058988062,0.00014405033,0.0003299662,0.0005910063,0.00025905337],"domain_scores_gemma":[0.9989429,0.0001857393,0.000021219113,0.00062775373,0.00009997189,0.00012238842],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0002688989,0.00007745254,0.00010956468,0.00009265843,0.00011420101,0.00015272558,0.0010189746,0.00005057811,0.0010428266],"category_scores_gemma":[0.00016450092,0.000062607905,0.00007756199,0.00050510373,0.000044678494,0.0003511487,0.0003222584,0.00017544457,0.0015330015],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000023997377,0.00062448985,0.037952743,0.00003807869,0.00006455698,0.000098244476,0.0007180337,0.00050634966,0.00038843666,0.0569449,0.022227576,0.8804342],"study_design_scores_gemma":[0.00068307726,0.00017617541,0.019775245,0.000031842483,0.0000092766695,0.000072736955,0.000048670292,0.90177196,0.0036120515,0.018225744,0.054827303,0.00076592196],"about_ca_topic_score_codex":0.000055347013,"about_ca_topic_score_gemma":0.0000067575884,"teacher_disagreement_score":0.9161171,"about_ca_system_score_codex":0.00003147098,"about_ca_system_score_gemma":0.0000127364,"threshold_uncertainty_score":0.99987036},"labels":[],"label_agreement":null},{"id":"W2118247093","doi":"10.5555/381473.381511","title":"Theory of software reliability based on components","year":2001,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":133,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Component (thermodynamics); Computer science; Software quality; Reliability engineering; Reliability theory; Reliability (semiconductor); Certification; Component-based software engineering; Software; Software system; Software engineering; Software development; Systems engineering; Engineering; Operating system","score_opus":0.03077410997373903,"score_gpt":0.2789169840823954,"score_spread":0.24814287410865637,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2118247093","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11843893,0.000013752732,0.8776364,0.0005607979,0.00004814444,0.000104633524,0.0000016329969,0.00018576797,0.0030099791],"genre_scores_gemma":[0.9581689,0.000007682054,0.041037686,0.00032861298,0.000011459135,0.0000068991158,0.000002977549,0.00000450287,0.00043125922],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9981919,0.0003092334,0.00025665236,0.00039430257,0.0006072859,0.00024063695],"domain_scores_gemma":[0.9971898,0.0012939537,0.000051407325,0.0011706526,0.0001893445,0.00010483175],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013108151,0.000103516635,0.0002032941,0.00014826105,0.00007579493,0.000037692902,0.0009348294,0.000058129695,0.00029982877],"category_scores_gemma":[0.0006817822,0.00007849115,0.0001592309,0.00070065795,0.00011913302,0.00015765392,0.00016157678,0.00014098652,0.00009727894],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003868962,0.0038100274,0.47750846,0.00019076289,0.00008189712,0.00004971904,0.00029388606,0.021739187,0.0010816475,0.08110396,0.0014988342,0.41225472],"study_design_scores_gemma":[0.00130943,0.0005893499,0.16274014,0.000093443174,0.0000165964,0.00000501181,0.000039539,0.7263639,0.008166775,0.09502694,0.0051402794,0.0005085835],"about_ca_topic_score_codex":0.000055295153,"about_ca_topic_score_gemma":0.0000024332119,"teacher_disagreement_score":0.83972996,"about_ca_system_score_codex":0.000043858803,"about_ca_system_score_gemma":0.00006685261,"threshold_uncertainty_score":0.32829148},"labels":[],"label_agreement":null},{"id":"W2119853816","doi":"10.1109/arms.1990.67946","title":"Radio network modeling using criticality and conditional probability theory","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Military College of Canada","funders":"","keywords":"Criticality; Reliability block diagram; Computer science; Simple (philosophy); Reliability (semiconductor); Maintainability; Block diagram; Theoretical computer science; Interpretation (philosophy); Failure mode, effects, and criticality analysis; Reliability theory; Fault tree analysis; Algorithm; Reliability engineering; Programming language; Engineering","score_opus":0.08803034541937693,"score_gpt":0.2953829979042292,"score_spread":0.2073526524848523,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2119853816","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.063186616,0.00030414524,0.93481934,0.0006867713,0.000027364837,0.00008980848,0.0000012376886,0.00010976134,0.00077498076],"genre_scores_gemma":[0.9057772,0.000017251703,0.09386614,0.00020651688,0.000066878965,0.0000047341823,9.789666e-7,0.0000028932036,0.00005738146],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99845296,0.0003064299,0.00021019208,0.00040587757,0.00031593474,0.00030860855],"domain_scores_gemma":[0.9987711,0.00054692174,0.000014619389,0.00039271527,0.00013483282,0.00013982576],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014416012,0.00008719618,0.00015153506,0.000035047236,0.0002569784,0.00017073787,0.00029443912,0.000054328837,0.0002996332],"category_scores_gemma":[0.0003784454,0.00007310077,0.000067288725,0.0003103411,0.00017398529,0.00037711216,0.00021431106,0.00015270407,0.00001498761],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000042409533,0.00015041487,0.004989919,0.000041588915,0.000033781696,0.000005341776,0.00012901337,0.07767704,0.000027198841,0.90813714,0.00016871493,0.0086355945],"study_design_scores_gemma":[0.00005384587,0.0000071847353,0.00017673199,0.0000041334347,0.000003532476,0.0000069884823,0.0000045968436,0.62775517,0.000007264983,0.3718935,0.000028262391,0.000058807647],"about_ca_topic_score_codex":0.000026238835,"about_ca_topic_score_gemma":0.0000024716148,"teacher_disagreement_score":0.84259063,"about_ca_system_score_codex":0.000046056844,"about_ca_system_score_gemma":0.000025225921,"threshold_uncertainty_score":0.32807732},"labels":[],"label_agreement":null},{"id":"W2123554630","doi":"10.1109/cmpsac.1993.404246","title":"An integrated environment for software reliability modeling","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science; IBM; Software engineering; Reliability (semiconductor); Software quality; Software; Software development; Operating system","score_opus":0.04141129552390837,"score_gpt":0.2739027051855415,"score_spread":0.23249140966163312,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2123554630","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.014629541,0.000067532455,0.983832,0.0007662597,0.0000359841,0.00024077487,0.0000042518136,0.00034864273,0.00007499285],"genre_scores_gemma":[0.51142186,0.000044942088,0.48787194,0.0001378345,0.00002580825,0.0000622203,0.0000076395,0.0000079677675,0.00041980698],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99830365,0.00009836886,0.00026801063,0.00063138286,0.00035146505,0.00034712697],"domain_scores_gemma":[0.99835235,0.00022480158,0.000026281472,0.0011011377,0.00011806515,0.00017733609],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006776453,0.00012993146,0.00017625537,0.00008398508,0.00018105908,0.0001629042,0.00092431076,0.00007923004,0.00038071838],"category_scores_gemma":[0.00031821185,0.000102799466,0.00013866469,0.0002673474,0.000047222024,0.0005492282,0.00012437171,0.00013878771,0.00012458065],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014014202,0.0010991768,0.0029630987,0.000055296157,0.000040873147,0.0000033236104,0.00060737977,0.33505693,0.00018602374,0.003448959,0.0012478795,0.655277],"study_design_scores_gemma":[0.00015649374,0.000114843846,0.000039253187,0.0000036653846,0.000004176355,9.4386047e-7,0.00002709333,0.9898493,0.0001752306,0.0038085156,0.0056820605,0.0001383918],"about_ca_topic_score_codex":0.00009533224,"about_ca_topic_score_gemma":0.000006339358,"teacher_disagreement_score":0.6551387,"about_ca_system_score_codex":0.000102222984,"about_ca_system_score_gemma":0.00002037685,"threshold_uncertainty_score":0.41920406},"labels":[],"label_agreement":null},{"id":"W2123971618","doi":"10.1109/syscon.2011.5929087","title":"A model driven framework for N-version programming","year":2011,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Royal Military College of Canada","funders":"","keywords":"Computer science; Fault tolerance; Software engineering; Software fault tolerance; Software development; Unified Modeling Language; Software quality; Redundancy (engineering); Software; Reliability engineering; Programming language; Distributed computing; Operating system","score_opus":0.0885224268306274,"score_gpt":0.32002976741441025,"score_spread":0.23150734058378286,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2123971618","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0027293009,0.000011789931,0.99572366,0.00027040162,0.000027569904,0.00017035907,4.7353427e-7,0.00018547969,0.00088094256],"genre_scores_gemma":[0.2823677,0.0000038860094,0.7172577,0.00007301627,0.000010939357,0.000030355504,6.0425685e-7,0.0000028454044,0.00025298572],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99917483,0.000017830924,0.00010275789,0.0002709041,0.00019401492,0.00023965983],"domain_scores_gemma":[0.9992501,0.00013951097,0.000021601418,0.0003960838,0.00010814001,0.00008455705],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002479485,0.00006005172,0.000089001995,0.00006692469,0.00010956125,0.00006682431,0.0006349332,0.00006287592,0.000033293694],"category_scores_gemma":[0.00016736047,0.0000476778,0.00010947997,0.0002794423,0.000028916647,0.0002437187,0.00019296423,0.000087096174,0.000049670587],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001788009,0.00026854422,0.0034308722,0.00004231263,0.00003979963,0.0000026121852,0.002088313,0.00063471374,0.000070792994,0.7572263,0.0008075751,0.23537028],"study_design_scores_gemma":[0.00007205485,0.00005699695,0.000068688176,0.000008976416,0.0000034539753,4.443409e-7,0.000029550365,0.8495497,0.00054116524,0.14866051,0.0009360839,0.000072358904],"about_ca_topic_score_codex":0.000024883058,"about_ca_topic_score_gemma":0.000003979303,"teacher_disagreement_score":0.848915,"about_ca_system_score_codex":0.000020448539,"about_ca_system_score_gemma":0.000046584006,"threshold_uncertainty_score":0.19442444},"labels":[],"label_agreement":null},{"id":"W2124044559","doi":"10.1145/337180.337194","title":"Broad-spectrum studies of log file analysis","year":2000,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":56,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Computer science; File system; Unix file types; Completeness (order theory); Web log analysis software; File format; Operating system; Database; Computer file; Stub file; Mathematics","score_opus":0.030501907914484854,"score_gpt":0.32187157756020407,"score_spread":0.2913696696457192,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124044559","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37681484,0.004730905,0.54001904,0.010236471,0.0001228068,0.0004289409,0.00009556551,0.000974471,0.06657695],"genre_scores_gemma":[0.9702853,0.00056699733,0.015536947,0.00013262268,0.000022873135,0.000011478669,0.000008828238,0.000003341405,0.013431624],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99878424,0.00008085256,0.00024055384,0.0003096252,0.00036906343,0.00021565038],"domain_scores_gemma":[0.99870414,0.00043791003,0.00003278518,0.00066658965,0.000099301746,0.000059305698],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00032935187,0.00007840748,0.00031595601,0.00025002882,0.00007094901,0.000040373172,0.0006625595,0.00003130395,0.013509878],"category_scores_gemma":[0.00011523696,0.00005833444,0.0002651516,0.0027755357,0.00010935918,0.00017777956,0.0001364957,0.00007214542,0.0002131217],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000036963014,0.0011489409,0.052719973,0.00015335159,0.011566387,0.000047295984,0.00352164,0.022101605,0.00024350984,0.031966466,0.08098811,0.79550576],"study_design_scores_gemma":[0.0010432198,0.00073570583,0.13352342,0.00006338347,0.0010041038,0.00001046612,0.00079156714,0.713488,0.016234698,0.06618122,0.06561096,0.0013132703],"about_ca_topic_score_codex":0.00024272894,"about_ca_topic_score_gemma":0.000115139825,"teacher_disagreement_score":0.7941925,"about_ca_system_score_codex":0.000023847617,"about_ca_system_score_gemma":0.00003215119,"threshold_uncertainty_score":0.9873919},"labels":[],"label_agreement":null},{"id":"W2128454637","doi":"10.1007/s10009-005-0209-6","title":"Software tools for safety-critical software development","year":2005,"lang":"en","type":"article","venue":"International Journal on Software Tools for Technology Transfer","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":24,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Software development; Software engineering; Package development process; Variety (cybernetics); Software construction; Software peer review; Software analytics; Software; Social software engineering; Personal software process; Verification and validation; Avionics software; Operating system; Engineering; Artificial intelligence","score_opus":0.03718945379551026,"score_gpt":0.33089255944516616,"score_spread":0.2937031056496559,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2128454637","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008360988,0.00033634115,0.9668025,0.021531437,0.0009785263,0.0006901661,0.00019860492,0.0010805826,0.000020887774],"genre_scores_gemma":[0.113939494,0.00018593746,0.8817133,0.0023421384,0.0006891686,0.00055064616,0.00011240861,0.00008011149,0.00038682594],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9951978,0.000083791696,0.0012791392,0.0009898549,0.0013839456,0.0010654671],"domain_scores_gemma":[0.9919865,0.004662997,0.00011325849,0.0006880027,0.0022277134,0.00032151953],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0012790665,0.00049650954,0.00063854613,0.001154801,0.0007831091,0.0012750103,0.0038579134,0.0005964128,0.000159598],"category_scores_gemma":[0.011769003,0.00045244355,0.0007027183,0.00060598104,0.00031140196,0.002363799,0.00024119725,0.0010549608,0.00014839313],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00035114164,0.0005037458,0.0007329546,0.000043802625,0.00044727104,0.000048815644,0.00016261687,0.00064086996,0.00009666994,0.031986203,0.0012771724,0.96370876],"study_design_scores_gemma":[0.004965976,0.0011433945,0.0010021603,0.0004557436,0.00009423805,0.0007083658,0.00009494412,0.0017122832,0.018205376,0.043338608,0.92697287,0.0013060567],"about_ca_topic_score_codex":0.0000011944214,"about_ca_topic_score_gemma":0.000013625372,"teacher_disagreement_score":0.9624027,"about_ca_system_score_codex":0.0007684914,"about_ca_system_score_gemma":0.00067516527,"threshold_uncertainty_score":0.99979275},"labels":[],"label_agreement":null},{"id":"W2129084475","doi":"10.1145/1385269.1385283","title":"A case study in database reliability","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University; IBM (Canada)","funders":"","keywords":"Reliability (semiconductor); Computer science; Database; Reputation; Work (physics); Engineering","score_opus":0.0734899394576537,"score_gpt":0.3357681851769771,"score_spread":0.2622782457193234,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2129084475","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9281249,0.000014786101,0.070794865,0.00021526191,0.000025594249,0.00023935763,6.8102406e-7,0.00011834753,0.00046622727],"genre_scores_gemma":[0.98846906,0.000008968454,0.0112087745,0.000058198206,0.0000102775575,0.0000250143,5.692591e-7,0.0000025498052,0.00021656457],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9983418,0.00026306583,0.00025407504,0.0005157119,0.00037640615,0.00024896784],"domain_scores_gemma":[0.998194,0.00034568008,0.00001821469,0.0012479445,0.00008591551,0.00010829356],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011776822,0.000083638326,0.00016057763,0.000166098,0.00013099947,0.00004090979,0.0005898267,0.000026593372,0.000062482606],"category_scores_gemma":[0.0004460233,0.000066676555,0.00005149278,0.0011940106,0.0000658343,0.00047995496,0.0004704395,0.00019189848,0.000081846585],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006471914,0.0025442715,0.94784313,0.000012636407,0.000010977266,0.038563635,0.0029625753,0.00021094352,0.000019098048,0.00046812117,0.00041598006,0.006942161],"study_design_scores_gemma":[0.005811396,0.0015322457,0.2810702,0.000030441335,0.000023148703,0.026063044,0.0077475333,0.6694892,0.00086344004,0.0034994555,0.0022278232,0.0016420876],"about_ca_topic_score_codex":0.0056152763,"about_ca_topic_score_gemma":0.0010555455,"teacher_disagreement_score":0.66927826,"about_ca_system_score_codex":0.00005665465,"about_ca_system_score_gemma":0.000118836484,"threshold_uncertainty_score":0.8488655},"labels":[],"label_agreement":null},{"id":"W2131005582","doi":"10.4018/978-1-60566-026-4.ch477","title":"An Overview of Software Engineering Process and Its Improvement","year":2009,"lang":"en","type":"book-chapter","venue":"IGI Global eBooks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Software Engineering Process Group; Software engineering; Personal software process; Social software engineering; Software construction; Standardization; Software development; Software requirements; Software; Software development process; Process (computing); Computer science; Engineering; Systems engineering; Operating system","score_opus":0.028833433941212062,"score_gpt":0.3001213830892294,"score_spread":0.27128794914801735,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2131005582","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.027247265,0.10391665,0.20018633,0.00069703377,0.0013466999,0.0073951306,0.00089965505,0.0045558303,0.6537554],"genre_scores_gemma":[0.9575029,0.000854521,0.022972416,0.000810855,0.0004370845,0.00007379982,0.000016712109,0.000115400275,0.017216314],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99802923,0.000012710006,0.00038580727,0.00062107743,0.0006458962,0.00030527156],"domain_scores_gemma":[0.99864477,0.00004680399,0.00015008866,0.0006674357,0.0002777247,0.00021316751],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00023093105,0.00031019232,0.00049672055,0.00010139268,0.000052242718,0.00011628746,0.0009013799,0.0002314036,0.000007834731],"category_scores_gemma":[0.00006274739,0.0002945578,0.00015390624,0.000055859662,0.000034509776,0.00015302715,0.00023153253,0.00023247111,0.0000084723115],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000053621816,0.000022211145,0.00003482766,0.00061027904,0.00007946524,0.000021054808,0.00006200982,0.00010322105,0.000065223154,0.82452404,0.000015384183,0.1744569],"study_design_scores_gemma":[0.001331839,0.0025881024,0.00094866974,0.0033454546,0.00029546136,0.00009410616,0.000014287936,0.0379781,0.0041489345,0.93683606,0.0096064815,0.0028125243],"about_ca_topic_score_codex":0.000017249762,"about_ca_topic_score_gemma":0.0000074603336,"teacher_disagreement_score":0.93025565,"about_ca_system_score_codex":0.00010170731,"about_ca_system_score_gemma":0.00019817211,"threshold_uncertainty_score":0.99995065},"labels":[],"label_agreement":null},{"id":"W2131072500","doi":"","title":"Teaching Software Quality Assurance in an Undergraduate Software Engineering Program","year":2007,"lang":"en","type":"article","venue":"Software Quality Professional Magazine","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Software quality assurance; Software quality analyst; Quality assurance; Software quality control; Software quality; Software engineering; Software peer review; Engineering management; Social software engineering; Personal software process; Software; Software construction; Verification and validation; Software development; Computer science; Engineering; Operations management; Operating system","score_opus":0.03615835656803074,"score_gpt":0.38738873109572486,"score_spread":0.3512303745276941,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2131072500","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.30413988,0.00027701876,0.6892353,0.0018476356,0.0007233547,0.00097256806,0.00004056212,0.0027168589,0.000046768306],"genre_scores_gemma":[0.4431427,0.000017880497,0.5541296,0.00063659274,0.00029470606,0.00021437134,0.00018005059,0.0000848777,0.0012992274],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98949826,0.0018958297,0.002172987,0.0019484315,0.0025641047,0.0019203745],"domain_scores_gemma":[0.9895304,0.00623974,0.00052347104,0.0021685981,0.00073860906,0.00079920073],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.014189245,0.0007579346,0.0011293596,0.00074299006,0.00065233983,0.0003784834,0.0025986768,0.00052625424,0.00004734059],"category_scores_gemma":[0.014196715,0.0007060398,0.0004396466,0.0021518157,0.00022737238,0.0022461407,0.0011851201,0.0024055922,0.000172438],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00020966162,0.0029686317,0.62557167,0.0013612498,0.000101216654,0.00015975846,0.0012340951,0.0010283762,0.0007274748,0.0135473795,0.00054715073,0.35254332],"study_design_scores_gemma":[0.0019270087,0.0003669468,0.9666416,0.00076429796,0.00002016747,0.000025290408,0.00019608215,0.0035307321,0.0003079405,0.017467296,0.007038271,0.0017143439],"about_ca_topic_score_codex":0.00048248842,"about_ca_topic_score_gemma":0.0011428465,"teacher_disagreement_score":0.35082898,"about_ca_system_score_codex":0.0005613239,"about_ca_system_score_gemma":0.0006713974,"threshold_uncertainty_score":0.9998959},"labels":[],"label_agreement":null},{"id":"W2134421244","doi":"10.1287/inte.1040.0067","title":"General Motors Optimizes Its Scheduling of Cold-Weather Tests","year":2004,"lang":"en","type":"article","venue":"INFORMS Journal on Applied Analytics","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"General Motors (Canada)","funders":"Oregon State University","keywords":"Warranty; Scheduling (production processes); Schedule; Operations research; Cold weather; Engineering; General motors; Computer science; Automotive engineering; Transport engineering; Operations management; Operating system","score_opus":0.023779067967838047,"score_gpt":0.28469505830260505,"score_spread":0.260915990334767,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2134421244","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.34937203,0.00013582807,0.6435668,0.0007976044,0.00016962882,0.00023299099,0.0000040319933,0.0001061368,0.005614983],"genre_scores_gemma":[0.90811974,0.00022923005,0.09091591,0.00032644573,0.00014408909,0.0000042256747,0.0000010842882,0.000013401404,0.0002458871],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976215,0.000016283795,0.00069576665,0.00023065726,0.0010030817,0.00043275513],"domain_scores_gemma":[0.9984209,0.00013829411,0.00033350557,0.0004695075,0.00035005296,0.00028775365],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008424954,0.00020240547,0.00039039578,0.00047310695,0.00019534607,0.00030204156,0.0010776003,0.00012128168,0.000028224149],"category_scores_gemma":[0.00019460698,0.00014842616,0.00027145082,0.0010166955,0.000086393324,0.0004087095,0.00017123445,0.0005672807,0.00009166482],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029765199,0.00021762389,0.00045265097,0.000028300621,0.00018736585,0.000027433194,0.0002750313,0.8418433,0.0021578623,0.14674613,0.000040164596,0.007994373],"study_design_scores_gemma":[0.008619471,0.0015885769,0.0024387292,0.00052210793,0.0002213562,0.0002586461,0.0004397042,0.7559995,0.15891665,0.060704965,0.008351039,0.0019392676],"about_ca_topic_score_codex":0.0000066882512,"about_ca_topic_score_gemma":0.0000016039063,"teacher_disagreement_score":0.5587477,"about_ca_system_score_codex":0.00019901573,"about_ca_system_score_gemma":0.00035279995,"threshold_uncertainty_score":0.60526437},"labels":[],"label_agreement":null},{"id":"W2134754630","doi":"10.1109/csmr.2004.1281425","title":"SM/sup CMM/ model to evaluate and improve the quality of the software maintenance process","year":2004,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Software maintenance; Computer science; Capability Maturity Model; Software engineering; LeanCMMI; Software quality; Software development process; Process (computing); Verification and validation; Software; Software development; Systems engineering; Reliability engineering; Engineering; Operating system; Operations management","score_opus":0.04360111583045167,"score_gpt":0.3548708212753319,"score_spread":0.31126970544488025,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2134754630","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.3066653,0.00004432469,0.6846092,0.008136602,0.000023444562,0.00030713365,0.0000041388876,0.00005048596,0.0001593774],"genre_scores_gemma":[0.98323774,0.000011157658,0.015374711,0.0008111128,0.000009604166,0.000038056438,1.7210581e-7,0.0000045611696,0.00051290495],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9981401,0.00013408082,0.00029334115,0.00040699876,0.0007366999,0.00028876908],"domain_scores_gemma":[0.9981561,0.000246349,0.000074766336,0.0010069901,0.00042601483,0.000089822985],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0018579473,0.000110661,0.00018495096,0.000035332185,0.00017768693,0.000103075516,0.0014732069,0.000042883315,0.0000055551754],"category_scores_gemma":[0.0012857318,0.00005141848,0.00010601378,0.00067703886,0.0001866086,0.00020932984,0.0006261491,0.00017358661,0.000008679687],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000105423096,0.00059923093,0.01814179,0.0006037519,0.0001841262,0.0000020688262,0.016321639,0.6771231,0.0049969666,0.13855757,0.0006158727,0.1427484],"study_design_scores_gemma":[0.0008492289,0.00012390969,0.013950392,0.000082368126,0.000020029529,0.000004516273,0.00041830735,0.658252,0.014137323,0.31174827,0.00008657742,0.00032707982],"about_ca_topic_score_codex":0.00033887094,"about_ca_topic_score_gemma":0.0001043494,"teacher_disagreement_score":0.67657244,"about_ca_system_score_codex":0.0000493248,"about_ca_system_score_gemma":0.0002923389,"threshold_uncertainty_score":0.27376094},"labels":[],"label_agreement":null},{"id":"W2136555272","doi":"10.1016/j.ress.2004.08.024","title":"An effective technique for the software requirements analysis of NPP safety-critical systems, based on software inspection, requirements traceability, and formal specification","year":2004,"lang":"en","type":"article","venue":"Reliability Engineering & System Safety","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Atomic Energy of Canada Limited","keywords":"Traceability; Requirements traceability; Computer science; Software requirements specification; Software requirements; Formal specification; Verification and validation; Software engineering; Requirements analysis; Formal methods; Software system; Reliability engineering; Software construction; Functional requirement; Software; Systems engineering; Requirement; Engineering; Programming language","score_opus":0.014854686408639926,"score_gpt":0.28640427972275434,"score_spread":0.27154959331411443,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2136555272","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009572631,0.00015726472,0.985402,0.00024968854,0.00030308153,0.0032810955,0.00016073078,0.00086237915,0.0000111230465],"genre_scores_gemma":[0.94345605,0.000017053382,0.05550991,0.000011159566,0.00006986044,0.0008473244,0.00005472253,0.000031872267,0.0000020345858],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9951374,0.0004645873,0.0013949763,0.0012117687,0.0011375548,0.00065374403],"domain_scores_gemma":[0.9927017,0.0034608457,0.00026327334,0.0023580526,0.0009579609,0.0002581844],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0070090997,0.00042376333,0.0009218546,0.0006010076,0.00061891176,0.00019572752,0.0010516288,0.00026452693,0.0000030253223],"category_scores_gemma":[0.003373835,0.00034059957,0.0004582393,0.002309729,0.00022398097,0.00087015197,0.00013980216,0.00037486674,0.0000025858228],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002966164,0.00061511237,0.014018298,0.0028229484,0.0005152958,0.0000024199376,0.00033075863,0.9594221,0.00081666413,0.018297158,0.000004679384,0.0028579081],"study_design_scores_gemma":[0.0009777925,0.000621723,0.10345786,0.000540679,0.0005267695,0.0000067988212,0.00013912782,0.89152473,0.0014057158,0.0002138678,0.0001367714,0.00044816628],"about_ca_topic_score_codex":0.00029259568,"about_ca_topic_score_gemma":0.000018187253,"teacher_disagreement_score":0.9338834,"about_ca_system_score_codex":0.0017083556,"about_ca_system_score_gemma":0.00017070002,"threshold_uncertainty_score":0.9999046},"labels":[],"label_agreement":null},{"id":"W2136805049","doi":"10.1109/iceccs.2004.1310912","title":"Independence transformations and tools for components","year":2004,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Independence (probability theory); Computer science; Component (thermodynamics); Continuation; Set (abstract data type); Reliability (semiconductor); Property (philosophy); Composition (language); Work (physics); Programming language; Engineering; Mathematics","score_opus":0.047896311656501836,"score_gpt":0.30077164818407925,"score_spread":0.25287533652757743,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2136805049","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05438101,0.000025927722,0.9423595,0.0025287683,0.000016214792,0.00014400078,0.0000016554262,0.00006493808,0.00047797238],"genre_scores_gemma":[0.93464833,0.00002525946,0.065122806,0.00013201183,0.0000060509647,0.00002104337,0.0000024237775,0.0000012737625,0.00004079489],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9994305,0.000010539389,0.00010162128,0.00014315802,0.00018418761,0.00013002871],"domain_scores_gemma":[0.9995279,0.00017033897,0.00001002584,0.0001629793,0.00006799034,0.00006077619],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00023030922,0.000041257987,0.000065959546,0.00005284082,0.00012022451,0.00029268162,0.000275118,0.000028932714,0.000005959398],"category_scores_gemma":[0.00007874108,0.00003363865,0.000036556587,0.00017111334,0.000028772904,0.001184265,0.000043240514,0.000053024676,0.000014074631],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000008759396,0.00020288586,0.00082327117,0.00007601133,0.00004130808,0.000002233367,0.0014074397,0.0014715836,0.002134615,0.6775369,0.000087421366,0.31620756],"study_design_scores_gemma":[0.0049650664,0.00047688585,0.039520558,0.00008331273,0.000027322572,0.000057822108,0.00043449237,0.23991777,0.028575942,0.6736213,0.011501684,0.00081784825],"about_ca_topic_score_codex":0.000042874126,"about_ca_topic_score_gemma":0.00001616294,"teacher_disagreement_score":0.8802673,"about_ca_system_score_codex":0.000018059201,"about_ca_system_score_gemma":0.00003971012,"threshold_uncertainty_score":0.28223357},"labels":[],"label_agreement":null},{"id":"W2137918864","doi":"10.1109/issre.1996.558672","title":"Automatic failure detection with Conditional-Belief supervisors","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bell (Canada); University of Waterloo","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Supervisor; Computer science; Formalism (music); Finite-state machine; Class (philosophy); Control (management); Software engineering; Reliability engineering; Artificial intelligence; Programming language; Engineering","score_opus":0.014198486979831067,"score_gpt":0.21869878461973033,"score_spread":0.20450029763989927,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2137918864","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.105155624,0.00003929838,0.8903395,0.0021658463,0.000020295072,0.00010651708,7.9430146e-7,0.0003837845,0.0017883203],"genre_scores_gemma":[0.9832353,0.000006643588,0.015540912,0.00018597391,0.000019972045,0.000020692329,0.0000017610149,0.0000039264914,0.0009848183],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9988694,0.00006347439,0.00012919732,0.00027972256,0.00044943782,0.00020873656],"domain_scores_gemma":[0.9992341,0.00014984385,0.000021740143,0.0003918949,0.000109948225,0.00009246785],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00015936905,0.00008580117,0.000108464395,0.00013017561,0.00015288843,0.00015248495,0.00039147053,0.00004262497,0.0015780181],"category_scores_gemma":[0.000055060904,0.00006131381,0.000054890093,0.00071942265,0.000055221022,0.00045140722,0.0000631812,0.000116307834,0.00076357403],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007074537,0.0011216601,0.0150122065,0.00021259092,0.00036705015,0.00010121623,0.0024745704,0.0021842744,0.002252007,0.05188738,0.009427023,0.91495293],"study_design_scores_gemma":[0.0002657745,0.00015277075,0.003067395,0.000010851906,0.0000069109974,0.000044320735,0.00006823035,0.9867599,0.0020541823,0.003349066,0.004051739,0.00016886424],"about_ca_topic_score_codex":0.000048350084,"about_ca_topic_score_gemma":0.00012661246,"teacher_disagreement_score":0.9845756,"about_ca_system_score_codex":0.000038248,"about_ca_system_score_gemma":0.00001596641,"threshold_uncertainty_score":0.9993347},"labels":[],"label_agreement":null},{"id":"W2138270590","doi":"10.1109/hase.2007.39","title":"Improving Reliability and Safety by Trading off Software Failure Criticalities","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Fault tolerance; Reliability engineering; Computer science; Failure rate; Reliability (semiconductor); Voting; Software quality; Software fault tolerance; Constraint (computer-aided design); Fault (geology); Class (philosophy); Process (computing); Software; Life-critical system; Distributed computing; Software development; Engineering; Artificial intelligence; Programming language","score_opus":0.010067642597161442,"score_gpt":0.2612918160596365,"score_spread":0.25122417346247505,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2138270590","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05685803,0.0003671031,0.9393388,0.0023032285,0.000047697,0.00012539093,0.0000059259964,0.00035580405,0.00059804274],"genre_scores_gemma":[0.86661005,0.000038647682,0.13238114,0.00034406153,0.000045211138,0.0000049520513,0.000004086129,0.000009428794,0.0005624548],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99779457,0.000105856336,0.00039440137,0.0006371105,0.0004933779,0.0005746588],"domain_scores_gemma":[0.9971701,0.0017181851,0.00003374289,0.0006187703,0.00017570879,0.00028347492],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025715379,0.00015801875,0.00024198163,0.00010500235,0.00031009954,0.0002904576,0.00056706986,0.00011882478,0.00008486257],"category_scores_gemma":[0.00180642,0.00013320979,0.000098154065,0.00046119385,0.00023942228,0.0006813787,0.00033573803,0.00028980107,0.000010541134],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004782659,0.00029627234,0.051609438,0.0004282545,0.000057083293,0.000032204756,0.0013332498,0.000013103668,0.0026240575,0.04983427,0.006780276,0.88694394],"study_design_scores_gemma":[0.004795221,0.0012461591,0.058940716,0.00034285954,0.00015514357,0.00025398305,0.0070752082,0.51563483,0.06048318,0.15007351,0.19593802,0.005061194],"about_ca_topic_score_codex":0.00025374853,"about_ca_topic_score_gemma":0.000055032502,"teacher_disagreement_score":0.8818828,"about_ca_system_score_codex":0.00009663727,"about_ca_system_score_gemma":0.00006515797,"threshold_uncertainty_score":0.5432138},"labels":[],"label_agreement":null},{"id":"W2139714001","doi":"10.14736/kyb-2016-5-0757","title":"A linear programming approach to error bounds for random walks in the quarter-plane","year":2016,"lang":"en","type":"preprint","venue":"Kybernetika","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Nederlandse Organisatie voor Wetenschappelijk Onderzoek","keywords":"Random walk; Markov chain; Mathematics; Approximation error; State space; Plane (geometry); Space (punctuation); Work (physics); Product (mathematics); Markov process; Distribution (mathematics); Stationary distribution; State (computer science); Quarter (Canadian coin); Statistical physics; Applied mathematics; Mathematical analysis; Computer science; Algorithm; Statistics; Physics; Geometry","score_opus":0.03855498802795353,"score_gpt":0.3167938980303294,"score_spread":0.2782389100023759,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2139714001","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0061489055,0.00035485555,0.9795258,0.0070261867,0.00029749385,0.003304406,0.000029881461,0.00020024352,0.0031122288],"genre_scores_gemma":[0.8293752,0.00004077536,0.16232955,0.00062675687,0.00086654717,0.004968228,0.000119494565,0.00005239508,0.001621042],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963356,0.00042479596,0.0005360321,0.0011557393,0.00079705205,0.0007507752],"domain_scores_gemma":[0.9968265,0.0009749134,0.00013436556,0.0017216395,0.0001812985,0.00016123135],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003127344,0.00033534964,0.000577231,0.00030458515,0.0001425982,0.00056224107,0.0033922372,0.00030483416,0.0000073700558],"category_scores_gemma":[0.00047854966,0.00020511505,0.0003716604,0.00054911256,0.00009478927,0.00011251691,0.00095194,0.0006489051,0.00007050552],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009511916,0.0040317774,0.006554006,0.0030466358,0.0007959962,0.000086069944,0.046617415,0.005553785,0.00014443586,0.05953028,0.03127407,0.84141433],"study_design_scores_gemma":[0.008750128,0.0009880582,0.0034568629,0.0011260955,0.00014990233,0.0000660885,0.00086172327,0.33774844,0.00024388895,0.1425397,0.50131303,0.0027560645],"about_ca_topic_score_codex":0.00027165786,"about_ca_topic_score_gemma":0.000085533204,"teacher_disagreement_score":0.8386583,"about_ca_system_score_codex":0.0001157991,"about_ca_system_score_gemma":0.00024298592,"threshold_uncertainty_score":0.8364349},"labels":[],"label_agreement":null},{"id":"W2139724964","doi":"10.4236/iim.2009.13022","title":"Probabilistic Analysis of a Robot System with Redundant Safety Units and Common-Cause Failures","year":2009,"lang":"en","type":"article","venue":"Intelligent Information Management","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Common cause failure; Probabilistic logic; Computer science; Common cause and special cause; Reliability engineering; Artificial intelligence; Engineering; Operations management","score_opus":0.019420864059841243,"score_gpt":0.26080186863994764,"score_spread":0.2413810045801064,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2139724964","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008128088,0.00004139104,0.988614,0.000559537,0.000019757766,0.0004850586,0.00000663129,0.00011692317,0.0020285747],"genre_scores_gemma":[0.99169284,0.00010011184,0.008015622,0.000084924504,0.0000034306695,0.00001674564,0.000036056063,0.0000019498996,0.000048310947],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9983116,0.000078992365,0.0006211523,0.00018854706,0.00059607334,0.00020367972],"domain_scores_gemma":[0.99865437,0.00010250116,0.00021358157,0.0006103697,0.00032871444,0.000090490976],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0006281559,0.00013555544,0.00030480706,0.00085783075,0.00011304496,0.00023638525,0.0005146242,0.000033923883,0.000009186918],"category_scores_gemma":[0.000038617043,0.000102713675,0.000068804395,0.0029608335,0.000060968025,0.00077599566,0.00017404072,0.000086509026,0.000015954776],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008429212,0.00011346525,0.0016924656,0.0005263042,0.0017839413,0.000009499906,0.0022754818,0.2803748,0.0000020425186,0.47201762,0.000071344286,0.24104874],"study_design_scores_gemma":[0.00030267637,0.00031259566,0.029852865,0.00018814836,0.00066014926,0.0000063780485,0.0016219278,0.96120787,0.00015156143,0.0006798791,0.004745947,0.00027001541],"about_ca_topic_score_codex":0.00010458605,"about_ca_topic_score_gemma":0.000041424184,"teacher_disagreement_score":0.98356473,"about_ca_system_score_codex":0.00012258564,"about_ca_system_score_gemma":0.00002751157,"threshold_uncertainty_score":0.4188542},"labels":[],"label_agreement":null},{"id":"W2140081622","doi":"10.1109/icsm.2005.46","title":"Explorative study to provide decision support for software release decisions","year":2005,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Software release life cycle; Reliability (semiconductor); Computer science; Software quality; Decision support system; Domain (mathematical analysis); Software; Reliability engineering; Risk analysis (engineering); Software development; Engineering; Data mining; Business","score_opus":0.05390194114157464,"score_gpt":0.36131853987127066,"score_spread":0.307416598729696,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2140081622","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.060559165,0.0000139734,0.9357508,0.0019091377,0.000060334438,0.0013245385,0.000005676399,0.0002727934,0.000103597115],"genre_scores_gemma":[0.38350463,0.000005835176,0.6139534,0.00059320504,0.00006768634,0.0003058698,0.0000035093706,0.0000109634275,0.0015548997],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99741685,0.00010143769,0.0004381536,0.00080680376,0.00079741806,0.00043931123],"domain_scores_gemma":[0.9952518,0.0026093572,0.000045723948,0.0011293321,0.00056165806,0.00040216147],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011505394,0.00016973566,0.00027093102,0.0003147663,0.0003106739,0.0002989147,0.00132511,0.000042200867,0.00012141436],"category_scores_gemma":[0.0056771035,0.0001287607,0.00017082559,0.0010868869,0.000025182295,0.0008547611,0.0006611524,0.00012721276,0.00071045593],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000073001305,0.0012186783,0.0023249965,0.00000312015,0.000030636893,0.000009263079,0.00222848,0.0018378801,0.000021765427,0.00084966957,0.0151093425,0.97629315],"study_design_scores_gemma":[0.008600138,0.014381911,0.019238602,0.00020251103,0.00014552352,0.00003109584,0.00901888,0.3078247,0.009166131,0.05491754,0.57349306,0.0029798883],"about_ca_topic_score_codex":0.00003131841,"about_ca_topic_score_gemma":0.00033860662,"teacher_disagreement_score":0.9733133,"about_ca_system_score_codex":0.00011739348,"about_ca_system_score_gemma":0.00024948703,"threshold_uncertainty_score":0.91317105},"labels":[],"label_agreement":null},{"id":"W2143595460","doi":"10.1109/eit.2007.4374474","title":"Performance analysis of service-oriented architectures with multi-factor sensitivity analysis","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Windsor","funders":"","keywords":"Computer science; Sensitivity (control systems); Factor (programming language); Software; Statistical analysis; Generalization; Reliability engineering; Service (business); Data mining; Engineering; Mathematics; Operating system; Programming language","score_opus":0.016987478742679987,"score_gpt":0.2800379638860908,"score_spread":0.26305048514341084,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2143595460","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.50907165,0.0000069312423,0.49059817,0.00009105986,0.000007396337,0.000051810286,0.0000054047327,0.00006914972,0.00009840916],"genre_scores_gemma":[0.9440318,0.0000046701,0.055704534,0.00013579606,0.00000972314,0.0000022394222,0.0000130016515,0.0000049847354,0.00009325765],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975093,0.00015503523,0.00040596147,0.00060697494,0.0008676082,0.00045511438],"domain_scores_gemma":[0.99727017,0.0006330978,0.00015033051,0.0011130186,0.0006386618,0.0001947025],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0013181991,0.00019242447,0.0006409643,0.0021141833,0.00012568531,0.00005312013,0.0005071714,0.00007440792,0.000082418046],"category_scores_gemma":[0.00009500426,0.00013319611,0.00042053187,0.020642262,0.00010375431,0.00017935556,0.00021813504,0.00019792057,0.000008130367],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000045732737,0.00023670465,0.88931817,0.000030967796,0.006743825,0.000013894876,0.00076322,0.08763689,0.00086574187,0.00012387508,0.0000012327301,0.014219738],"study_design_scores_gemma":[0.000111849375,0.000034017332,0.49738038,0.0000028286697,0.0007132431,8.7741813e-7,0.000029081015,0.4979888,0.0036250204,0.0000015830597,0.000012447745,0.0000998924],"about_ca_topic_score_codex":0.001606927,"about_ca_topic_score_gemma":0.014454435,"teacher_disagreement_score":0.43496013,"about_ca_system_score_codex":0.000051608968,"about_ca_system_score_gemma":0.00006944407,"threshold_uncertainty_score":0.99179226},"labels":[],"label_agreement":null},{"id":"W2144309645","doi":"10.1109/issre.1994.341343","title":"A framework for reliability estimation","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":11,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Reliability (semiconductor); Computer science; Estimation; Software quality; Reliability engineering; Software; Reliability theory; Random testing; Data mining; Machine learning; Programming language; Software development; Engineering; Test case; Failure rate; Systems engineering","score_opus":0.04937257370593322,"score_gpt":0.3290800312891988,"score_spread":0.2797074575832656,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2144309645","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.001891298,0.000039430837,0.99116856,0.0055000423,0.000053274704,0.00016433677,6.404641e-7,0.0002187911,0.00096363656],"genre_scores_gemma":[0.33655763,0.0000082937595,0.6627604,0.00016425086,0.000019905809,0.00003675947,5.657344e-7,0.0000021712624,0.0004500655],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99910206,0.00003915674,0.00014766776,0.0002963434,0.0002229298,0.00019186812],"domain_scores_gemma":[0.99828476,0.0008927089,0.000023133854,0.0006106758,0.000119660355,0.00006905514],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00045049784,0.000057937374,0.000096155185,0.000052719675,0.00010879068,0.00012347901,0.00047521445,0.00006032804,0.0002205826],"category_scores_gemma":[0.0018784961,0.000046762336,0.00009703655,0.00042153086,0.000032408516,0.00029253858,0.00008205958,0.000087625915,0.00018080322],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000002724681,0.00020304223,0.0011223397,0.000043931763,0.000012085427,6.509537e-7,0.00028946254,0.0017168182,0.000012696408,0.57253516,0.0042657596,0.41979533],"study_design_scores_gemma":[0.000052788717,0.00003065702,0.00025377507,0.000004015767,0.0000015257541,5.160623e-7,0.000003004763,0.74117297,0.00017985345,0.2561158,0.0021344298,0.000050656065],"about_ca_topic_score_codex":0.0000109947105,"about_ca_topic_score_gemma":9.330505e-7,"teacher_disagreement_score":0.7394562,"about_ca_system_score_codex":0.000033083423,"about_ca_system_score_gemma":0.000010357993,"threshold_uncertainty_score":0.24152249},"labels":[],"label_agreement":null},{"id":"W2147243633","doi":"10.5539/cis.v3n2p145","title":"Incorporating Generalized Modified Weibull TEF in to Software Reliability Growth Model and Analysis of Optimal Release Policy","year":2010,"lang":"en","type":"article","venue":"Computer and Information Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Software quality; Weibull distribution; Reliability engineering; Software release life cycle; Reliability (semiconductor); Context (archaeology); Software; Software reliability testing; Process (computing); Software development; Statistics","score_opus":0.013553900545923425,"score_gpt":0.28350762836583077,"score_spread":0.26995372781990734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2147243633","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.48953265,0.0000023272994,0.5099359,0.00034337482,0.000019917998,0.000086403976,0.0000044219455,0.000026431091,0.00004856171],"genre_scores_gemma":[0.714483,0.0000098384435,0.28519785,0.0002858847,0.000009611704,0.0000070166234,0.0000036037125,0.0000012105272,0.0000019585036],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99816394,0.00004536219,0.0005520124,0.00037307295,0.0005876411,0.00027799176],"domain_scores_gemma":[0.9982883,0.00016144764,0.00014375405,0.0005340006,0.0006065306,0.00026598398],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0020129525,0.00012128559,0.00029338137,0.0019191237,0.00020151866,0.0004032162,0.0008032741,0.000058693775,0.0000014341376],"category_scores_gemma":[0.0010043103,0.00010523797,0.00005843999,0.0059763063,0.00037989323,0.0047959085,0.00084827666,0.00019498813,0.0000019880038],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028078086,0.00008759976,0.053724986,0.00008760764,0.00002913555,8.628059e-7,0.005739378,0.724064,0.0012493829,0.109843165,0.000030908097,0.105114914],"study_design_scores_gemma":[0.00018866808,0.00003654152,0.059594307,0.0000066955367,0.000007839611,0.0000015646824,0.000010839041,0.9380647,0.00027391044,0.0016888031,0.00001483854,0.000111312664],"about_ca_topic_score_codex":0.0003076233,"about_ca_topic_score_gemma":0.000017694225,"teacher_disagreement_score":0.22495036,"about_ca_system_score_codex":0.000043680546,"about_ca_system_score_gemma":0.00036681496,"threshold_uncertainty_score":0.42914796},"labels":[],"label_agreement":null},{"id":"W2148571320","doi":"10.1145/1368088.1368187","title":"Experience applying the SPIN model checker to an industrial telecommunications system","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Computer science; Model checking; Telecommunications; Programming language","score_opus":0.15827392233734286,"score_gpt":0.3484139023140468,"score_spread":0.19013997997670393,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2148571320","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17029473,0.000021508644,0.8261994,0.0009050492,0.00003604589,0.0004187656,6.710101e-7,0.00022945425,0.0018943829],"genre_scores_gemma":[0.92892593,0.0000094213965,0.0699179,0.00022721718,0.00004537992,0.00047046845,0.000001039066,0.0000046097493,0.00039801598],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99854475,0.0001645315,0.00024253839,0.0003320759,0.00044364695,0.00027246174],"domain_scores_gemma":[0.9977622,0.00010454591,0.000033103523,0.0018114538,0.00012458532,0.00016408182],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005934564,0.0000915608,0.00013100775,0.00008874152,0.0007324611,0.00016343173,0.0025684007,0.00006071037,0.000012809318],"category_scores_gemma":[0.00015885035,0.00006054574,0.00006584135,0.0009904071,0.0000976634,0.00041990197,0.0006085323,0.00023662744,0.000090909],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000043655906,0.0011808566,0.021787353,0.000044800934,0.00014249046,0.00003013144,0.06180024,0.24372724,0.0036557477,0.42422566,0.0077701863,0.23559162],"study_design_scores_gemma":[0.00009218036,0.000023886269,0.0002822514,0.0000082368415,0.00000178977,0.000012685373,0.0008227598,0.99563295,0.000845674,0.000064278385,0.002105691,0.00010761891],"about_ca_topic_score_codex":0.00034119128,"about_ca_topic_score_gemma":0.000031782904,"teacher_disagreement_score":0.7586312,"about_ca_system_score_codex":0.00008221018,"about_ca_system_score_gemma":0.0001672775,"threshold_uncertainty_score":0.5633573},"labels":[],"label_agreement":null},{"id":"W2154454232","doi":"10.1109/qsic.2008.46","title":"On Failure Propagation in Component-Based Software Systems","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Component (thermodynamics); Reliability engineering; Computer science; Reliability (semiconductor); Reliability theory; Failure mode and effects analysis; Representation (politics); Software quality; IT service continuity; Masking (illustration); Propagation of uncertainty; Software system; Software; Distributed computing; Engineering; Failure rate; Software development; Computer network; Algorithm","score_opus":0.022903390631058528,"score_gpt":0.25058010306007544,"score_spread":0.2276767124290169,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2154454232","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16399102,0.00003242869,0.83436114,0.0009074021,0.00005427587,0.00022278671,7.5709636e-7,0.00023833496,0.00019185737],"genre_scores_gemma":[0.9888095,0.000003218329,0.010792872,0.00012383334,0.000014526511,0.000032109037,0.0000059784556,0.0000043221858,0.0002136404],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9985252,0.0001585709,0.00021049376,0.0003414888,0.00053604273,0.00022821483],"domain_scores_gemma":[0.9989682,0.0003440666,0.00003508173,0.00047840748,0.00010100297,0.000073273695],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00034247132,0.00009240534,0.000155284,0.00024911875,0.000110911744,0.00007730422,0.00049532886,0.00005861035,0.000020982754],"category_scores_gemma":[0.00021104026,0.000072966846,0.000058830905,0.0007391332,0.000046154244,0.00020650358,0.000059643527,0.00015905287,0.00017966921],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012906265,0.0036777123,0.32883072,0.00066335197,0.00009595119,0.0007936196,0.0014785175,0.44105962,0.0022360508,0.16011947,0.020157473,0.040758457],"study_design_scores_gemma":[0.00069852703,0.00015358586,0.014946122,0.00009116381,0.0000015502087,0.000012678867,0.00002411451,0.9808565,0.0014639711,0.00060639513,0.0009070352,0.00023840633],"about_ca_topic_score_codex":0.0002255944,"about_ca_topic_score_gemma":0.000033355856,"teacher_disagreement_score":0.8248185,"about_ca_system_score_codex":0.00010156953,"about_ca_system_score_gemma":0.00012515638,"threshold_uncertainty_score":0.29755014},"labels":[],"label_agreement":null},{"id":"W2156854037","doi":"10.1109/tr.2004.842087","title":"Stochastic Fault Tree Analysis With Self-Loop Basic Events","year":2005,"lang":"en","type":"article","venue":"IEEE Transactions on Reliability","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Fault tree analysis; Computer science; Loop (graph theory); Event (particle physics); Stochastic process; Moment-generating function; Stochastic modelling; Tree (set theory); Event tree; Probability-generating function; Probability density function; Mathematics; Reliability engineering; Statistics; Engineering","score_opus":0.011636418826411908,"score_gpt":0.26086771630046246,"score_spread":0.24923129747405054,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2156854037","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.17398548,0.00001691092,0.82340395,0.0015036084,0.00008365281,0.00031540476,0.000014898264,0.00048968504,0.0001864025],"genre_scores_gemma":[0.97613555,0.0000191656,0.023115119,0.00014942745,0.00003345965,0.00011299975,0.0000027834303,0.0000129977125,0.00041852196],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9962693,0.00034285148,0.0005134526,0.0011654338,0.0011178737,0.00059109565],"domain_scores_gemma":[0.9964862,0.00069083524,0.00010054516,0.0020268604,0.00036134018,0.00033422522],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009723354,0.00032627126,0.0005190986,0.0006141456,0.00041117912,0.00012638951,0.0010655781,0.00015326022,0.00016692659],"category_scores_gemma":[0.000044977747,0.0002616654,0.0005651528,0.0040836018,0.0001331793,0.00064051413,0.0000063600505,0.0006022927,0.0002290332],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008915565,0.0021969131,0.001181029,0.000033678985,0.0008890405,0.000004440423,0.00038701697,0.8752316,0.000048751594,0.00006122706,0.00004082641,0.1198363],"study_design_scores_gemma":[0.0011051376,0.00045938307,0.012905201,0.00002698404,0.0008871387,0.000009181337,0.00004192554,0.9805349,0.0025919406,0.00046123948,0.00036860755,0.0006083533],"about_ca_topic_score_codex":0.00018268026,"about_ca_topic_score_gemma":0.0005513429,"teacher_disagreement_score":0.8021501,"about_ca_system_score_codex":0.00040914555,"about_ca_system_score_gemma":0.00021280491,"threshold_uncertainty_score":0.99998355},"labels":[],"label_agreement":null},{"id":"W2158309598","doi":"10.1007/978-3-642-39512-3_8","title":"From Software Models to Dependability Analysis Models","year":2013,"lang":"en","type":"book-chapter","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"","keywords":"Dependability; Computer science; Reliability engineering; Software engineering; Engineering","score_opus":0.054534635096940565,"score_gpt":0.26954270005118663,"score_spread":0.21500806495424607,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2158309598","genre_codex":"methods","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00008493956,0.00024137898,0.87381786,0.00069777976,0.00008278984,0.00049382675,0.00012582613,0.0005835645,0.123872034],"genre_scores_gemma":[0.022462618,0.0001766272,0.4415977,0.0009235526,0.00025643004,0.00012276144,0.00024461807,0.00008915629,0.5341265],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99407905,0.000103018996,0.00090915215,0.0024125215,0.0017995015,0.00069672853],"domain_scores_gemma":[0.992941,0.00086359784,0.00018671418,0.004333865,0.00090213964,0.00077268],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00067237875,0.0006786236,0.001432524,0.0010726128,0.00019534292,0.00070361985,0.0036290223,0.00065951573,0.004662105],"category_scores_gemma":[0.000109175926,0.00058279705,0.001433115,0.00076869346,0.00011369989,0.0012226094,0.0020474338,0.00073073,0.0024950367],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002009845,0.00018570588,0.000101777274,0.000069944406,0.0056894645,0.000041086216,0.0010918974,0.33772144,0.0000055948412,0.45497695,0.0115196295,0.18857642],"study_design_scores_gemma":[0.000059041948,0.000025055191,0.000012560115,0.000018859992,0.00027306267,3.8723147e-7,0.0000048804623,0.45573428,0.0000110792325,0.540127,0.0032633683,0.00047046447],"about_ca_topic_score_codex":0.0054909443,"about_ca_topic_score_gemma":0.0007381105,"teacher_disagreement_score":0.43222016,"about_ca_system_score_codex":0.00032202326,"about_ca_system_score_gemma":0.0002960916,"threshold_uncertainty_score":0.99966234},"labels":[],"label_agreement":null},{"id":"W2159457860","doi":"10.1109/cmpsac.1978.810370","title":"Analysis of software design modeled by multiple finite state machines","year":2005,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Bell (Canada)","funders":"","keywords":"Computer science; Finite-state machine; Correctness; Deadlock; State (computer science); Distributed computing; Software; Programming language; Inter-process communication; Process (computing); Surprise; Extended finite-state machine; Software design; Software engineering; Software development","score_opus":0.024132223877171843,"score_gpt":0.2756181590862247,"score_spread":0.2514859352090529,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2159457860","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.025412206,0.00018159843,0.97370327,0.0003456971,0.000009538427,0.00010112263,0.000021404332,0.00017150366,0.000053671443],"genre_scores_gemma":[0.72623914,0.00006349704,0.27244428,0.00012598971,0.0000073524775,0.000011600474,0.00001575147,0.0000058940177,0.0010865026],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9981521,0.00019410637,0.00038343912,0.00043849126,0.00052324316,0.00030863433],"domain_scores_gemma":[0.9973636,0.0014431427,0.000086642125,0.00074852345,0.00023149581,0.0001265922],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00067402254,0.0001392431,0.00038682995,0.0004778474,0.000084369276,0.00009263527,0.0009270939,0.000049461152,0.00016839165],"category_scores_gemma":[0.00051540503,0.000110095716,0.00028246437,0.0025463041,0.00005817735,0.00041401226,0.00021283378,0.00011078219,0.00003927457],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014915688,0.0002214387,0.025097577,0.00001050095,0.00068739895,0.0000022792892,0.0002907061,0.8104745,0.00034673244,0.00007955735,0.00088164816,0.16189273],"study_design_scores_gemma":[0.00019253274,0.000031960375,0.0008167731,0.0000024387807,0.00008137502,2.7231573e-7,0.0000036175052,0.9953183,0.0026259527,0.0004159134,0.00038351558,0.00012734663],"about_ca_topic_score_codex":0.00037773844,"about_ca_topic_score_gemma":0.00015735716,"teacher_disagreement_score":0.70125896,"about_ca_system_score_codex":0.000037452835,"about_ca_system_score_gemma":0.00005954269,"threshold_uncertainty_score":0.4489573},"labels":[],"label_agreement":null},{"id":"W2161564801","doi":"10.1109/ccece.2002.1013027","title":"Evaluation of several Efron bootstrap methods to estimate error measures for software metrics","year":2003,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Statistics; Kurtosis; Confidence interval; Statistic; Nonparametric statistics; CDF-based nonparametric confidence interval; Skewness; Computer science; Robust confidence intervals; Software quality; Sample size determination; Context (archaeology); Sample (material); Population; Mathematics; Software","score_opus":0.18637125764209603,"score_gpt":0.48427364652479254,"score_spread":0.2979023888826965,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2161564801","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0038418327,0.00022850033,0.9945294,0.0002643111,0.00011159689,0.0005210532,0.0000029115424,0.000094231276,0.00040616342],"genre_scores_gemma":[0.13934322,0.000002471339,0.86032844,0.000068456204,0.000008821082,0.00007697015,0.0000017246466,0.000006837836,0.00016305283],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99657303,0.000888643,0.00033622535,0.00043060194,0.0014581362,0.00031337605],"domain_scores_gemma":[0.9962649,0.0010633477,0.00007174634,0.000593393,0.0018589225,0.00014770194],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.013947529,0.0001226949,0.0002584713,0.0003747497,0.000102425016,0.00008586913,0.00057435496,0.0000729721,0.00005640526],"category_scores_gemma":[0.0160502,0.00010236262,0.00016903107,0.0015305262,0.000031286458,0.00025121134,0.000087299755,0.000080386366,0.000009750231],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000147084465,0.00020874236,0.0018674996,0.00008907882,0.00012779237,4.2982438e-7,0.00038759378,0.07791221,0.0013262247,0.011801301,0.0011406298,0.9051238],"study_design_scores_gemma":[0.0007505499,0.0002811434,0.003656952,0.000018624823,0.00016179078,0.000003217972,0.000047984944,0.8964159,0.04708711,0.047036868,0.004238274,0.00030161245],"about_ca_topic_score_codex":0.000048472622,"about_ca_topic_score_gemma":0.000022522765,"teacher_disagreement_score":0.9048222,"about_ca_system_score_codex":0.00012848702,"about_ca_system_score_gemma":0.00038279625,"threshold_uncertainty_score":0.99223804},"labels":[],"label_agreement":null},{"id":"W2162019083","doi":"10.1109/cnsr.2007.56","title":"Risk Analysis in Critical Infrastructure Systems based on the Astrolabe Methodology","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of New Brunswick","funders":"Atlantic Canada Opportunities Agency","keywords":"Astrolabe; Critical infrastructure; Risk analysis (engineering); Computer science; Process (computing); Risk management; Computer security; Business","score_opus":0.051903835308598795,"score_gpt":0.3764612487390822,"score_spread":0.3245574134304834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2162019083","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.028612472,0.000035429388,0.96785325,0.0020309452,0.00007591007,0.00009457309,0.0000021961614,0.00006329233,0.0012319353],"genre_scores_gemma":[0.93613434,0.0000036753588,0.0633009,0.0004853917,0.00003093452,0.000007859647,0.0000011424609,0.000003192164,0.000032546777],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9965596,0.0016548468,0.00034443452,0.000448989,0.0005493438,0.00044273701],"domain_scores_gemma":[0.984789,0.014011945,0.000046466706,0.0009149064,0.00014278175,0.00009488697],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008457811,0.000119147844,0.00031473642,0.0006400059,0.00012751707,0.00015666184,0.0009568709,0.00011817513,0.00013419671],"category_scores_gemma":[0.004794599,0.00006905311,0.00020077363,0.0035637312,0.00012919126,0.0001221167,0.00012882098,0.00051158614,0.000027075497],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006058946,0.00020881701,0.48604774,0.000025145204,0.0003602162,0.00008103649,0.00027927343,0.230254,0.000096557276,0.2587831,0.00049513916,0.023308372],"study_design_scores_gemma":[0.00010432056,0.000054268803,0.14979613,0.000003965405,0.000045704688,0.0000014498017,0.00009271534,0.84713423,0.00013093553,0.0022979826,0.00025580358,0.00008248347],"about_ca_topic_score_codex":0.00056404603,"about_ca_topic_score_gemma":0.0002146887,"teacher_disagreement_score":0.9075219,"about_ca_system_score_codex":0.00007641063,"about_ca_system_score_gemma":0.00006256369,"threshold_uncertainty_score":0.5739929},"labels":[],"label_agreement":null},{"id":"W2162836425","doi":"10.1109/issre.1991.145373","title":"Metrics evaluation of software reliability growth models","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Metric (unit); Reliability (semiconductor); Computer science; Software quality; Software metric; Field (mathematics); Software; Reliability engineering; Predictive validity; Term (time); Data mining; Machine learning; Software development; Statistics; Mathematics; Engineering; Programming language","score_opus":0.10571111484273858,"score_gpt":0.3088034808344209,"score_spread":0.2030923659916823,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2162836425","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.026570218,0.0003721917,0.967516,0.0006630801,0.00005054812,0.0001957019,0.0000016460243,0.00015518634,0.004475419],"genre_scores_gemma":[0.9143364,0.00008036878,0.085338615,0.000044641027,0.000011498964,0.000015507665,0.0000011333274,0.00000383422,0.00016802723],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99690896,0.0003328728,0.00033784314,0.0004165424,0.0017810817,0.00022270199],"domain_scores_gemma":[0.9965789,0.0005867646,0.000073186,0.0008958266,0.001770827,0.000094494775],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003442721,0.00009380715,0.00019795298,0.00027852412,0.00007141442,0.000058268215,0.0008386239,0.00007331157,0.00036417608],"category_scores_gemma":[0.003670801,0.00007682149,0.00014383727,0.0021751823,0.00006956923,0.0007228843,0.00023497545,0.00012239147,0.00005569115],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000072816542,0.0016163507,0.03310097,0.00021759835,0.00012272401,0.0000025816555,0.0013329188,0.07156928,0.00010506616,0.091070086,0.0066224546,0.79423267],"study_design_scores_gemma":[0.00017725448,0.00003465119,0.00081747054,0.000004272718,0.000016670236,7.249469e-7,0.000008105781,0.9027031,0.00085628795,0.09525431,0.000046663296,0.00008049263],"about_ca_topic_score_codex":0.00013217166,"about_ca_topic_score_gemma":0.0000045823176,"teacher_disagreement_score":0.8877662,"about_ca_system_score_codex":0.000111401874,"about_ca_system_score_gemma":0.00006530095,"threshold_uncertainty_score":0.4394557},"labels":[],"label_agreement":null},{"id":"W2163293317","doi":"10.1109/hase.2008.14","title":"At What Level of Granularity Should We be Componentizing for Software Reliability?","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Granularity; Decomposition; Component (thermodynamics); Computer science; Reliability engineering; Reliability (semiconductor); Software system; Software; Component-based software engineering; Architectural pattern; Software construction; Programming language; Engineering","score_opus":0.29396059064155045,"score_gpt":0.3513395991599596,"score_spread":0.05737900851840916,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2163293317","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11752243,0.0005084889,0.8762262,0.004932364,0.000154985,0.00038021154,0.000021397282,0.00019744155,0.000056506546],"genre_scores_gemma":[0.68166256,0.0009885015,0.31471387,0.00044413834,0.000044703127,0.00005701184,0.000025682268,0.000018445491,0.002045092],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9974346,0.00013552942,0.0005005918,0.00067522866,0.0007988438,0.00045518027],"domain_scores_gemma":[0.996622,0.0014482801,0.000101617625,0.0011607525,0.00049770076,0.00016960275],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010574339,0.00018408366,0.00044129897,0.00016697998,0.00042300014,0.00010652561,0.001138684,0.0001390165,0.00006750031],"category_scores_gemma":[0.0007802664,0.00015592147,0.00042238095,0.00061525335,0.00029082436,0.0011430065,0.00070506707,0.00018713683,0.000017929591],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00077247154,0.007179023,0.42254668,0.0049897716,0.001003026,0.000169758,0.018270165,0.005719437,0.024491481,0.062095594,0.06394859,0.38881397],"study_design_scores_gemma":[0.008055106,0.0015728745,0.06787598,0.00075777073,0.00020602312,0.00030668775,0.0011655003,0.4061848,0.26659435,0.11796021,0.125673,0.0036477102],"about_ca_topic_score_codex":0.0002011722,"about_ca_topic_score_gemma":0.00007422438,"teacher_disagreement_score":0.56414014,"about_ca_system_score_codex":0.00011842358,"about_ca_system_score_gemma":0.00012191739,"threshold_uncertainty_score":0.6358293},"labels":[],"label_agreement":null},{"id":"W2164917791","doi":"10.1109/phm.2008.4711453","title":"Improving preciseness of time to failure predictions: Application to APU starter","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":19,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"National Research Council Canada","funders":"National Research Council Canada; Astellas Pharma US","keywords":"Prognostics; Cluster analysis; Computer science; Component (thermodynamics); Aerospace; Support vector machine; Data mining; Machine learning; Reliability engineering; Artificial intelligence; Engineering","score_opus":0.009138767481495032,"score_gpt":0.24081457646551568,"score_spread":0.23167580898402065,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2164917791","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.028546007,0.000007185368,0.96906537,0.0015070218,0.000013103644,0.00032005185,0.0000028785769,0.0001304001,0.00040799234],"genre_scores_gemma":[0.85997224,0.000003824719,0.1365182,0.00024902012,0.000057211873,0.00015942687,0.0000039643746,0.0000075791836,0.0030285327],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99884516,0.000039406026,0.00020895303,0.0003481139,0.00037927635,0.00017910101],"domain_scores_gemma":[0.9988229,0.00010936208,0.000029813242,0.0006447336,0.00024624253,0.00014692443],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00025831704,0.00006989819,0.00013161678,0.00017982424,0.000098233686,0.000034888504,0.0005864522,0.000039493225,0.000098031305],"category_scores_gemma":[0.00012318531,0.00006002738,0.000056660265,0.0009588233,0.000024349129,0.0002495046,0.000242859,0.00005934653,0.0006639084],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007173507,0.0010157231,0.016429402,0.00018557384,0.0001472462,0.000007923401,0.0046717376,0.03769023,0.15292872,0.0029124108,0.0423173,0.741622],"study_design_scores_gemma":[0.00045458294,0.00055661873,0.01677478,0.000044151035,0.000017576238,0.000022606333,0.000089394496,0.888513,0.055379543,0.0008969996,0.036710165,0.0005405785],"about_ca_topic_score_codex":0.00016558524,"about_ca_topic_score_gemma":0.000020949794,"teacher_disagreement_score":0.85082275,"about_ca_system_score_codex":0.000041636387,"about_ca_system_score_gemma":0.000074042175,"threshold_uncertainty_score":0.85334206},"labels":[],"label_agreement":null},{"id":"W2166054436","doi":"10.1145/1921705.1921706","title":"Measurement of software requirements derived from system reliability requirements","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Reliability engineering; Software requirements specification; Non-functional testing; Computer science; Software quality; Reliability (semiconductor); Functional requirement; Software requirements; System requirements; Software system; Software; Software reliability testing; System requirements specification; Verification and validation; Avionics software; Requirements analysis; Software construction; Software development; Software engineering; Requirement; Engineering; Operating system","score_opus":0.04205296260434291,"score_gpt":0.29026509441791554,"score_spread":0.24821213181357263,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2166054436","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.45065078,0.000060618786,0.54692304,0.00026996393,0.0005375695,0.00035640184,0.000009033758,0.00039982583,0.00079278415],"genre_scores_gemma":[0.9053423,0.0000059354147,0.09446485,0.000041952444,0.000051048144,0.0000335857,0.00000505277,0.000010906583,0.000044359855],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9953021,0.00024942055,0.00081548403,0.000871197,0.0022873762,0.00047443132],"domain_scores_gemma":[0.99588054,0.00019880678,0.00022773321,0.0022672086,0.001173333,0.00025238318],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0027521355,0.00023100158,0.0004451054,0.00015120607,0.00017441582,0.00012923089,0.0018991814,0.00015338468,0.00015772144],"category_scores_gemma":[0.0014637318,0.000192091,0.00023737304,0.0006241491,0.00015753935,0.0006239705,0.00069757533,0.00032607568,0.00011244397],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009421303,0.0015903105,0.36093915,0.0007603955,0.0005881055,0.000026341386,0.0007164851,0.00013911365,0.53366554,0.0088927755,0.0010573023,0.091530286],"study_design_scores_gemma":[0.0038029647,0.0005324068,0.27796888,0.0007313629,0.00018198525,0.000008223721,0.0004335702,0.052728835,0.64773256,0.011099006,0.0031083308,0.0016718743],"about_ca_topic_score_codex":0.0012789828,"about_ca_topic_score_gemma":0.00018660986,"teacher_disagreement_score":0.45469153,"about_ca_system_score_codex":0.0002317562,"about_ca_system_score_gemma":0.00026498103,"threshold_uncertainty_score":0.78332436},"labels":[],"label_agreement":null},{"id":"W2168813770","doi":"10.1109/sess.1997.596004","title":"Fundamental Principles Of Software Engineering","year":2005,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Computer science; Fist; Software engineering; Software; Point (geometry); Value (mathematics); Software development; Position paper; Engineering ethics; Management science; Data science; Engineering; World Wide Web","score_opus":0.02106128218866731,"score_gpt":0.2544934242789839,"score_spread":0.2334321420903166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2168813770","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15149842,0.00010693276,0.84709543,0.00037955158,0.000026720341,0.000049483122,8.101139e-7,0.00019329779,0.00064936397],"genre_scores_gemma":[0.7321272,0.000008628759,0.26728326,0.000030782696,0.000023651537,0.0000028481297,5.2284906e-7,0.0000024090589,0.0005207006],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9992706,0.000011733359,0.00015375669,0.00015828146,0.0002534018,0.0001521947],"domain_scores_gemma":[0.9994315,0.00013416211,0.000021006454,0.00030899904,0.000047298665,0.0000570338],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00019992176,0.00005347384,0.000095797455,0.000084164116,0.00002540475,0.0000363768,0.0004678238,0.000023644534,0.00010043717],"category_scores_gemma":[0.00016583088,0.000044811743,0.000069013906,0.0002753728,0.000020047624,0.00024009294,0.00021885826,0.000060279133,0.000051818788],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009111033,0.00072781154,0.21553266,0.00022420647,0.00019347758,0.00001042101,0.0012364179,0.067891605,0.007418087,0.24589305,0.000658123,0.46020502],"study_design_scores_gemma":[0.00039205706,0.00010061437,0.04096619,0.000038769365,0.000006994094,0.0000113132855,0.000036161684,0.8224765,0.06052825,0.00022028857,0.0749004,0.00032242565],"about_ca_topic_score_codex":0.000015578986,"about_ca_topic_score_gemma":0.000005801578,"teacher_disagreement_score":0.7545849,"about_ca_system_score_codex":0.00003224888,"about_ca_system_score_gemma":0.000031097683,"threshold_uncertainty_score":0.18273698},"labels":[],"label_agreement":null},{"id":"W2168965642","doi":"10.1109/ssiri.2010.12","title":"Failure Type-Aware Reliability Assessment with Component Failure Dependency","year":2010,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Reliability engineering; Component (thermodynamics); Dependency (UML); Reliability (semiconductor); Computer science; Criticality; Engineering; Artificial intelligence","score_opus":0.00841511832364373,"score_gpt":0.27468364771842346,"score_spread":0.26626852939477974,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2168965642","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21453239,0.000009292526,0.7720075,0.010375725,0.00020255084,0.00037565714,0.0000032825885,0.0005154951,0.0019780719],"genre_scores_gemma":[0.85158473,0.0000043947634,0.14779902,0.00012003904,0.000050799426,0.00002279069,0.000009271661,0.000010482109,0.00039845594],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9971639,0.00015137847,0.00030912782,0.0008364771,0.0010442301,0.000494891],"domain_scores_gemma":[0.9969536,0.0003035058,0.000073429044,0.0017673413,0.0006065678,0.00029553924],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00090628467,0.00023120547,0.0002991044,0.0001334705,0.00025982084,0.0003508541,0.0014996965,0.00016129682,0.00069827633],"category_scores_gemma":[0.00013263118,0.00015668833,0.00012088158,0.00095803785,0.00015467027,0.00060024863,0.00047796016,0.00093109853,0.00018917392],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000095479874,0.0032959548,0.7237735,0.00040014467,0.00043952503,0.00026292607,0.00091511407,0.0024680924,0.021682192,0.17024575,0.01926645,0.05715491],"study_design_scores_gemma":[0.0044732434,0.0023659589,0.24963693,0.00016641556,0.00014210121,0.0004112049,0.0011313902,0.50178254,0.016697032,0.037605666,0.18165807,0.003929465],"about_ca_topic_score_codex":0.00026842093,"about_ca_topic_score_gemma":0.0016106189,"teacher_disagreement_score":0.63705236,"about_ca_system_score_codex":0.000075618875,"about_ca_system_score_gemma":0.00042570182,"threshold_uncertainty_score":0.7645636},"labels":[],"label_agreement":null},{"id":"W2169071419","doi":"10.1007/s11704-012-2904-2","title":"The use of mathematics in software quality assurance","year":2012,"lang":"en","type":"article","venue":"Frontiers of Computer Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"JDA Software (Canada)","funders":"Office of Naval Research","keywords":"Computer science; Software quality assurance; Quality assurance; Software quality analyst; Software quality; Software engineering; Verification and validation; Software; Quality (philosophy); Software quality control; Software construction; Software metric; Software system; Software development; Programming language; Mathematics; Statistics; Engineering","score_opus":0.053524351866737784,"score_gpt":0.3098367404852625,"score_spread":0.25631238861852473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169071419","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10004586,0.00048806172,0.8986245,0.00024639253,0.00044851523,0.0001082851,0.0000016030516,0.00002688866,0.000009909119],"genre_scores_gemma":[0.41363487,0.000039418166,0.5862652,0.000023470064,0.000016600425,0.0000031707389,1.3256015e-7,0.000002224711,0.0000149040525],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9974698,0.0001570862,0.00054161536,0.0002796167,0.00105035,0.00050156325],"domain_scores_gemma":[0.9970689,0.001101679,0.0002468136,0.0011080228,0.00034987307,0.00012472145],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0038050339,0.000102928185,0.0002921269,0.00022120948,0.00016156853,0.00014965246,0.002491516,0.000038020513,6.584344e-7],"category_scores_gemma":[0.0011827565,0.00007206492,0.00010219668,0.0019620268,0.00095079973,0.0016672511,0.0006495264,0.00013533205,0.0000029234104],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000008906067,0.0006493889,0.74332863,0.0001300005,0.00003355241,0.0000012663858,0.0035682963,0.0047808504,0.00039438048,0.024095712,0.003917909,0.21909109],"study_design_scores_gemma":[0.00032837555,0.00009398508,0.44528347,0.00009703391,0.000006420955,0.0000048392317,0.0000996229,0.5383117,0.0056100413,0.0054046316,0.00443213,0.00032774787],"about_ca_topic_score_codex":0.00007353618,"about_ca_topic_score_gemma":0.000011663204,"teacher_disagreement_score":0.53353083,"about_ca_system_score_codex":0.000058157148,"about_ca_system_score_gemma":0.0002206104,"threshold_uncertainty_score":0.4629898},"labels":[],"label_agreement":null},{"id":"W2169118413","doi":"10.1109/icsssm.2006.320532","title":"Availability evaluation of multicomponent systems under repair policies: a generalized kronecker-based approach","year":2006,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Kronecker delta; Component (thermodynamics); Computer science; Automaton; Tensor (intrinsic definition); Theoretical computer science; Distributed computing; Mathematical optimization; Algebra over a field; Mathematics","score_opus":0.06422696575890857,"score_gpt":0.3187077949055028,"score_spread":0.2544808291465942,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2169118413","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.405213,0.0002316491,0.5920519,0.00020586816,0.00005145694,0.0005106266,0.0000027440542,0.0001890189,0.0015437389],"genre_scores_gemma":[0.9421091,0.000002052508,0.057507154,0.00003502361,0.000036806414,0.000087922104,0.000018877097,0.0000065787362,0.00019647075],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99572307,0.0010099452,0.0006369493,0.0005904649,0.0017150061,0.00032459383],"domain_scores_gemma":[0.99735266,0.00031042632,0.00014294086,0.0012663425,0.0008430056,0.000084642525],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0044669295,0.00015688686,0.00034589495,0.00022043311,0.00010154039,0.00010178338,0.0005887529,0.00009463949,0.000055704633],"category_scores_gemma":[0.00021389562,0.00012373876,0.0002703424,0.0007965534,0.00013875899,0.00019068005,0.00013603525,0.00010457813,0.000025398143],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020645513,0.001825093,0.013983538,0.00021926548,0.00009039199,5.2020755e-7,0.00014348158,0.94604105,0.003803873,0.031639513,0.00074479147,0.0014878267],"study_design_scores_gemma":[0.0007891827,0.0000332781,0.010497482,0.000010981819,0.000033825392,0.0000010525938,0.000050746323,0.9859024,0.0014474557,0.0009079627,0.00018906218,0.0001365336],"about_ca_topic_score_codex":0.008622484,"about_ca_topic_score_gemma":0.00005769895,"teacher_disagreement_score":0.5368961,"about_ca_system_score_codex":0.0003450817,"about_ca_system_score_gemma":0.00034172428,"threshold_uncertainty_score":0.99797916},"labels":[],"label_agreement":null},{"id":"W2170874362","doi":"10.5539/mas.v3n11p71","title":"Reliability Growth Testing Based on Dynamic Planning Methodology","year":2009,"lang":"en","type":"article","venue":"Modern Applied Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Reliability (semiconductor); Reliability engineering; Computer science; Test plan; Test (biology); Work (physics); Engineering; Mathematics; Statistics","score_opus":0.07585846233438953,"score_gpt":0.3475916472983961,"score_spread":0.27173318496400656,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2170874362","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03287962,0.000013689011,0.9596273,0.0008349117,0.000057465062,0.00018724348,6.8284e-7,0.00035286858,0.0060462244],"genre_scores_gemma":[0.7099229,4.6124643e-7,0.28925678,0.000778674,0.000012090849,0.0000118033395,5.575452e-7,0.000003884349,0.000012793023],"study_design_codex":"bench_or_experimental","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.996229,0.0001726966,0.0003137072,0.0013166316,0.0011951071,0.00077285344],"domain_scores_gemma":[0.9964044,0.0017278065,0.00009764081,0.0012488993,0.00028115706,0.00024006191],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0057021906,0.0001968354,0.00028584056,0.0004291771,0.00065207074,0.00030496137,0.0025567075,0.00008119282,0.000004643937],"category_scores_gemma":[0.0029316866,0.00016891793,0.00007251585,0.0030500705,0.00050533796,0.0003647147,0.00025912552,0.00040932573,0.0000380603],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005854757,0.00044454733,0.0027721243,0.00003432144,0.000005139534,0.000026786172,0.0010201692,0.16401117,0.43328166,0.026471728,0.000033270015,0.37184054],"study_design_scores_gemma":[0.0001444737,0.000114128045,0.017689502,0.0000129877635,0.000002375658,0.0000023100845,0.000008152003,0.8809445,0.0032557384,0.097644046,0.000009398268,0.00017239158],"about_ca_topic_score_codex":0.000016761509,"about_ca_topic_score_gemma":7.17166e-7,"teacher_disagreement_score":0.7169333,"about_ca_system_score_codex":0.00023315402,"about_ca_system_score_gemma":0.00042157588,"threshold_uncertainty_score":0.68882734},"labels":[],"label_agreement":null},{"id":"W2186159877","doi":"","title":"Application of Tabular Methods to the Speciflcation and Veriflcation of a Nuclear Reactor Shutdown System","year":2001,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Suite; Software; Software engineering; Shutdown; Process (computing); Nuclear power; Systems engineering; Computer science; Work (physics); Selection (genetic algorithm); Engineering; Operating system; Artificial intelligence; Mechanical engineering","score_opus":0.01813443198710177,"score_gpt":0.3180393741988065,"score_spread":0.2999049422117047,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2186159877","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05180668,0.00004551143,0.9459201,0.0012776466,0.000017957098,0.0002372082,6.062137e-7,0.000040488794,0.0006537947],"genre_scores_gemma":[0.8627199,0.000032611046,0.13715245,0.000030860356,0.0000191069,0.000012085365,0.0000011365862,0.0000030575056,0.000028799144],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99905705,0.00016526946,0.00023715194,0.00020960861,0.00024442453,0.00008649844],"domain_scores_gemma":[0.9988401,0.00019896486,0.000091950205,0.0006164802,0.00020746578,0.000045032502],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011907682,0.000051274586,0.00012898076,0.00009352827,0.000044465873,0.00003182348,0.00041953952,0.000034096833,0.000008254585],"category_scores_gemma":[0.00013811929,0.000035053316,0.000037625312,0.00076461304,0.000035900015,0.0001411536,0.00011937481,0.000045706947,0.00001291799],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028176559,0.00012380687,0.004482744,0.00014701099,0.000044933622,2.8536505e-7,0.0020210755,0.000271202,0.065411136,0.26137784,0.0002695602,0.6658222],"study_design_scores_gemma":[0.0002516971,0.00017804929,0.061913215,0.00004864634,0.00003136137,0.000009929935,0.0011583427,0.8714421,0.023652263,0.0013941119,0.03973626,0.00018403672],"about_ca_topic_score_codex":0.00030546985,"about_ca_topic_score_gemma":0.000010823364,"teacher_disagreement_score":0.8711709,"about_ca_system_score_codex":0.00003530737,"about_ca_system_score_gemma":0.000020936803,"threshold_uncertainty_score":0.14294328},"labels":[],"label_agreement":null},{"id":"W2206726861","doi":"10.5267/j.dsl.2015.6.005","title":"Application of fuzzy-MOORA method: Ranking of components for reliability estimation of component-based software systems","year":2015,"lang":"en","type":"article","venue":"Decision Science Letters","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Component (thermodynamics); Ranking (information retrieval); Fuzzy logic; Reliability (semiconductor); Computer science; Estimation; Reliability engineering; Data mining; Software; Mathematics; Mathematical optimization; Engineering; Artificial intelligence; Systems engineering","score_opus":0.051522974188548246,"score_gpt":0.3613228720203835,"score_spread":0.30979989783183526,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2206726861","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.30898556,0.000024070707,0.69005615,0.00034953293,0.0001239556,0.0004204379,0.000008947437,0.000028513428,0.0000028464544],"genre_scores_gemma":[0.5887781,4.725508e-7,0.41114002,0.00004365837,0.0000056321414,0.000023138156,0.0000053219983,0.0000032961368,3.809354e-7],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9956979,0.00023513894,0.0009388222,0.000622976,0.0022257043,0.00027947032],"domain_scores_gemma":[0.9938433,0.0028387643,0.00062203605,0.001238592,0.0013037933,0.00015347738],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009747353,0.00013040831,0.0004847454,0.0006234601,0.000118112795,0.000074754156,0.0018900783,0.000060901206,6.410491e-7],"category_scores_gemma":[0.0037510653,0.000111451234,0.00016677855,0.0023316469,0.0005672474,0.00061986834,0.0002521895,0.00009328038,0.0000029624293],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012139675,0.00024395711,0.012672584,0.0002448467,0.000011223224,4.0924502e-7,0.00034800696,0.7613951,0.110848695,0.0017062925,0.00012942079,0.11227809],"study_design_scores_gemma":[0.0006300827,0.00007890427,0.0060051996,0.00008899541,0.000010293199,0.0000010898754,0.000032314536,0.96277237,0.0271026,0.0031154228,0.000061496656,0.000101246325],"about_ca_topic_score_codex":0.00027539668,"about_ca_topic_score_gemma":8.421658e-7,"teacher_disagreement_score":0.27979252,"about_ca_system_score_codex":0.00014704982,"about_ca_system_score_gemma":0.00022445458,"threshold_uncertainty_score":0.45448494},"labels":[],"label_agreement":null},{"id":"W2224341679","doi":"","title":"Software estimation: from wishful thinking to sound engineering","year":2015,"lang":"en","type":"article","venue":"Conferencia Iberoamericana de Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Wishful thinking; Sound (geography); Computer science; Estimation; Engineering; Systems engineering; Acoustics; Psychology; Cognitive psychology","score_opus":0.024112958078131654,"score_gpt":0.2545124777852503,"score_spread":0.23039951970711867,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2224341679","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12070961,0.00030000112,0.87647605,0.00023657223,0.00032770348,0.00021044211,0.000014546598,0.0017007185,0.000024376322],"genre_scores_gemma":[0.48101726,0.000005243609,0.51848143,0.00019785344,0.00014783545,0.00006369186,0.000022456468,0.00004265441,0.00002158096],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99662364,0.00007110482,0.00049978046,0.000859868,0.0009794282,0.0009661681],"domain_scores_gemma":[0.996298,0.0009873535,0.000101256264,0.0011460182,0.0004057014,0.0010616323],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00067671196,0.00045231072,0.0005799968,0.00046631688,0.00014613246,0.0006153241,0.0018144145,0.00016887937,0.000043548403],"category_scores_gemma":[0.0066190064,0.0005008955,0.00018128687,0.0016037107,0.000041361785,0.0006886637,0.0006759843,0.00056097464,0.00018929271],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015822498,0.00009752632,0.04169316,0.00009718869,0.00029338236,0.00013751547,0.005987128,0.84922796,0.00032660653,0.0034291318,0.0011327576,0.09756182],"study_design_scores_gemma":[0.0007757956,0.00020296751,0.03493013,0.0002973475,0.00007199696,0.00004150879,0.00018862438,0.9519429,0.001053176,0.0049574026,0.0039079576,0.0016301983],"about_ca_topic_score_codex":0.0006914942,"about_ca_topic_score_gemma":0.000025392163,"teacher_disagreement_score":0.36030766,"about_ca_system_score_codex":0.00063599285,"about_ca_system_score_gemma":0.00057237025,"threshold_uncertainty_score":0.9997443},"labels":[],"label_agreement":null},{"id":"W2230115085","doi":"10.12794/metadc700064","title":"An Evaluation of Negative Reinforcement During Error Correction Procedures","year":2014,"lang":"en","type":"dissertation","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Global Affairs Canada","funders":"","keywords":"Reinforcement; Computer science; Psychology; Social psychology","score_opus":0.02814144576090903,"score_gpt":0.35587687193640105,"score_spread":0.327735426175492,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2230115085","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8710443,0.00007813228,0.11091033,0.00005131806,0.0010447972,0.0011518846,0.0000010570442,0.00024079412,0.015477428],"genre_scores_gemma":[0.9946335,0.000017660757,0.0012340317,0.00000828967,0.00005475516,0.00010580571,0.00013817968,0.000009316561,0.0037984764],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9971052,0.000272654,0.00041108378,0.000472805,0.0015547853,0.00018346094],"domain_scores_gemma":[0.9973236,0.000111715315,0.00030084953,0.00059663155,0.0015943414,0.000072917486],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001424094,0.00016433622,0.00026218744,0.0003447132,0.00012840844,0.00009861615,0.0005859062,0.00015864542,0.00019941248],"category_scores_gemma":[0.0010276961,0.0001438176,0.000116599054,0.0005141595,0.000021299442,0.00043307088,0.000033611475,0.00018326238,0.000019055602],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00061282195,0.001215148,0.0037501268,0.004457961,0.0009316831,0.0000024382698,0.026918896,0.34789863,0.023199445,0.0076653203,0.0020954374,0.5812521],"study_design_scores_gemma":[0.00037708538,0.00019114437,0.018659327,0.00018358766,0.00008030901,8.9374635e-7,0.00089211215,0.89258724,0.08544078,0.0013481743,0.000012843567,0.00022649232],"about_ca_topic_score_codex":0.00034996384,"about_ca_topic_score_gemma":0.0009899641,"teacher_disagreement_score":0.5810256,"about_ca_system_score_codex":0.00016119539,"about_ca_system_score_gemma":0.00054359116,"threshold_uncertainty_score":0.58647114},"labels":[],"label_agreement":null},{"id":"W2242225095","doi":"10.2139/ssrn.1026842","title":"Dynamic Optimization of Software Enhancement Effort","year":2007,"lang":"en","type":"article","venue":"SSRN Electronic Journal","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science; Software; Operating system","score_opus":0.005892796696382247,"score_gpt":0.2681912104826593,"score_spread":0.26229841378627705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2242225095","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0355206,0.0009921915,0.9630022,0.00020485405,0.00007878746,0.00007003976,2.121747e-7,0.00003537327,0.00009570875],"genre_scores_gemma":[0.9290015,0.002196924,0.068279885,0.00003083225,0.000031457064,0.0000017729083,0.0000025785666,0.000008140718,0.00044689982],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9973118,0.00004387059,0.00038411858,0.0002079302,0.0005699176,0.0014824123],"domain_scores_gemma":[0.9991087,0.000102491074,0.00016179001,0.0003016644,0.00024041695,0.00008495563],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.004078581,0.00009800598,0.00016500903,0.00023085985,0.00014128328,0.000051291638,0.0007252642,0.0000569841,0.000025896734],"category_scores_gemma":[0.00014389148,0.00008507732,0.00013607868,0.00060320867,0.000043518532,0.00030244037,0.00010116865,0.0007559258,0.000010789769],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009150698,0.00060605415,0.012523137,0.000054433058,0.0005092283,0.00001578902,0.00046184656,0.06733035,0.0013073366,0.1018195,0.00003837104,0.81524247],"study_design_scores_gemma":[0.0027237004,0.002778677,0.0056982916,0.0001647334,0.00011009663,0.0007272048,0.0009939335,0.5344221,0.010925755,0.43924835,0.0011521884,0.001054968],"about_ca_topic_score_codex":0.000020161851,"about_ca_topic_score_gemma":0.0001083279,"teacher_disagreement_score":0.89472234,"about_ca_system_score_codex":0.0007306113,"about_ca_system_score_gemma":0.0011616421,"threshold_uncertainty_score":0.34693524},"labels":[],"label_agreement":null},{"id":"W2251589446","doi":"10.1145/2559636","title":"Proceedings of the 2014 ACM/IEEE international conference on Human-robot interaction","year":2014,"lang":"en","type":"paratext","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":138,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Computer science; Multidisciplinary approach; Human–robot interaction; Robot; Process (computing); Rebuttal; Pleasure; Empirical research; Artificial intelligence; Theme (computing); World Wide Web; Sociology; Psychology; Political science; Social science","score_opus":0.125300995503274,"score_gpt":0.4034465881089634,"score_spread":0.2781455926056894,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2251589446","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006380622,0.000035119756,0.08763943,0.007969323,0.0070493026,0.0006490152,0.000020874939,0.00013969527,0.89011663],"genre_scores_gemma":[0.6810561,0.00022179687,0.00260515,0.0004912532,0.0007354103,0.00006290708,0.000028511706,0.00002514324,0.31477374],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99771136,0.000059248192,0.00042806848,0.0006128061,0.00096094713,0.00022757942],"domain_scores_gemma":[0.99772155,0.0002153794,0.0003724341,0.0007616146,0.0008650085,0.000063983745],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00051912013,0.00022597692,0.00034856878,0.00026205878,0.00013167057,0.00036455746,0.0046597896,0.0002187762,0.0025251661],"category_scores_gemma":[0.00022982527,0.0001460453,0.00027479566,0.00030508867,0.00013899835,0.00028542968,0.0008523701,0.000686204,0.0026339474],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014319725,0.00019615484,0.00021950238,0.00014940415,0.0001574723,2.453131e-7,0.00017440374,0.00025517654,0.0054708146,0.047352567,0.9400825,0.0059274556],"study_design_scores_gemma":[0.00172087,0.0012232191,0.0034347253,0.00323704,0.00012817708,0.000036688267,0.00032614038,0.1455272,0.1273847,0.024364674,0.69020253,0.0024140049],"about_ca_topic_score_codex":0.00014079364,"about_ca_topic_score_gemma":0.00001494246,"teacher_disagreement_score":0.67467546,"about_ca_system_score_codex":0.00012569073,"about_ca_system_score_gemma":0.00011649181,"threshold_uncertainty_score":0.9983867},"labels":[],"label_agreement":null},{"id":"W2277826145","doi":"","title":"Exploratory study on an innovative use of COSMIC-FFP for early quality assessment","year":2007,"lang":"en","type":"dissertation","venue":"Spectrum Research Repository (Concordia University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"COSMIC cancer database; Computer science; Scale (ratio); Reliability engineering; Context (archaeology); Systems engineering; Engineering; Physics; Astrophysics","score_opus":0.12433944255881382,"score_gpt":0.4080491565695253,"score_spread":0.28370971401071143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2277826145","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9829987,0.000017069673,0.009451196,0.00006580608,0.00046347242,0.00205377,0.00003519379,0.0001481988,0.004766545],"genre_scores_gemma":[0.99219936,0.000020851845,0.0006045108,0.000008049959,0.00014454784,0.000030638483,0.000107857784,0.000043600514,0.006840605],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99071527,0.0030786977,0.00080935063,0.0016482743,0.002744265,0.0010041721],"domain_scores_gemma":[0.99021554,0.0029220264,0.0005493907,0.0022167785,0.0036611306,0.00043511562],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0052007614,0.00043699716,0.00090273947,0.0035322453,0.000875276,0.00050635455,0.0026977314,0.00041748447,0.000008130257],"category_scores_gemma":[0.0006443374,0.0004724313,0.00035223775,0.005315596,0.00033308115,0.0015311955,0.0004228512,0.0018129734,0.000008804409],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00919635,0.016796412,0.7920867,0.0015149172,0.0037640315,0.0021145958,0.029922768,0.00024082288,0.021588571,0.10137505,0.001185782,0.020213978],"study_design_scores_gemma":[0.0019221317,0.008480664,0.94081116,0.00019099467,0.000071386225,0.000001553442,0.017522588,0.00066973345,0.02548804,0.0011458362,0.0028240944,0.00087180495],"about_ca_topic_score_codex":0.010539271,"about_ca_topic_score_gemma":0.008056162,"teacher_disagreement_score":0.14872445,"about_ca_system_score_codex":0.0013837332,"about_ca_system_score_gemma":0.0032742885,"threshold_uncertainty_score":0.9997727},"labels":[],"label_agreement":null},{"id":"W2286276931","doi":"10.1007/978-3-642-38911-5_4","title":"Modeling Early Availability Requirements Using Aspect-Oriented Use Case Maps","year":2013,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University; University of Ottawa","funders":"","keywords":"Computer science; Dependability; Requirements elicitation; Software engineering; Requirements analysis; Notation; Non-functional requirement; Functional requirement; Unified Modeling Language; Software development; Software; Programming language; Software construction","score_opus":0.06769988310583246,"score_gpt":0.3026688174043542,"score_spread":0.23496893429852173,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2286276931","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.024013095,0.00022413649,0.97364223,0.00015020349,0.000811702,0.0006515589,0.000010768814,0.00022670138,0.00026962825],"genre_scores_gemma":[0.48191828,0.000036383783,0.51701933,0.00032307225,0.000298347,0.000013325682,0.0000069691605,0.00004920106,0.0003350886],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9928066,0.00015369341,0.0010173414,0.0027492687,0.0020783197,0.0011947564],"domain_scores_gemma":[0.9945172,0.0007242993,0.00026096823,0.0029914796,0.001077993,0.0004280327],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0020825865,0.0007008133,0.00079831877,0.0012069036,0.0006372559,0.0017543787,0.0030595458,0.0004520203,0.000083316605],"category_scores_gemma":[0.0006348245,0.0006437147,0.0003069972,0.001427628,0.000883612,0.0024816466,0.0027859046,0.0013539679,0.00013796915],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019209407,0.00021485894,0.0014497471,0.00017069999,0.00010477013,0.0025649897,0.0016617096,0.55261254,0.00034661862,0.010525099,0.0000152484945,0.43031454],"study_design_scores_gemma":[0.00020531117,0.00010057673,0.000023786664,0.00024294805,0.00001618937,0.0003119408,3.166897e-7,0.95534533,0.0001778914,0.04281355,0.00012956643,0.00063258887],"about_ca_topic_score_codex":0.0018264505,"about_ca_topic_score_gemma":0.00021264881,"teacher_disagreement_score":0.45790517,"about_ca_system_score_codex":0.0010277001,"about_ca_system_score_gemma":0.0008068154,"threshold_uncertainty_score":0.9996014},"labels":[],"label_agreement":null},{"id":"W2289544687","doi":"10.1007/978-3-319-09940-8_2","title":"Detecting Program Execution Phases Using Heuristic Search","year":2014,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":15,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Computer science; Heuristic; Artificial intelligence","score_opus":0.04170009338257575,"score_gpt":0.3274090775574338,"score_spread":0.28570898417485807,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2289544687","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0011828471,0.00031763042,0.99646103,0.0001827213,0.00053576817,0.0005197827,0.0000017546128,0.00034603567,0.0004524306],"genre_scores_gemma":[0.4855732,0.00003410266,0.51344407,0.00021910502,0.00053554587,0.00001313144,0.000004015066,0.00003351809,0.0001433342],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9946279,0.00012682793,0.0005859177,0.0018459983,0.0018133416,0.0010000147],"domain_scores_gemma":[0.9961987,0.0012413573,0.00020386676,0.0015544911,0.0005413763,0.0002601761],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0024256152,0.0004773577,0.0006005633,0.0012509917,0.000578173,0.001150907,0.0032257596,0.00032242562,0.00002280515],"category_scores_gemma":[0.00059753354,0.000436662,0.00022627081,0.0013823957,0.00096023805,0.000426367,0.0016067876,0.001203182,0.000034447916],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000036293497,0.000040035713,0.00009213207,0.00006576323,0.000011629736,0.00004312314,0.00017855444,0.049711365,0.00020973025,0.0013692296,0.0000019596998,0.9482728],"study_design_scores_gemma":[0.00016054674,0.00024573316,0.000031277552,0.00037396798,0.00001131493,0.00006928822,1.6995148e-7,0.9709711,0.0018155547,0.025324192,0.000514597,0.0004822779],"about_ca_topic_score_codex":0.00010748181,"about_ca_topic_score_gemma":0.00004115363,"teacher_disagreement_score":0.94779056,"about_ca_system_score_codex":0.00052730134,"about_ca_system_score_gemma":0.00075821957,"threshold_uncertainty_score":0.999886},"labels":[],"label_agreement":null},{"id":"W2294411928","doi":"10.1016/j.apm.2016.02.040","title":"Bayes, E-Bayes and robust Bayes prediction of a future observation under precautionary prediction loss functions with applications","year":2016,"lang":"en","type":"article","venue":"Applied Mathematical Modelling","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":32,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Bayes' theorem; Hyperparameter; Prior probability; Bayes error rate; Bayes factor; Bayesian probability; Machine learning; Computer science; Naive Bayes classifier; Artificial intelligence; Bayes estimator; Bayesian programming; Bayes' rule; Bayesian hierarchical modeling; Bayes classifier; Mathematics","score_opus":0.031161287106028168,"score_gpt":0.2230426331882349,"score_spread":0.19188134608220672,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2294411928","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013110077,0.000092594215,0.9846292,0.0010343824,0.000022901715,0.000553232,0.00003713914,0.00022931756,0.00029112294],"genre_scores_gemma":[0.7738067,0.00019150479,0.22519858,0.000024945668,0.00014368075,0.00045738195,0.000027984413,0.00001772561,0.00013152418],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9982142,0.00004192721,0.00043511076,0.0005388949,0.00052694243,0.0002429102],"domain_scores_gemma":[0.99847245,0.0005205015,0.00012748108,0.00053721707,0.00022068982,0.00012166901],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00043380444,0.00017111459,0.00024704493,0.00014332133,0.0002458607,0.0000667444,0.00026014037,0.00014254941,0.000027017017],"category_scores_gemma":[0.000020031668,0.00011337748,0.00006018526,0.00055740017,0.00018481794,0.0004698365,0.0001039408,0.00015021425,0.000016952645],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000048390997,0.00042108563,0.0009921653,0.00035040706,0.00013856488,4.879251e-7,0.0003486383,0.13614702,0.0015710994,0.8389251,0.0000813068,0.020975698],"study_design_scores_gemma":[0.00040929927,0.00007924104,0.0006463304,0.00013456939,0.00006541691,0.000012249662,0.00013317821,0.75979936,0.00049527374,0.23783346,0.0002467786,0.00014482785],"about_ca_topic_score_codex":0.000004060081,"about_ca_topic_score_gemma":0.0000014854098,"teacher_disagreement_score":0.7606966,"about_ca_system_score_codex":0.00006818689,"about_ca_system_score_gemma":0.000073822885,"threshold_uncertainty_score":0.46233994},"labels":[],"label_agreement":null},{"id":"W2298190639","doi":"10.7282/t3br8rhn","title":"Flexible modeling and simulating mission availability within the operational framework for Canadian Naval platforms","year":2011,"lang":"en","type":"article","venue":"Rutgers University Community Repository (Rutgers University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Systems engineering; Computer science; Aeronautics; Engineering; Operations research","score_opus":0.060255611926483856,"score_gpt":0.24422923143940412,"score_spread":0.18397361951292027,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2298190639","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08836052,0.000054124204,0.7705553,0.0006474699,0.0003309666,0.0008036361,0.000027016496,0.00037944224,0.13884152],"genre_scores_gemma":[0.96680695,0.00001624446,0.03239735,0.000089774796,0.000028435687,2.921178e-7,0.000013277112,0.000012171336,0.0006355033],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.997809,0.0005925808,0.00025108934,0.0005058346,0.00033832242,0.0005031902],"domain_scores_gemma":[0.9966269,0.0010506712,0.0001449428,0.0011961766,0.00045361093,0.0005276953],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0012684414,0.00026040178,0.00029197946,0.000554949,0.005819262,0.00020542087,0.0021455344,0.00027694608,0.000015952503],"category_scores_gemma":[0.00031058997,0.00025429646,0.00022019517,0.0011288577,0.00044040737,0.0012769018,0.00069764577,0.0010570325,0.000004552038],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0037548882,0.0034601763,0.08798948,0.0013679712,0.00377837,0.00092851603,0.2395947,0.22143815,0.0759153,0.34275743,0.0011049649,0.017910045],"study_design_scores_gemma":[0.0014464129,0.00038734757,0.00016550903,0.00018932452,0.00018597553,0.00002837459,0.002754535,0.91628766,0.05273287,0.024491245,0.00032381268,0.001006924],"about_ca_topic_score_codex":0.052055303,"about_ca_topic_score_gemma":0.00011870754,"teacher_disagreement_score":0.87844646,"about_ca_system_score_codex":0.0006863663,"about_ca_system_score_gemma":0.000919531,"threshold_uncertainty_score":0.99999094},"labels":[],"label_agreement":null},{"id":"W2314078445","doi":"10.2514/6.2008-3444","title":"Optimising Mission Investment Through System Availability Selection","year":2008,"lang":"en","type":"article","venue":"SpaceOps 2008 Conference","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Space Agency","funders":"","keywords":"Selection (genetic algorithm); Computer science; Investment (military); Artificial intelligence","score_opus":0.06341982973072915,"score_gpt":0.2882791648715246,"score_spread":0.22485933514079545,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2314078445","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.10246366,0.00031123706,0.88291717,0.0012074568,0.00020608962,0.00034419526,0.0000023258062,0.00066399056,0.011883853],"genre_scores_gemma":[0.9235137,0.00015348248,0.07498275,0.00014373222,0.00005885481,0.000014682906,0.0000038074638,0.0000094689585,0.0011195184],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.997219,0.00031075897,0.00035036803,0.00078098546,0.0008339504,0.0005049222],"domain_scores_gemma":[0.9981298,0.00016334848,0.00011869692,0.0008283496,0.0005287132,0.00023107823],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00059933646,0.00021923968,0.00031470624,0.00012676713,0.0005707169,0.00018904613,0.0008107464,0.0001360897,0.00009591768],"category_scores_gemma":[0.00028194784,0.00019228026,0.00012807144,0.0011195561,0.00019792031,0.0007599778,0.00028444678,0.00031471235,0.00022221949],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00029971392,0.0034582985,0.23575763,0.0022049795,0.0007056441,0.0007795676,0.034408577,0.026887435,0.03340631,0.5334688,0.05856501,0.07005802],"study_design_scores_gemma":[0.0010699765,0.00041625672,0.016053567,0.00033406654,0.000031281113,0.0004816533,0.0005233535,0.92719364,0.026648717,0.004590266,0.021598188,0.0010590161],"about_ca_topic_score_codex":0.0005566097,"about_ca_topic_score_gemma":0.000024146148,"teacher_disagreement_score":0.9003062,"about_ca_system_score_codex":0.00033135555,"about_ca_system_score_gemma":0.0006012353,"threshold_uncertainty_score":0.7840961},"labels":[],"label_agreement":null},{"id":"W2314148015","doi":"10.1111/j.1937-5956.2010.01215.x","title":"Optimal Enhancement and Lifetime of Software Systems: A Control Theoretic Analysis","year":2011,"lang":"en","type":"article","venue":"Production and Operations Management","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science; Realization (probability); Lag; Maturity (psychological); Software; Process (computing); Control (management); Optimal control; Quality (philosophy); Risk analysis (engineering); Industrial engineering; Mathematical optimization; Operations research; Artificial intelligence; Mathematics; Statistics","score_opus":0.01353451273869412,"score_gpt":0.24160062437126223,"score_spread":0.2280661116325681,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2314148015","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05915417,0.0005332435,0.9391366,0.00042555714,0.00006576071,0.00039421904,0.0000016218618,0.000044110697,0.00024472116],"genre_scores_gemma":[0.95126665,0.00049634237,0.04744968,0.000023614031,0.000015471207,0.00009209645,0.0000025518812,0.0000026489495,0.00065095257],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99893904,0.00010833834,0.00024135299,0.0003774775,0.00021360425,0.0001201987],"domain_scores_gemma":[0.999349,0.000019995665,0.000039866016,0.0003933238,0.00015149757,0.000046343797],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00060652464,0.00008306585,0.00017612183,0.00031034913,0.00016027706,0.00010945053,0.00016401481,0.000019990426,0.00004227744],"category_scores_gemma":[0.00007110271,0.000069230395,0.000045280365,0.00057590473,0.00010037655,0.00023827812,0.00012917402,0.000045350003,0.000005537239],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010215507,0.0016464522,0.03077443,0.0010734801,0.0062875473,0.000012204865,0.00947786,0.103466414,0.00035905314,0.7770986,0.00038521778,0.06931658],"study_design_scores_gemma":[0.0011213954,0.0005617393,0.035692044,0.00011661053,0.0021026633,0.000013319595,0.0015897357,0.9520082,0.0027644068,0.002254691,0.0011202411,0.0006549465],"about_ca_topic_score_codex":0.000059576454,"about_ca_topic_score_gemma":0.0000070200254,"teacher_disagreement_score":0.8921125,"about_ca_system_score_codex":0.000016511167,"about_ca_system_score_gemma":0.000011068705,"threshold_uncertainty_score":0.28231335},"labels":[],"label_agreement":null},{"id":"W2322933176","doi":"10.3182/20120523-3-ro-2023.00312","title":"Algorithms for Generating Minimal Cutsets in R-networks","year":2012,"lang":"en","type":"article","venue":"IFAC Proceedings Volumes","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Correctness; Binary decision diagram; Benchmark (surveying); Set (abstract data type); Reliability (semiconductor); Computer science; Reduction (mathematics); Process (computing); Algorithm; Computation; Binary number; Theoretical computer science; Mathematics; Arithmetic","score_opus":0.030982293906998953,"score_gpt":0.30290310448482083,"score_spread":0.27192081057782186,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2322933176","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4455583,0.0013559029,0.55084366,0.00074716605,0.0003092782,0.00045903568,0.0000028957036,0.0002514443,0.0004723304],"genre_scores_gemma":[0.83779126,0.000029940647,0.16120735,0.000114505034,0.00045281416,0.00013081612,0.0000027978824,0.000012489738,0.00025802574],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9981454,0.000014405656,0.00030843914,0.00040437852,0.00032329868,0.00080407166],"domain_scores_gemma":[0.9992115,0.0001743698,0.00007832564,0.00015398313,0.00020403929,0.0001778132],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001246,0.00015228655,0.00023616367,0.0001775851,0.00016957152,0.0002831625,0.0006496581,0.00011480304,0.000009820754],"category_scores_gemma":[0.00048112764,0.00014319865,0.00011354396,0.00068504974,0.00005185747,0.0011273999,0.00024965743,0.00020089692,0.000016690941],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019843053,0.0004080678,0.5457806,0.00016790789,0.00006425312,0.0000023658065,0.003873951,0.0005316789,0.0016367204,0.009119798,0.0056979065,0.4326969],"study_design_scores_gemma":[0.00029852448,0.00006844724,0.009630705,0.000029814977,0.000006382194,0.0000075734097,0.00014012936,0.9838332,0.0011097867,0.0005674253,0.0040699667,0.00023806265],"about_ca_topic_score_codex":0.000037764963,"about_ca_topic_score_gemma":0.0000047240146,"teacher_disagreement_score":0.9833015,"about_ca_system_score_codex":0.00008119417,"about_ca_system_score_gemma":0.000040975738,"threshold_uncertainty_score":0.5839471},"labels":[],"label_agreement":null},{"id":"W2327755579","doi":"10.2514/6.2000-1227","title":"Analyzing real cost of past orbital failures for satellite test effectiveness and insurance","year":2000,"lang":"en","type":"article","venue":"18th International Communications Satellite Systems Conference and Exhibit","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Space Agency","funders":"","keywords":"Computer science; Satellite; Test (biology); Reliability engineering; Engineering; Geology; Aerospace engineering","score_opus":0.03636082191364724,"score_gpt":0.31534338725651023,"score_spread":0.278982565342863,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2327755579","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89643437,0.012508221,0.074866064,0.0026415922,0.00020170085,0.0018691539,0.00033791037,0.00020318455,0.01093782],"genre_scores_gemma":[0.9836469,0.013082711,0.0028262404,0.000009933739,0.00003488166,0.00013872786,0.00006111648,0.000008552571,0.00019090559],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99832284,0.0002722768,0.00052064314,0.00039573608,0.00027334222,0.00021515126],"domain_scores_gemma":[0.9961513,0.002005759,0.00016085133,0.0009333919,0.00064670114,0.00010194763],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009529561,0.00016552355,0.00033682684,0.00019675221,0.00024660674,0.00039324598,0.001240723,0.00008309618,0.000009839233],"category_scores_gemma":[0.00016871876,0.0001538908,0.0000803335,0.000337628,0.00031353516,0.00053184404,0.00029735398,0.00016540207,0.000009397064],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007702567,0.00024430724,0.47285828,0.00037613884,0.00021307272,0.0000024506724,0.0011322418,0.00015846893,0.0019297797,0.07381948,0.000015604966,0.44917315],"study_design_scores_gemma":[0.0018687124,0.00037443062,0.80475396,0.0013997504,0.000051526487,0.000093605544,0.0006299282,0.12279648,0.0013132745,0.0041031563,0.06177964,0.00083555473],"about_ca_topic_score_codex":0.0007562548,"about_ca_topic_score_gemma":0.00011780679,"teacher_disagreement_score":0.44833758,"about_ca_system_score_codex":0.0000448491,"about_ca_system_score_gemma":0.0000689798,"threshold_uncertainty_score":0.62754846},"labels":[],"label_agreement":null},{"id":"W2331387892","doi":"10.4204/eptcs.157.3","title":"Automatic Device Driver Synthesis Project: a Work-in-Progress Report","year":2014,"lang":"en","type":"article","venue":"Electronic Proceedings in Theoretical Computer Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Work (physics); Computer science; Engineering; Systems engineering; Mechanical engineering","score_opus":0.008160420653458223,"score_gpt":0.27745660586082815,"score_spread":0.2692961852073699,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2331387892","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6564852,0.000114004026,0.33761698,0.0026724245,0.00016114685,0.00061734085,1.5081929e-7,0.000464351,0.0018684117],"genre_scores_gemma":[0.9421119,0.000011052666,0.057523247,0.00014472676,0.000065348606,0.0001239088,2.1686272e-7,0.000012426696,0.000007140288],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99440736,0.00012527882,0.00067465456,0.0015584637,0.0014620115,0.0017722563],"domain_scores_gemma":[0.9978341,0.00074512133,0.00015327417,0.0007149498,0.00034158517,0.00021095686],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.008681424,0.0002911275,0.00047658634,0.0009470056,0.0002562385,0.0007870674,0.003964609,0.00011777144,0.000020838837],"category_scores_gemma":[0.0020696232,0.00024639032,0.0001111559,0.0077720657,0.0020584255,0.0011164128,0.0012563526,0.00073648675,0.000034121862],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010579465,0.0001863644,0.010922924,0.00004914945,0.0000067750116,0.000034563764,0.000694759,0.000024653267,0.00004465484,0.84780556,0.000026512585,0.1401935],"study_design_scores_gemma":[0.00024125702,0.00018215104,0.008244586,0.00023704581,0.000006384288,0.00014030054,0.000016201848,0.88635665,0.001474082,0.10247214,0.00023555415,0.0003936442],"about_ca_topic_score_codex":0.000013320102,"about_ca_topic_score_gemma":0.0000057237025,"teacher_disagreement_score":0.886332,"about_ca_system_score_codex":0.00061721826,"about_ca_system_score_gemma":0.0006762972,"threshold_uncertainty_score":0.9999988},"labels":[],"label_agreement":null},{"id":"W2345191262","doi":"","title":"Space Software Validation using Abstract Interpretation","year":2009,"lang":"en","type":"article","venue":"HAL (Le Centre pour la Communication Scientifique Directe)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Geomembrane Technologies (Canada)","funders":"","keywords":"Software deployment; Software; Interpretation (philosophy); Computer science; Computation; Verification and validation; Abstract interpretation; Space (punctuation); Reliability engineering; Systems engineering; Data mining; Engineering; Software engineering; Algorithm; Operating system; Operations management; Programming language","score_opus":0.01815376046744835,"score_gpt":0.26773527139042436,"score_spread":0.249581510922976,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2345191262","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12238599,0.00014844695,0.86735153,0.005967784,0.000060080893,0.00012990055,0.0000019339739,0.00027162346,0.003682682],"genre_scores_gemma":[0.808704,0.000045796274,0.19049706,0.0000829643,0.000012806432,0.0000042459424,0.000029015693,0.0000076213882,0.00061649247],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9967492,0.0016159727,0.0003329783,0.000511837,0.00050078996,0.00028917773],"domain_scores_gemma":[0.9958675,0.0011198547,0.00019716604,0.0012771271,0.0013970482,0.00014131493],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003893189,0.00015222651,0.00017804165,0.00020633115,0.00032966404,0.0005875402,0.0010940921,0.00009467985,0.000059193597],"category_scores_gemma":[0.0024451355,0.00015778613,0.000138728,0.00084995176,0.00007832086,0.00071699556,0.00021109452,0.00023416725,0.000055704488],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000018572237,0.0010590377,0.0053771576,0.000075227435,0.00007136896,0.000012359917,0.013663689,0.0019127768,0.031988118,0.14867778,0.00065248495,0.79649144],"study_design_scores_gemma":[0.00070202973,0.000002128664,0.024940362,0.00085144397,0.000034443463,0.00002600074,0.00012961862,0.71885175,0.21674435,0.034033522,0.0029976596,0.0006866952],"about_ca_topic_score_codex":0.00027706503,"about_ca_topic_score_gemma":0.000052600873,"teacher_disagreement_score":0.79580474,"about_ca_system_score_codex":0.000116431176,"about_ca_system_score_gemma":0.00014907985,"threshold_uncertainty_score":0.64343315},"labels":[],"label_agreement":null},{"id":"W2350159543","doi":"","title":"A Method of Data Flow Analysis Based on Visual C","year":2008,"lang":"en","type":"article","venue":"Microcomputer applications","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Data-flow analysis; Slicing; Program slicing; Data flow diagram; Key (lock); Software; Flow (mathematics); Code (set theory); Program optimization; Programming language; Program analysis; Theoretical computer science; Data mining; Algorithm; Parallel computing; Database; Computer graphics (images); Operating system; Compiler","score_opus":0.04909596633653873,"score_gpt":0.3688014728840901,"score_spread":0.31970550654755137,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2350159543","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00035093195,0.00003009726,0.9985073,0.0005873594,0.0000029509542,0.00024842934,0.00003787698,0.00010951354,0.00012553022],"genre_scores_gemma":[0.045206115,0.000009582569,0.9542038,0.00027142008,0.00004202881,0.000077734454,0.00014275273,0.0000060667444,0.00004052346],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99812436,0.00014813592,0.0003363953,0.0007606151,0.00041280157,0.00021767843],"domain_scores_gemma":[0.99707025,0.00055269356,0.00009445266,0.001983358,0.00019614337,0.000103122],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005534826,0.00012647465,0.00030810465,0.00053816056,0.00019214903,0.000050230487,0.0025163116,0.000057300687,0.000026256504],"category_scores_gemma":[0.0000036342112,0.000116945484,0.0002038354,0.0035560702,0.000079172736,0.00021178398,0.00062289066,0.00014046273,0.00006618564],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006701798,0.0010482346,0.0035416544,0.000030849227,0.00044283352,0.000004371247,0.00017485664,0.04098891,0.0011233073,0.0012707764,0.0017524373,0.94961506],"study_design_scores_gemma":[0.00015082482,0.000029396891,0.003050547,0.0000032920912,0.000065630884,0.0000038237167,0.0000016343829,0.96031445,0.0017797011,0.00019878379,0.03428758,0.00011432367],"about_ca_topic_score_codex":0.000053408254,"about_ca_topic_score_gemma":0.000006209876,"teacher_disagreement_score":0.94950074,"about_ca_system_score_codex":0.000029648045,"about_ca_system_score_gemma":0.00013604313,"threshold_uncertainty_score":0.47688982},"labels":[],"label_agreement":null},{"id":"W2352401519","doi":"","title":"Research for Identifying Potential Vulnerabilities Method Based On Fault Injection Algorithm","year":2005,"lang":"en","type":"article","venue":"Microcomputer applications","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Algorithm; Fault (geology); Computer security; Seismology; Geology","score_opus":0.05393402660825343,"score_gpt":0.40615310779162217,"score_spread":0.35221908118336875,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2352401519","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00022193225,0.000056764784,0.9944625,0.003689544,0.00002915179,0.0011104783,0.000018468909,0.00026440478,0.00014673773],"genre_scores_gemma":[0.016900776,0.0000054335424,0.97991407,0.00037925536,0.00057811965,0.0018023786,0.000037490543,0.000017633743,0.00036483092],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99719584,0.00038868823,0.00037421857,0.00088122755,0.0006223158,0.000537699],"domain_scores_gemma":[0.996998,0.0012293867,0.00006232044,0.0008670416,0.00070717756,0.00013604778],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025431823,0.00016542357,0.00020587802,0.00062771526,0.0009364815,0.0005750119,0.0012296349,0.0001190888,0.00002012308],"category_scores_gemma":[0.0000132447285,0.00016363703,0.00022255052,0.0012948322,0.00009757437,0.00035303007,0.00027650915,0.0004279907,0.00016383569],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000053436424,0.00028951553,0.000008116801,0.000036673537,0.000020499423,3.569402e-7,0.00017770035,0.04995553,0.0012610367,0.0045514875,0.0024576075,0.94123614],"study_design_scores_gemma":[0.0003013955,0.000081528364,0.00007764193,0.000011886306,0.000006334456,0.000006028405,0.000036733672,0.8180891,0.006008767,0.006911229,0.16831905,0.00015028541],"about_ca_topic_score_codex":0.00006928829,"about_ca_topic_score_gemma":0.00001034746,"teacher_disagreement_score":0.9410859,"about_ca_system_score_codex":0.00024387015,"about_ca_system_score_gemma":0.00015039487,"threshold_uncertainty_score":0.72027534},"labels":[],"label_agreement":null},{"id":"W2356194526","doi":"","title":"Noninterference-based Integrity Measurement Model","year":2010,"lang":"en","type":"article","venue":"Journal of Sichuan University","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"L'Alliance Boviteq","funders":"","keywords":"Computer science; Process (computing); Data integrity; Scheme (mathematics); Point (geometry); Distributed computing; Inter-process communication; Signal integrity; Reliability engineering; Programming language; Computer security; Operating system; Engineering; Mathematics; Printed circuit board","score_opus":0.04340801443563873,"score_gpt":0.2470507482511037,"score_spread":0.203642733815465,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2356194526","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24091138,0.000008937919,0.75678194,0.0013674978,0.00013281431,0.000026911597,0.0000010112379,0.000019082157,0.0007504369],"genre_scores_gemma":[0.9602018,0.000007354543,0.039631013,0.000048826867,0.00003251024,2.7336858e-8,1.12245466e-7,0.0000018991288,0.000076431475],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9986588,0.00008483207,0.00017554134,0.00014812103,0.000746594,0.0001861568],"domain_scores_gemma":[0.99809635,0.0000856121,0.00016413315,0.00032736914,0.0011392492,0.00018731625],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015378316,0.00007973279,0.00017739447,0.00025108486,0.00012972728,0.00006895218,0.0012950518,0.00006999876,0.000030613344],"category_scores_gemma":[0.00024940158,0.00006822613,0.00020619355,0.00035492255,0.00008866606,0.0004837225,0.00016284139,0.0008226612,0.000010101076],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0015296368,0.0069047534,0.11154015,0.00031240023,0.0014175458,0.001679501,0.005263866,0.05470802,0.26944274,0.22981414,0.019599771,0.2977875],"study_design_scores_gemma":[0.0018917836,0.00048409987,0.009556228,0.00010644807,0.00008010052,0.000042291722,0.00023323408,0.95457345,0.017184375,0.008415009,0.0070489733,0.0003839871],"about_ca_topic_score_codex":0.00003113763,"about_ca_topic_score_gemma":0.00007840165,"teacher_disagreement_score":0.89986545,"about_ca_system_score_codex":0.00016045409,"about_ca_system_score_gemma":0.0007135457,"threshold_uncertainty_score":0.35740992},"labels":[],"label_agreement":null},{"id":"W2360252811","doi":"","title":"A Software Model Checking Mothed and Its Prototype System","year":2010,"lang":"en","type":"article","venue":"Microcomputer applications","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Model checking; Software; Software engineering; Software quality; Verification and validation; Programming language; Reliability engineering; Software development","score_opus":0.012080063173159104,"score_gpt":0.25984315689340937,"score_spread":0.24776309372025027,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2360252811","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016700896,0.00007328414,0.98115534,0.0004152644,0.000010348867,0.0010405134,0.0000051500165,0.00048401335,0.00011521803],"genre_scores_gemma":[0.36630026,0.0000057280186,0.6324681,0.00009467496,0.000076778764,0.000942939,0.000004344829,0.000012188729,0.000094985815],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9987175,0.000027425398,0.0002246819,0.0005621037,0.00019698712,0.00027133702],"domain_scores_gemma":[0.99890363,0.000091810136,0.000059323396,0.0006013924,0.00019969884,0.00014413099],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00028394192,0.00013805255,0.00016894472,0.00013166906,0.00029693593,0.00026812436,0.00091813353,0.00009521964,0.0000033048484],"category_scores_gemma":[0.000004165115,0.00012827938,0.00006843286,0.00047821947,0.00004365786,0.00024364365,0.0004355454,0.00029481613,0.000089427514],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000072944094,0.00036176317,0.0018317729,0.0006338002,0.0000973163,0.0000046642585,0.0012659368,0.0019053447,0.08407832,0.23758245,0.00039345163,0.6718379],"study_design_scores_gemma":[0.000198735,0.000016429862,0.0004053156,0.000020476717,0.000009036934,0.00004337465,0.0000069477446,0.97377115,0.004478625,0.0027830903,0.018045617,0.00022118217],"about_ca_topic_score_codex":0.000007265752,"about_ca_topic_score_gemma":0.000004686349,"teacher_disagreement_score":0.97186583,"about_ca_system_score_codex":0.000028559958,"about_ca_system_score_gemma":0.00008641856,"threshold_uncertainty_score":0.5231081},"labels":[],"label_agreement":null},{"id":"W2372050858","doi":"","title":"EXERGY ANALYSIS AND PRELIMINARY EXPERIMENTAL INVESTIGATION ON A MIXTURE AUTO- CASCADE REFRIGERATOR","year":2001,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"CAE (Canada)","funders":"","keywords":"Zeotropic mixture; Exergy; Cascade; Refrigerant; Thermodynamics; Refrigeration; Exergy efficiency; Refrigerator car; Heat pump and refrigeration cycle; Process engineering; Materials science; Gas compressor; Chemistry; Physics; Chromatography","score_opus":0.022072898685144827,"score_gpt":0.2816664879434276,"score_spread":0.25959358925828274,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2372050858","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8982552,0.00034257903,0.098433375,0.0016200956,0.000035179222,0.00011929899,9.935314e-7,0.000204358,0.0009889045],"genre_scores_gemma":[0.9895531,0.000048959628,0.008678241,0.00050278654,0.000034686407,0.000024463425,0.0000072730836,0.0000047496906,0.0011457279],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9984938,0.00013635757,0.00019334661,0.00052437186,0.00041728027,0.00023482881],"domain_scores_gemma":[0.99903685,0.0001590099,0.0000408883,0.0005069688,0.000059934882,0.0001963307],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002835801,0.00013047048,0.00019943039,0.0003732196,0.00017265971,0.00020226031,0.00035026902,0.00008880359,0.00012419758],"category_scores_gemma":[0.000071995004,0.00010356723,0.00011419781,0.0017543286,0.000079211175,0.00034978494,0.00017224705,0.0001402624,0.000030915067],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007107783,0.0037184926,0.3489302,0.00016964097,0.0050060893,0.0010801295,0.024826247,0.0056471056,0.19462347,0.08236743,0.019993735,0.31292665],"study_design_scores_gemma":[0.0007633891,0.0011381662,0.07818862,0.000034726996,0.0001723335,0.00005575569,0.0004052481,0.67749316,0.23337266,0.0026955928,0.0049587023,0.0007216554],"about_ca_topic_score_codex":0.00022089135,"about_ca_topic_score_gemma":0.00002734458,"teacher_disagreement_score":0.67184603,"about_ca_system_score_codex":0.00006280339,"about_ca_system_score_gemma":0.000044201777,"threshold_uncertainty_score":0.4223349},"labels":[],"label_agreement":null},{"id":"W2386317215","doi":"","title":"Prokaryotic expression purification and identification of human amino-terminal pro-brain natriuretic peptide and its stability","year":2005,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"CAE (Canada)","funders":"","keywords":"Fusion protein; Affinity chromatography; Recombinant DNA; Molecular biology; lac operon; Western blot; Chemistry; Peptide; Chromatography; Biology; Biochemistry; Gene","score_opus":0.0305768336012813,"score_gpt":0.30692059648386283,"score_spread":0.27634376288258156,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2386317215","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96763057,0.0002538042,0.030022074,0.0015368358,0.000010883502,0.00038636054,0.0000016047339,0.0000639907,0.00009386778],"genre_scores_gemma":[0.99580806,0.000024081164,0.0036719663,0.000018744126,0.00002321347,0.00003232935,0.0000043727737,0.0000040738537,0.000413157],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9982993,0.00019428115,0.00043331916,0.000501034,0.00039878415,0.00017327593],"domain_scores_gemma":[0.99886036,0.00020860188,0.00013515467,0.0004795231,0.00022041486,0.00009593304],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012680438,0.00010207368,0.00016477512,0.00014691573,0.00015565006,0.00012899266,0.00033587325,0.00006505403,0.000022303491],"category_scores_gemma":[0.00039624027,0.000085692234,0.00003743378,0.00030566315,0.000121548735,0.0006720852,0.00017996583,0.00011754366,0.000006856469],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001697681,0.00048176714,0.030707518,0.00062221236,0.00001706897,0.0000017924137,0.002528865,0.000024324725,0.90139645,0.01121828,0.00015873404,0.052826017],"study_design_scores_gemma":[0.000449653,0.00018610527,0.19859289,0.00006641381,0.000018920946,0.000015722078,0.0001592353,0.1443972,0.65217,0.0035591542,0.00012698956,0.00025768974],"about_ca_topic_score_codex":0.000012968038,"about_ca_topic_score_gemma":0.000011288922,"teacher_disagreement_score":0.24922642,"about_ca_system_score_codex":0.000037345566,"about_ca_system_score_gemma":0.000040609924,"threshold_uncertainty_score":0.3494428},"labels":[],"label_agreement":null},{"id":"W2395908282","doi":"10.1145/2889160.2889250","title":"Assessing the process of an Eastern European software SME using systemic analysis, GQM, and reliability growth models","year":2016,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Innopolis University","keywords":"Computer science; Documentation; Process management; Workflow; Scope (computer science); Process (computing); Quality (philosophy); New product development; Product (mathematics); Productivity; Knowledge management; Business; Marketing","score_opus":0.039053374199657244,"score_gpt":0.3258543778175788,"score_spread":0.28680100361792155,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2395908282","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47807893,0.00005104261,0.5214625,0.00015703667,0.000011474204,0.00008267553,0.0000015818101,0.00007671519,0.00007806065],"genre_scores_gemma":[0.98371744,0.000014574314,0.01616488,0.00003001209,0.000024275243,0.0000037944583,9.457411e-7,0.000010756085,0.000033318734],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99684024,0.0009576473,0.00050547806,0.00070956425,0.0006600274,0.0003270465],"domain_scores_gemma":[0.9972834,0.00057598046,0.00019387077,0.0011825791,0.0006118199,0.00015237828],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0029286053,0.00017360467,0.00037735177,0.00023225787,0.00022580178,0.00040746378,0.0012529186,0.000053299977,0.0000068688664],"category_scores_gemma":[0.00044150156,0.000087256165,0.0001622364,0.0014626518,0.00024783227,0.0022580738,0.0004479605,0.0001241943,0.000003519632],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019835921,0.00043152418,0.8807344,0.0007505159,0.0006444621,0.000022167604,0.003277839,0.02350301,0.0041541113,0.002963893,0.000006365755,0.083491914],"study_design_scores_gemma":[0.00023779152,0.000048586502,0.027031036,0.00013954868,0.00015495036,0.000020573036,0.00039209792,0.96239734,0.00082373805,0.008515888,0.0000017528558,0.00023667746],"about_ca_topic_score_codex":0.00020530995,"about_ca_topic_score_gemma":0.000008924939,"teacher_disagreement_score":0.93889433,"about_ca_system_score_codex":0.000054074644,"about_ca_system_score_gemma":0.00012307485,"threshold_uncertainty_score":0.39291826},"labels":[],"label_agreement":null},{"id":"W2396528502","doi":"10.1109/cmpsac.2004.1342660","title":"Optimizing the planning and executing of software independent verification and validation (IV&V) in mature organizations","year":2004,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Capability Maturity Model Integration; LeanCMMI; Verification and validation; Computer science; Software quality; Capability Maturity Model; Software engineering; Maturity (psychological); Software; Process (computing); Reliability (semiconductor); Quality (philosophy); Process management; Software development process; Software development; Engineering; Operating system; Operations management","score_opus":0.02137549025939334,"score_gpt":0.2825144821756106,"score_spread":0.26113899191621726,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2396528502","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.35866106,0.00017881217,0.64040446,0.00061266345,0.000009798959,0.00007333531,3.8485032e-7,0.000030519503,0.000028968514],"genre_scores_gemma":[0.9217031,0.000037687958,0.07819581,0.000029773553,0.0000062308222,0.0000034395594,0.000004608088,0.0000029889793,0.000016370777],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9992988,0.000050592746,0.00016366412,0.00019561264,0.00018926327,0.000102054335],"domain_scores_gemma":[0.9994347,0.00016102387,0.0000527287,0.00021796742,0.0001059968,0.00002756202],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00044875077,0.000053574087,0.00008055806,0.00010612341,0.00012816855,0.00012157922,0.00019148306,0.000046609905,0.0000038094634],"category_scores_gemma":[0.0004057051,0.00003952194,0.000010476157,0.00072450767,0.000041532585,0.00027387723,0.00014506781,0.0001207208,0.000001083463],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000020155056,0.0002801556,0.5307132,0.00026943447,0.00008030269,0.000007461203,0.063321516,0.29110053,0.015905198,0.05714117,0.000052776802,0.041108117],"study_design_scores_gemma":[0.0030104779,0.00017010163,0.74341697,0.0007395626,0.00006660649,0.00010126521,0.0077712773,0.1410209,0.056767587,0.045461096,0.00048140553,0.0009927683],"about_ca_topic_score_codex":0.00011164838,"about_ca_topic_score_gemma":0.000021845875,"teacher_disagreement_score":0.56304204,"about_ca_system_score_codex":0.000028780196,"about_ca_system_score_gemma":0.00004877747,"threshold_uncertainty_score":0.1611658},"labels":[],"label_agreement":null},{"id":"W2424668608","doi":"10.1109/syscon.2016.7490556","title":"Formal analysis of fault tree using probabilistic model checking: A solar array case study","year":2016,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Fault tree analysis; Computer science; Probabilistic logic; Probabilistic CTL; Markov chain; Model checking; Reliability engineering; Markov process; Markov model; Reliability (semiconductor); Markov decision process; Probabilistic analysis of algorithms; Real-time computing; Algorithm; Theoretical computer science; Power (physics); Engineering; Machine learning; Artificial intelligence; Mathematics","score_opus":0.07025469306954886,"score_gpt":0.33604602812185747,"score_spread":0.2657913350523086,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2424668608","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.47842285,0.0000057862303,0.5212567,0.000046319532,0.0000073989168,0.00013233407,0.0000020448865,0.000043863434,0.00008270384],"genre_scores_gemma":[0.96972686,0.0000021494516,0.030084798,0.000013675385,0.000009636054,0.00001347127,4.1524257e-7,0.0000057547545,0.00014324868],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9977479,0.00015769879,0.0004886332,0.00054454576,0.0006449398,0.00041629194],"domain_scores_gemma":[0.99796164,0.0003396467,0.00011670909,0.0010665702,0.00037496738,0.00014048017],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012916283,0.00015309938,0.00045822267,0.00059748336,0.00015987411,0.00009197948,0.00073762564,0.00005789349,0.00003320748],"category_scores_gemma":[0.00037207894,0.000092019945,0.00032738218,0.002463861,0.00009414251,0.00074774184,0.0003211762,0.00009069707,0.0000045741567],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000088258086,0.0065530892,0.22553127,0.00017958674,0.007898923,0.0012009876,0.014861697,0.48753795,0.032167863,0.008273763,0.000029080893,0.21567753],"study_design_scores_gemma":[0.00033281636,0.00010652513,0.00053426967,0.000008150123,0.00039268142,0.000031372194,0.00038407385,0.99627763,0.0010170185,0.00076732854,0.0000023671564,0.00014576604],"about_ca_topic_score_codex":0.00067531347,"about_ca_topic_score_gemma":0.0008018311,"teacher_disagreement_score":0.50873965,"about_ca_system_score_codex":0.00010004007,"about_ca_system_score_gemma":0.0001689243,"threshold_uncertainty_score":0.37524647},"labels":[],"label_agreement":null},{"id":"W2462116092","doi":"","title":"Making trade-offs among security and other requirements during system design","year":2012,"lang":"en","type":"dissertation","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Risk analysis (engineering); Computer security; Security testing; Threat model; Computer science; Viewpoints; Software security assurance; Security through obscurity; Security information and event management; Security service; Cloud computing security; Business; Information security","score_opus":0.05915873302006781,"score_gpt":0.3277577736487615,"score_spread":0.2685990406286937,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2462116092","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79387486,0.0019868608,0.1844998,0.000060739705,0.00081407075,0.0012210399,0.0000043994146,0.0009464756,0.01659178],"genre_scores_gemma":[0.99571246,0.00001548032,0.0030509513,0.000013553275,0.000093894516,0.000043473836,0.0000064798205,0.000023164946,0.0010405503],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99761003,0.00027116487,0.00037618983,0.0006064168,0.000673418,0.00046278583],"domain_scores_gemma":[0.9989163,0.00013991699,0.00016951421,0.0005757063,0.000057284284,0.00014123975],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00094394374,0.00025578495,0.00035799242,0.00024549887,0.0002580844,0.00034330602,0.0006934794,0.00026033653,0.000040037827],"category_scores_gemma":[0.000053127602,0.00022322607,0.00012143008,0.00036446177,0.00003506322,0.0005156866,0.00011865256,0.0003098143,0.000022291677],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00047759814,0.0015935436,0.66142,0.036125116,0.003657406,0.00033711692,0.089683816,0.00036137088,0.008515619,0.07197757,0.001442773,0.124408066],"study_design_scores_gemma":[0.0025162133,0.00027174572,0.7671055,0.007000799,0.0005910527,0.00009719656,0.010489645,0.17112106,0.030768871,0.0037817352,0.0015418987,0.00471426],"about_ca_topic_score_codex":0.00015272466,"about_ca_topic_score_gemma":0.000089291636,"teacher_disagreement_score":0.20183761,"about_ca_system_score_codex":0.00014415316,"about_ca_system_score_gemma":0.000059574635,"threshold_uncertainty_score":0.9102895},"labels":[],"label_agreement":null},{"id":"W2462934991","doi":"10.1016/j.advengsoft.2016.06.008","title":"State of the practice for mesh generation and mesh processing software","year":2016,"lang":"en","type":"article","venue":"Advances in Engineering Software","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":11,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Usability; Maintainability; Software engineering; Correctness; Software; Analytic hierarchy process; Reliability engineering; Programming language; Operating system; Operations research; Engineering","score_opus":0.009672580553957708,"score_gpt":0.2764194619632309,"score_spread":0.2667468814092732,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2462934991","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01088546,0.002955465,0.9852722,0.00042296376,0.00012831067,0.00021433514,0.000008027332,0.00011082075,0.0000023972968],"genre_scores_gemma":[0.56251496,0.001076356,0.43610835,0.000056040968,0.000054564898,0.0000998526,0.0000014188951,0.00001876715,0.000069687376],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.998852,0.000042006406,0.00025725117,0.0003229162,0.00027783323,0.00024799973],"domain_scores_gemma":[0.99818385,0.0010966209,0.00010234668,0.0003528309,0.00022030805,0.00004402703],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00047263192,0.00011898944,0.00016964092,0.00009618064,0.00007758514,0.000052311956,0.00044145028,0.000038437112,9.0336636e-7],"category_scores_gemma":[0.0043425104,0.00007484475,0.000053157593,0.00045377662,0.00005830777,0.001472796,0.00015784997,0.000093528644,5.4386686e-7],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001019857,0.000046661436,0.0061425236,0.0003616227,0.00001551873,0.0000016599864,0.0003906582,0.022993175,0.0013001664,0.0009922099,0.000022876471,0.9677227],"study_design_scores_gemma":[0.0049697356,0.0007675025,0.016286079,0.0037030003,0.00011974453,0.0000935586,0.00015468281,0.76898974,0.080632456,0.029860426,0.09206609,0.0023570075],"about_ca_topic_score_codex":0.0000040565,"about_ca_topic_score_gemma":0.0000112409425,"teacher_disagreement_score":0.9653657,"about_ca_system_score_codex":0.00005868927,"about_ca_system_score_gemma":0.00007028379,"threshold_uncertainty_score":0.5198704},"labels":[],"label_agreement":null},{"id":"W2506113519","doi":"10.4018/978-1-4666-0155-0.ch013","title":"From an Ad-Hoc to a Proactive Software Release Plan and Deployment","year":2012,"lang":"en","type":"book-chapter","venue":"IGI Global eBooks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Software deployment; Upgrade; Software; Aerospace; Social software engineering; Software development; Computer science; Engineering; Process management; Engineering management; Systems engineering; Software engineering; Software construction","score_opus":0.02885777167189501,"score_gpt":0.2822997240884614,"score_spread":0.25344195241656636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2506113519","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.023564344,0.014676239,0.35105702,0.0024025633,0.0016557317,0.0070837466,0.003122674,0.0043677753,0.59206986],"genre_scores_gemma":[0.59746295,0.0004032015,0.3060589,0.008317009,0.0035695548,0.00077035837,0.00020770972,0.0005100238,0.08270029],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9970126,0.000077042874,0.00035107206,0.0011482304,0.00084685534,0.0005641552],"domain_scores_gemma":[0.9973347,0.00014705016,0.00012506406,0.0012891334,0.00017941036,0.0009246401],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00023351038,0.0004607456,0.0005402282,0.00012509158,0.00016577322,0.00032778663,0.0010681475,0.00035983743,0.000028153641],"category_scores_gemma":[0.00007969018,0.00042600828,0.0001775405,0.000050977786,0.000108256834,0.00023150555,0.0011963921,0.0004100447,0.00026108036],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011648164,0.000077293196,0.00022440283,0.000035818328,0.0003321749,0.00010768365,0.0010083059,0.0000043871873,0.0000127952835,0.1963321,0.0008482988,0.8009003],"study_design_scores_gemma":[0.000982876,0.000944782,0.0015626337,0.0004789527,0.0003077193,0.000062985666,0.00007562571,0.0007299288,0.0001976043,0.8820622,0.110490024,0.0021046433],"about_ca_topic_score_codex":0.00032815526,"about_ca_topic_score_gemma":0.00018457438,"teacher_disagreement_score":0.79879564,"about_ca_system_score_codex":0.0003628404,"about_ca_system_score_gemma":0.0002811689,"threshold_uncertainty_score":0.99981916},"labels":[],"label_agreement":null},{"id":"W2507969500","doi":"10.1007/978-3-319-44817-6_2","title":"Implementing the New ISO/IEC 29110 Systems Engineering Process Standard in a Small Public Transportation Company","year":2016,"lang":"en","type":"book-chapter","venue":"Communications in computer and information science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":11,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Capability Maturity Model Integration; Maturity (psychological); Process (computing); Engineering management; Engineering; Information security management system; Work (physics); Systems engineering; Computer science; Software development process; Software development; Software","score_opus":0.0454251061000584,"score_gpt":0.30455693402509826,"score_spread":0.25913182792503986,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2507969500","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005332898,0.00056955084,0.9860003,0.0035658511,0.00014361257,0.0006165274,0.00001961698,0.00010096212,0.008450281],"genre_scores_gemma":[0.9366106,0.0036124017,0.05777294,0.0005267525,0.00012820488,0.00017688438,0.00010352508,0.000027468219,0.0010412202],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976688,0.000047762464,0.0009222168,0.00030625606,0.00063144916,0.000423483],"domain_scores_gemma":[0.9969906,0.0004593545,0.000308361,0.0016137229,0.0005050424,0.00012291636],"candidate_categories":["scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0031415988,0.00019908337,0.00028562904,0.0010991776,0.00044921917,0.001352164,0.004070112,0.00009167291,0.000005177518],"category_scores_gemma":[0.00010549382,0.00014954736,0.000056179673,0.0009506725,0.0003411312,0.004789626,0.00077569997,0.00046398767,0.000012539593],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000024925378,0.000012167185,0.00066550344,0.000102764825,0.000011788645,4.7673836e-7,0.004043618,0.0028328786,0.0000026170064,0.75633526,0.00005676749,0.23593365],"study_design_scores_gemma":[0.0004799868,0.000042559677,0.0023647149,0.0006002399,0.0000061058518,0.000008635319,0.000103296436,0.84155506,0.00000732262,0.0059334594,0.14853473,0.0003638782],"about_ca_topic_score_codex":0.000061735474,"about_ca_topic_score_gemma":0.00021446988,"teacher_disagreement_score":0.9360773,"about_ca_system_score_codex":0.00022441923,"about_ca_system_score_gemma":0.0007043038,"threshold_uncertainty_score":0.9996845},"labels":[],"label_agreement":null},{"id":"W2519334373","doi":"10.1002/j.2334-5837.2016.00203.x","title":"Developing and implementing systems engineering and project management processes at CSiT – A small Canadian company in public transportation","year":2016,"lang":"en","type":"article","venue":"INCOSE International Symposium","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Business; Engineering management; Public transport; Telecommunications; Transport engineering; Engineering","score_opus":0.026468774030009332,"score_gpt":0.26201949282078313,"score_spread":0.2355507187907738,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2519334373","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9221509,0.0001848763,0.07114912,0.005605587,0.00014068541,0.00037206005,0.000019682568,0.00007021272,0.00030686284],"genre_scores_gemma":[0.9964262,0.0001853121,0.0031555858,0.00003099508,0.000023147211,0.000071867405,0.00000994947,0.0000055401492,0.00009136594],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9989246,0.000020695516,0.0002414708,0.00030674614,0.00023380936,0.00027268974],"domain_scores_gemma":[0.9995126,0.00012241805,0.00004647341,0.0000995923,0.00013608983,0.000082829916],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00040691392,0.00009241153,0.00010117906,0.00036952118,0.00008237923,0.00024976494,0.0003130573,0.000030045116,0.0000025481602],"category_scores_gemma":[0.00004462778,0.00007569493,0.000014797748,0.00031648081,0.000017822,0.00039720008,0.00011278091,0.000044381708,0.000002335648],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000023792658,0.000101623846,0.7818585,0.0013977381,0.0004284355,0.00027553178,0.003734626,0.00069770205,0.0045170593,0.15329313,0.00009222608,0.053579606],"study_design_scores_gemma":[0.004652577,0.00015446618,0.5751665,0.0032377355,0.00005046291,0.00022410142,0.0007383945,0.22911583,0.0036151395,0.002083376,0.17898802,0.001973363],"about_ca_topic_score_codex":0.011455683,"about_ca_topic_score_gemma":0.06738102,"teacher_disagreement_score":0.22841813,"about_ca_system_score_codex":0.00044917045,"about_ca_system_score_gemma":0.00012327355,"threshold_uncertainty_score":0.99512714},"labels":[],"label_agreement":null},{"id":"W2538976203","doi":"10.1109/sibcon.2005.1611189","title":"Automatic Detection of Local and Global Software Failures","year":2006,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Session (web analytics); Software; Software bug; Reliability engineering; Real-time computing; Software engineering; Operating system; Engineering","score_opus":0.005874874903907173,"score_gpt":0.24015331666410444,"score_spread":0.23427844176019727,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2538976203","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15704301,0.00010529626,0.8423695,0.00015965216,0.000014636958,0.00003575999,6.003644e-7,0.00013044034,0.00014109456],"genre_scores_gemma":[0.9686016,0.0000025075017,0.031331737,0.000014406426,0.000008023188,0.0000025148536,3.9966108e-7,0.0000011980743,0.00003762061],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99926656,0.000041855397,0.00014721777,0.0001769063,0.00024209284,0.00012536353],"domain_scores_gemma":[0.9995491,0.00010306048,0.00002751626,0.0002148268,0.00006915206,0.000036353773],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00016696718,0.00005377309,0.00010418195,0.00004503303,0.000049097005,0.00005515029,0.0002057805,0.00003868121,0.000017944863],"category_scores_gemma":[0.000052255997,0.000042835098,0.00004290904,0.00041043837,0.00009310202,0.00017034763,0.000118961216,0.000038987477,0.000009604387],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000020929376,0.00008202474,0.019881468,0.00007698508,0.000016922211,0.0000033995602,0.00002066723,0.00042163816,0.0004405156,0.013967829,0.00014923641,0.9649372],"study_design_scores_gemma":[0.00036252962,0.00015588022,0.17890881,0.000023868159,0.000013712498,0.000029855766,0.0000648898,0.74765533,0.01756273,0.05436192,0.0006530766,0.00020738435],"about_ca_topic_score_codex":0.0013754938,"about_ca_topic_score_gemma":0.00036504547,"teacher_disagreement_score":0.96472985,"about_ca_system_score_codex":0.000030405236,"about_ca_system_score_gemma":0.000028099515,"threshold_uncertainty_score":0.20793444},"labels":[],"label_agreement":null},{"id":"W2547521629","doi":"","title":"Adaptive Practice on Software Reliability Based on IEEE Std. 1633 in Frequent Requirement Modifications.","year":2010,"lang":"en","type":"article","venue":"Computer and Information Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science; Reliability engineering; Software development; Software requirements; Software quality; Reliability (semiconductor); Software construction; Software; Software reliability testing; Software development process; Verification and validation; Avionics software; Software engineering; Software sizing; Software release life cycle; Software deployment; Operating system; Engineering","score_opus":0.025974002994658935,"score_gpt":0.30663890803126487,"score_spread":0.2806649050366059,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2547521629","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0340342,0.0000035749586,0.9609604,0.0021238176,0.00039984848,0.00034071252,0.000004062388,0.00011093429,0.0020224457],"genre_scores_gemma":[0.82851607,0.000012036882,0.1690697,0.002328922,0.000030033603,0.00003404632,0.0000027695069,0.0000020958632,0.0000042940337],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9974589,0.00010741701,0.0004519518,0.0004689931,0.0011841453,0.00032862028],"domain_scores_gemma":[0.99731094,0.00065968296,0.00015588893,0.0009866345,0.0006973621,0.00018947232],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0032763008,0.00014265177,0.00014837952,0.0006391857,0.00036544973,0.00063861825,0.0010777807,0.0000631325,0.000009351158],"category_scores_gemma":[0.0011308935,0.00011916162,0.000045277793,0.0016835654,0.00038140005,0.007840613,0.00019200942,0.00042772098,0.000091547685],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013099823,0.0010388802,0.0033024806,0.00008039324,0.000010362567,0.000005638803,0.004431388,0.16566223,0.00033728982,0.14626303,0.0012485724,0.67748874],"study_design_scores_gemma":[0.000311454,0.0002649885,0.013871742,0.000028052687,0.0000014232904,0.0000025252698,0.000035479075,0.9802982,0.00074413593,0.0007337551,0.0035556115,0.00015263347],"about_ca_topic_score_codex":0.00004872243,"about_ca_topic_score_gemma":0.0000034388886,"teacher_disagreement_score":0.814636,"about_ca_system_score_codex":0.00016433094,"about_ca_system_score_gemma":0.00040522302,"threshold_uncertainty_score":0.61582106},"labels":[],"label_agreement":null},{"id":"W2551931692","doi":"10.7287/peerj.preprints.2601v1","title":"Anatomy of a crash repository","year":2016,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Crash; Computer science; Parsing; Debugging; Metadata; Data science; Programming language; World Wide Web","score_opus":0.010626706529955912,"score_gpt":0.2863821910409885,"score_spread":0.2757554845110326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2551931692","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0809125,0.000044829558,0.9049967,0.0018987572,0.00004801594,0.00003878649,2.6578175e-7,0.000106284824,0.0119539],"genre_scores_gemma":[0.98780745,0.000011002117,0.009232475,0.000027642373,0.000011572786,0.000002292223,2.6351668e-8,0.0000012240278,0.0029063374],"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9993345,0.000043499687,0.00012503413,0.00016263529,0.00022844794,0.0001058875],"domain_scores_gemma":[0.999159,0.00020938524,0.000024862704,0.00045794647,0.000100795405,0.00004805937],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00021029553,0.000032973447,0.000077929864,0.00006432996,0.000025335303,0.000018354936,0.0004919702,0.000021970036,0.00005032238],"category_scores_gemma":[0.00009931914,0.000017688133,0.00006127959,0.00025829455,0.000054628956,0.00017532929,0.00014610453,0.000022178385,0.000039722516],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000061142596,0.0002549504,0.19535248,0.000043554504,0.00008850536,0.000029480809,0.00018633188,0.0000049455984,0.07676595,0.21267062,0.006063477,0.5085336],"study_design_scores_gemma":[0.001131892,0.00028962476,0.10289922,0.00012389357,0.000013401733,0.00003402248,0.000038999024,0.012034323,0.80073845,0.047569714,0.03464899,0.00047744001],"about_ca_topic_score_codex":0.000033946497,"about_ca_topic_score_gemma":0.0000021687906,"teacher_disagreement_score":0.9068949,"about_ca_system_score_codex":0.000014319805,"about_ca_system_score_gemma":0.00004738933,"threshold_uncertainty_score":0.09142113},"labels":[],"label_agreement":null},{"id":"W2575734139","doi":"10.1016/j.scico.2016.12.001","title":"Preface to the special issue on advances in software measurement","year":2017,"lang":"en","type":"article","venue":"Science of Computer Programming","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Computer science; Software engineering; Software; Data science; Programming language","score_opus":0.03234619304457469,"score_gpt":0.32610259047257245,"score_spread":0.29375639742799775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2575734139","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.023108007,0.00022878118,0.9647377,0.008265199,0.0015757203,0.0009821098,8.833877e-7,0.000112949245,0.0009886326],"genre_scores_gemma":[0.7302114,0.000027117858,0.26813114,0.00015889909,0.0013751327,0.00004887538,1.5068188e-7,0.000006769049,0.000040522464],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99658924,0.00007302913,0.0002755581,0.0006298099,0.0019020112,0.000530329],"domain_scores_gemma":[0.99767417,0.000104883875,0.00014548864,0.0015278503,0.0004050757,0.00014251792],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003836535,0.00012857308,0.00020213342,0.00025344745,0.0008043093,0.0010054782,0.0053403117,0.000027004326,0.000006065218],"category_scores_gemma":[0.0008748343,0.000087665656,0.00007647587,0.00093145756,0.0005333058,0.0011836329,0.0012307487,0.00018084934,0.000048487273],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000005264049,0.00008190054,0.0023012615,0.00001471565,0.0000024783028,0.0000025621662,0.00064030616,0.002698783,0.000026562244,0.001212984,0.0001151349,0.99289805],"study_design_scores_gemma":[0.0011811091,0.0019954003,0.13405918,0.0010995413,0.000012026274,0.000010291812,0.00022323431,0.10833717,0.014864804,0.007866233,0.7292543,0.0010967246],"about_ca_topic_score_codex":0.00005174357,"about_ca_topic_score_gemma":0.000120579636,"teacher_disagreement_score":0.9918013,"about_ca_system_score_codex":0.00012567698,"about_ca_system_score_gemma":0.00024490684,"threshold_uncertainty_score":0.9923717},"labels":[],"label_agreement":null},{"id":"W2576160198","doi":"10.1109/quatic.2016.021","title":"A Fully Automated Approach to Discovering Nondeterminism in State Machine Diagrams","year":2016,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Computer science; Leverage (statistics); Model checking; Abstract state machines; State (computer science); Set (abstract data type); False positive paradox; Programming language; Finite-state machine; Theoretical computer science; Artificial intelligence","score_opus":0.01426978587479598,"score_gpt":0.27059275465323757,"score_spread":0.2563229687784416,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2576160198","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.18882036,0.00001414123,0.808113,0.0010101297,0.000033149765,0.00016813609,0.0000031689347,0.00049364497,0.0013442916],"genre_scores_gemma":[0.97126895,0.000015932435,0.027398594,0.00014364353,0.000010200055,0.000044594886,0.000001015607,0.000007349727,0.0011097072],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9983263,0.00009775497,0.00025783377,0.0005128504,0.00037027075,0.0004349643],"domain_scores_gemma":[0.9989416,0.00019259214,0.000025694591,0.00062424457,0.00004017589,0.00017569512],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005515625,0.00012569534,0.00019780594,0.00023941956,0.000045579825,0.0001711379,0.00088979217,0.00003527223,0.000015846066],"category_scores_gemma":[0.00016162434,0.00007344891,0.00006888736,0.0008713222,0.000035165893,0.00044346982,0.0005446329,0.00008307191,0.00015336968],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000032511507,0.00074323843,0.04687588,0.00006500582,0.00005099818,0.00007185224,0.0021302009,0.0019947132,0.00744614,0.006478816,0.0004989392,0.9336117],"study_design_scores_gemma":[0.00085000717,0.0001573642,0.04320564,0.00006165008,0.000002579699,0.0000117522995,0.000026006199,0.9474885,0.0038274035,0.0025880707,0.0013332109,0.000447812],"about_ca_topic_score_codex":0.00034918025,"about_ca_topic_score_gemma":0.00015209474,"teacher_disagreement_score":0.9454938,"about_ca_system_score_codex":0.00008377583,"about_ca_system_score_gemma":0.000059222966,"threshold_uncertainty_score":0.29951596},"labels":[],"label_agreement":null},{"id":"W2587084920","doi":"","title":"Classifying And Predicting Software Security Vulnerabilities based on Reproducibility","year":2017,"lang":"en","type":"dissertation","venue":"QSpace (Queen's University Library)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Canada Research Chairs","keywords":"Reproducibility; Computer science; Software; Software security assurance; Computer security; Data mining; Information security; Statistics; Mathematics; Operating system","score_opus":0.013112757596827607,"score_gpt":0.23680625563489907,"score_spread":0.22369349803807145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2587084920","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.7990282,0.0003958808,0.026629051,0.10463002,0.0025401309,0.003460701,0.00045777945,0.00678534,0.056072854],"genre_scores_gemma":[0.8456838,0.0003707312,0.013618075,0.00016075473,0.00025471975,0.000008571652,0.00058394566,0.00008504146,0.13923433],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99479884,0.0006853409,0.00029870108,0.0027923808,0.0008479869,0.00057673635],"domain_scores_gemma":[0.9935568,0.0011650721,0.00040662818,0.004320662,0.00021238302,0.0003384522],"candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.0008095604,0.00048588862,0.0006643477,0.00059808313,0.0013832406,0.0008407197,0.0021120794,0.0005301581,0.00011947224],"category_scores_gemma":[0.002588005,0.00052381936,0.0003381094,0.0005453259,0.00026705125,0.0027237749,0.00055278203,0.0013108521,0.00001659877],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0013793075,0.0016304422,0.8067802,0.008065982,0.0007290318,0.0009219619,0.013040583,0.0016673561,0.0000110349365,0.009346394,0.093009464,0.06341822],"study_design_scores_gemma":[0.006367741,0.0025682945,0.5342182,0.007875938,0.0009600691,0.0000041752705,0.016565017,0.061852433,0.016783446,0.049185492,0.29379535,0.009823876],"about_ca_topic_score_codex":0.0041698297,"about_ca_topic_score_gemma":0.00016603149,"teacher_disagreement_score":0.27256206,"about_ca_system_score_codex":0.00019978672,"about_ca_system_score_gemma":0.0009249078,"threshold_uncertainty_score":0.99991685},"labels":[],"label_agreement":null},{"id":"W2587695526","doi":"10.29173/cais7","title":"Problems With COTS Software: A Case Study","year":2013,"lang":"en","type":"article","venue":"Proceedings of the Annual Conference of CAIS / Actes du congrès annuel de l ACSI","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Software; Pace; Software engineering; Commercial off-the-shelf; Construct (python library); Computer science; Studio; Engineering; Operating system; Telecommunications","score_opus":0.02412307536920328,"score_gpt":0.2530033733102407,"score_spread":0.22888029794103742,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2587695526","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99488825,0.00006461915,0.0020407857,0.0012828922,0.000041599902,0.0010085726,0.000013667119,0.00011216659,0.0005474569],"genre_scores_gemma":[0.9953435,0.000018973498,0.004045728,0.000081819286,0.00003061501,0.00015269963,6.009043e-7,0.000017835206,0.00030819376],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9972828,0.000060146893,0.00056226656,0.00058177754,0.0009589211,0.0005540713],"domain_scores_gemma":[0.94381154,0.00033882805,0.0005614742,0.0005512071,0.054507382,0.00022955761],"candidate_categories":["metaresearch","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.00094239117,0.00030131597,0.00056389876,0.00022550953,0.00022784142,0.0043675634,0.0032462957,0.00011325626,0.00002878516],"category_scores_gemma":[0.0095439255,0.0001950143,0.00016726162,0.0012560757,0.0004695734,0.011738703,0.0012892826,0.00038972334,0.000010355011],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004765669,0.0012218695,0.89918905,0.0004957623,0.00033737475,0.000050721916,0.06370257,0.000045811194,0.0018536902,0.0023722223,0.0014050149,0.029278278],"study_design_scores_gemma":[0.01104708,0.015890216,0.71267825,0.002756667,0.0009930321,0.0075447108,0.07881942,0.04085957,0.045535102,0.05100961,0.027825546,0.0050407816],"about_ca_topic_score_codex":0.002315963,"about_ca_topic_score_gemma":0.000078604826,"teacher_disagreement_score":0.18651076,"about_ca_system_score_codex":0.000056483874,"about_ca_system_score_gemma":0.00036774503,"threshold_uncertainty_score":0.9987991},"labels":[],"label_agreement":null},{"id":"W2589239371","doi":"10.71781/10596","title":"Software stability assessment using multiple prediction models","year":2003,"lang":"fr","type":"dissertation","venue":"Open MIND","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Stability (learning theory); Computer science; Machine learning","score_opus":0.16537823285430694,"score_gpt":0.3908664435388779,"score_spread":0.22548821068457095,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2589239371","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.40226522,0.00034773196,0.5924784,0.00008095454,0.0009325755,0.001386177,0.00022026351,0.0000099400795,0.0022787054],"genre_scores_gemma":[0.41863972,0.00013568223,0.57246596,0.00002835519,0.000120436955,0.00013035373,0.001114281,0.000057786016,0.00730742],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99359417,0.001130746,0.0010890039,0.0019705605,0.0014020851,0.0008134167],"domain_scores_gemma":[0.9955358,0.00076527597,0.00045978703,0.0016228134,0.0012302306,0.00038612937],"candidate_categories":["metaepi_narrow","scholarly_communication","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.003014177,0.00054705865,0.00083405175,0.00023306062,0.0009123144,0.0022357532,0.0023516598,0.0006237901,0.002990468],"category_scores_gemma":[0.0009132619,0.0005789681,0.0003845588,0.0012300322,0.00017642858,0.002832792,0.00070921157,0.000983969,0.00019786943],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014406539,0.002774764,0.07959821,0.00061560527,0.000614586,0.000058747402,0.010782866,0.121512644,0.0024756654,0.00040752292,0.00013037423,0.7808849],"study_design_scores_gemma":[0.0006444777,0.00014438681,0.0031602294,0.0003627909,0.00017187826,0.000010070927,0.0015181688,0.9867554,0.002599856,0.0013746199,0.0026989751,0.0005591574],"about_ca_topic_score_codex":0.0025955846,"about_ca_topic_score_gemma":0.0014144712,"teacher_disagreement_score":0.8652427,"about_ca_system_score_codex":0.0010290894,"about_ca_system_score_gemma":0.0031406179,"threshold_uncertainty_score":0.99966615},"labels":[],"label_agreement":null},{"id":"W2594988625","doi":"10.1177/1748006x17694494","title":"Formal reliability analysis of oil and gas pipelines","year":2017,"lang":"en","type":"article","venue":"Proceedings of the Institution of Mechanical Engineers Part O Journal of Risk and Reliability","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"Qatar National Research Fund","keywords":"Pipeline transport; Reliability (semiconductor); Computer science; Monte Carlo method; Pipeline (software); Reliability engineering; Environmental science; Engineering; Mathematics; Statistics","score_opus":0.012012727037487446,"score_gpt":0.2573335610753184,"score_spread":0.24532083403783095,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2594988625","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.968816,0.00027194893,0.029437378,0.0011000554,0.00021255879,0.00006033836,0.000015803018,0.000010066371,0.00007587708],"genre_scores_gemma":[0.9875296,0.0027757932,0.009635536,0.0000052670616,0.000039107374,0.0000014796526,2.4088263e-7,0.000003248223,0.000009722854],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975798,0.000050071423,0.0011189502,0.00028630704,0.00075066235,0.0002142066],"domain_scores_gemma":[0.9958525,0.00038858072,0.0014596127,0.00051023706,0.0016207005,0.00016838842],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.005417891,0.00015909717,0.0007892074,0.00024897722,0.000285488,0.00006271474,0.0011739682,0.00015189957,0.0000036083204],"category_scores_gemma":[0.010086042,0.00009922034,0.00052127475,0.0005863416,0.0007973233,0.0009397256,0.00056814856,0.00043242774,1.0011812e-7],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0017113979,0.0029313823,0.47345015,0.0045066504,0.003213534,0.0000042688166,0.0033708606,0.062203407,0.014033372,0.14247742,0.00031081992,0.29178673],"study_design_scores_gemma":[0.003739384,0.0013172448,0.21239467,0.0011509818,0.004322018,0.000066132045,0.0006331225,0.61800355,0.092806764,0.062540695,0.002250081,0.00077534345],"about_ca_topic_score_codex":0.00010529718,"about_ca_topic_score_gemma":0.0000038319577,"teacher_disagreement_score":0.55580014,"about_ca_system_score_codex":0.00004084393,"about_ca_system_score_gemma":0.000112747955,"threshold_uncertainty_score":0.99825245},"labels":[],"label_agreement":null},{"id":"W2595846720","doi":"10.71781/10046","title":"Determination of software quality through a generic model","year":2003,"lang":"en","type":"dissertation","venue":"Open MIND","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Software quality; Quality (philosophy); Computer science; Software engineering; Software; Software development; Programming language; Philosophy; Epistemology","score_opus":0.11019280490903317,"score_gpt":0.41271736633247363,"score_spread":0.3025245614234405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2595846720","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09048484,0.00032579666,0.89678556,0.000039349783,0.00016864149,0.0005439604,0.00003200267,0.000006845435,0.011612994],"genre_scores_gemma":[0.020483209,0.00011773183,0.95290965,0.00003118274,0.000017688064,0.00006470595,0.00036801168,0.000020641475,0.025987199],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976337,0.00023070659,0.00056363,0.00067704235,0.0006579389,0.0002369895],"domain_scores_gemma":[0.9979751,0.00018089674,0.0003879819,0.0009430591,0.00045086155,0.00006211118],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00088097464,0.00019865621,0.00047974108,0.00012161283,0.00010988101,0.00028362087,0.0020707727,0.00025284992,0.00024157978],"category_scores_gemma":[0.0005060813,0.00019077351,0.00018706087,0.00071888533,0.00003579069,0.0006942395,0.0002148166,0.0002255056,0.00007802487],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022304035,0.00022187977,0.00010936909,0.00016326141,0.000042371285,0.0000042448182,0.0027697172,0.00067755935,0.00056378124,0.00042971908,0.00013681805,0.994859],"study_design_scores_gemma":[0.002581349,0.0005334566,0.0023761124,0.0010304983,0.00040102925,0.000020614641,0.0019633318,0.74831665,0.13756476,0.081603155,0.01990057,0.0037084816],"about_ca_topic_score_codex":0.00014975661,"about_ca_topic_score_gemma":0.00024048425,"teacher_disagreement_score":0.9911505,"about_ca_system_score_codex":0.000075179814,"about_ca_system_score_gemma":0.00066207483,"threshold_uncertainty_score":0.77795184},"labels":[],"label_agreement":null},{"id":"W2597503483","doi":"10.1109/saner.2017.7884643","title":"An exploratory study on library aging by monitoring client usage in a software ecosystem","year":2017,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":20,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"Japan Society for the Promotion of Science","keywords":"Leverage (statistics); Computer science; Software; Ecosystem; World Wide Web; Code (set theory); Data science; Operating system; Ecology","score_opus":0.0409794645603257,"score_gpt":0.32662985896629604,"score_spread":0.28565039440597034,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2597503483","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.97625387,0.0000524929,0.022390408,0.00023623435,0.0001844517,0.0002654451,0.000002578896,0.0003235984,0.00029090763],"genre_scores_gemma":[0.9974695,0.000014976464,0.0021825954,0.00003473313,0.000073411284,0.000056039793,0.0000011137706,0.000012697681,0.00015493644],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99777323,0.00033704375,0.0002825794,0.0006774779,0.00056082616,0.00036884804],"domain_scores_gemma":[0.99747354,0.0002042992,0.000086999484,0.0020135823,0.00003425991,0.00018729614],"candidate_categories":["scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0009526709,0.00015222162,0.00023396983,0.00020638168,0.00044409686,0.0013530818,0.0023927018,0.00005084617,0.00001561787],"category_scores_gemma":[0.000141625,0.00012951854,0.00006365255,0.00024191002,0.000025096337,0.0027375775,0.00068934896,0.0002581194,0.000075634314],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000047131603,0.0007042678,0.9714119,0.000017207189,0.0000143825355,0.00006787169,0.0019021727,0.00013462327,0.00009144288,0.000082641636,0.00015160219,0.025417184],"study_design_scores_gemma":[0.0031674935,0.0022090082,0.88872874,0.00060845626,0.000019679994,0.0000048220313,0.0124329515,0.06708551,0.019650592,0.0013033883,0.003095905,0.0016934379],"about_ca_topic_score_codex":0.00006827283,"about_ca_topic_score_gemma":0.00006220332,"teacher_disagreement_score":0.08268314,"about_ca_system_score_codex":0.000067697176,"about_ca_system_score_gemma":0.00006894205,"threshold_uncertainty_score":0.9996836},"labels":[],"label_agreement":null},{"id":"W2610973354","doi":"10.3166/jesa.49.449-469","title":"A Kaufmann reliability network approach based on components off-the-shelf to design safety instrumented systems","year":2016,"lang":"en","type":"article","venue":"Journal Européen des Systèmes Automatisés","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Reliability engineering; Reliability (semiconductor); Computer science; Off the shelf; Engineering; Software engineering; Physics","score_opus":0.033719570192338046,"score_gpt":0.26028200872234875,"score_spread":0.2265624385300107,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2610973354","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.021099754,0.00026832655,0.9730617,0.0030492002,0.0005904423,0.00094512076,0.000011129637,0.00040608548,0.00056828116],"genre_scores_gemma":[0.9139308,0.00009833987,0.08446052,0.0005412884,0.00033344704,0.000082047525,0.0000028704128,0.00004827105,0.000502445],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99134624,0.0036134939,0.0012840269,0.00081951515,0.0018855408,0.0010511756],"domain_scores_gemma":[0.9943649,0.0021043296,0.00046500485,0.0017095882,0.0007325739,0.0006235916],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0071085272,0.00042647377,0.0006949068,0.0002711411,0.0011705887,0.0010132109,0.002691643,0.0001377259,0.00002965614],"category_scores_gemma":[0.0018894281,0.00023291424,0.0003643499,0.0016980055,0.0002360937,0.0006241728,0.0004753282,0.00054188364,0.00031828714],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034664688,0.00071100186,0.0078061535,0.00024090511,0.00028578588,0.00010580415,0.0004696685,0.38056728,0.0004426975,0.0026364247,0.010401686,0.59598595],"study_design_scores_gemma":[0.000980732,0.00057583133,0.11557197,0.0007983332,0.00003364301,0.00021564325,0.00002652766,0.8738139,0.00005274026,0.0016554396,0.00587999,0.00039524375],"about_ca_topic_score_codex":0.000046428904,"about_ca_topic_score_gemma":0.000001581565,"teacher_disagreement_score":0.892831,"about_ca_system_score_codex":0.00089807704,"about_ca_system_score_gemma":0.00033236644,"threshold_uncertainty_score":0.9770416},"labels":[],"label_agreement":null},{"id":"W2616886411","doi":"","title":"The extended operational profile model for usage-based software testing","year":2004,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Computer science; Software quality; Software development; Verification and validation; Software; Software reliability testing; Software construction; Software sizing; Software engineering; Software development process; Reliability engineering; Engineering; Programming language","score_opus":0.05463305285816743,"score_gpt":0.3101263040904218,"score_spread":0.25549325123225436,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2616886411","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014870464,0.000031611675,0.9945844,0.0031605794,0.000027708915,0.00030195265,0.0000043889854,0.00020180653,0.00020050933],"genre_scores_gemma":[0.23972902,0.0000010565378,0.75874066,0.00040716413,0.000029912517,0.00017317961,0.0000055236724,0.000005458102,0.00090802176],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99877155,0.000032661614,0.00018502546,0.00032363023,0.00039147492,0.00029566977],"domain_scores_gemma":[0.9978996,0.0011409976,0.000029754958,0.00044450013,0.0004111426,0.000074010495],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007303337,0.00008680146,0.00008786887,0.000043455555,0.0011645013,0.0003054468,0.0007835211,0.000036998117,0.000008434496],"category_scores_gemma":[0.001744755,0.00005402176,0.00008889707,0.00039748303,0.000068632165,0.00025933533,0.00017194654,0.000095113755,0.00001990622],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014235884,0.00020947312,0.00070424535,0.000031703752,0.00002560447,0.000002009984,0.00019309184,0.79207057,0.00049056654,0.099802926,0.0005984886,0.10585711],"study_design_scores_gemma":[0.0003053811,0.000043113527,0.0001531942,0.0000073398355,0.0000022518686,9.955802e-7,0.000004975973,0.9661093,0.0013597033,0.031673115,0.0002585449,0.00008208746],"about_ca_topic_score_codex":0.000029306191,"about_ca_topic_score_gemma":0.000020559846,"teacher_disagreement_score":0.23824197,"about_ca_system_score_codex":0.00006937371,"about_ca_system_score_gemma":0.00065351947,"threshold_uncertainty_score":0.89565206},"labels":[],"label_agreement":null},{"id":"W2644985284","doi":"10.1002/qre.2180","title":"A new reliability analysis method for repairable systems with multifunction modes based on goal‐oriented methodology","year":2017,"lang":"en","type":"article","venue":"Quality and Reliability Engineering International","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Ministry of Industry and Information Technology of the People's Republic of China; National Natural Science Foundation of China","keywords":"Fault tree analysis; Reliability (semiconductor); Computer science; Reliability engineering; Operator (biology); Monte Carlo method; Process (computing); Bitwise operation; Complex system; Function (biology); Engineering; Mathematics; Artificial intelligence","score_opus":0.056029965069334005,"score_gpt":0.3752264359369813,"score_spread":0.3191964708676473,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2644985284","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.01256132,0.000024854326,0.9826617,0.0034024937,0.0004693256,0.00045375363,0.000040101153,0.0002549872,0.00013142212],"genre_scores_gemma":[0.36708122,0.000008783243,0.6322636,0.000060910035,0.00010575913,0.00010876048,0.000033104596,0.000012182909,0.0003256692],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9965202,0.0005187909,0.00065437454,0.001187405,0.00074156304,0.00037766175],"domain_scores_gemma":[0.9929877,0.003811286,0.00030715577,0.0019526964,0.0006985328,0.00024261046],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0075810966,0.00028030225,0.0006447196,0.00037817188,0.00043476906,0.00048430022,0.0010095959,0.00020807306,0.000016640864],"category_scores_gemma":[0.0081011215,0.00022616176,0.0003678782,0.00040662318,0.000098482684,0.00055084744,0.00020237481,0.00034476523,0.0000025926117],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032842078,0.00018894068,0.018073125,0.00021375109,0.00042381935,0.00000163053,0.00011222107,0.9514777,0.00014186697,0.024722204,0.000062734514,0.0042535802],"study_design_scores_gemma":[0.0008095186,0.0001887269,0.04491984,0.0000439867,0.000116782656,0.000002010249,0.000023378907,0.9501835,0.0002982539,0.0010336711,0.0021228231,0.0002574716],"about_ca_topic_score_codex":0.0029712263,"about_ca_topic_score_gemma":0.00003973414,"teacher_disagreement_score":0.3545199,"about_ca_system_score_codex":0.00024333941,"about_ca_system_score_gemma":0.00015799608,"threshold_uncertainty_score":0.96983844},"labels":[],"label_agreement":null},{"id":"W2759524920","doi":"10.1109/rew.2017.15","title":"A Distance-Based GRL Approach to Goal Model Refinement and Alternative Selection","year":2017,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Computer science; Selection (genetic algorithm); Goal orientation; Artificial intelligence","score_opus":0.03987311456047987,"score_gpt":0.31053091134069377,"score_spread":0.2706577967802139,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2759524920","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015697248,0.000008459697,0.9759399,0.0017480183,0.000018039598,0.00012630499,9.884436e-7,0.000055664765,0.00640532],"genre_scores_gemma":[0.8485415,0.000004507369,0.15014967,0.00019528987,0.00002009456,0.000035383768,7.9812844e-7,0.0000027053723,0.0010500603],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.998965,0.000031343443,0.00010258752,0.0003993724,0.00031654214,0.00018510387],"domain_scores_gemma":[0.99918276,0.000034778586,0.000038608305,0.00051778345,0.00010991754,0.000116138835],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003394997,0.000077387784,0.000107188964,0.00007044526,0.0003642148,0.00043864275,0.00062546186,0.000024524683,0.000004044934],"category_scores_gemma":[0.00011288622,0.000060934803,0.000037189122,0.00011782033,0.000050773408,0.00026061872,0.00024820975,0.00008015451,0.0000090135145],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014805183,0.0014157817,0.052998126,0.00016975167,0.00018279708,0.0000066718585,0.0018614922,0.25454694,0.0014778638,0.42601424,0.005235448,0.25594285],"study_design_scores_gemma":[0.00019234317,0.000044239354,0.0023350678,0.00000552513,0.0000024768794,4.467834e-7,0.0000057630514,0.99324656,0.00089935894,0.0026012077,0.0005843468,0.000082655104],"about_ca_topic_score_codex":0.00038611534,"about_ca_topic_score_gemma":0.00006820049,"teacher_disagreement_score":0.83284426,"about_ca_system_score_codex":0.00006842214,"about_ca_system_score_gemma":0.00006291135,"threshold_uncertainty_score":0.4229842},"labels":[],"label_agreement":null},{"id":"W2767694364","doi":"10.1109/ase.2017.8115702","title":"TREM: A tool for mining timed regular specifications from system traces","year":2017,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Software engineering; Embedded system; Programming language","score_opus":0.08879109471473576,"score_gpt":0.3107424937200098,"score_spread":0.22195139900527402,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2767694364","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06309629,0.00008297536,0.9302206,0.0030003504,0.000109585504,0.00025169866,0.000011324395,0.00021043106,0.003016747],"genre_scores_gemma":[0.81815815,0.000006813612,0.17826454,0.000018254175,0.00012343984,0.000070349626,0.0000062672093,0.000005205655,0.0033469547],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9988857,0.000045406632,0.00020295291,0.00040032543,0.00026149274,0.00020415816],"domain_scores_gemma":[0.99764663,0.00048107284,0.00009379507,0.0015731188,0.00013554901,0.00006983546],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00049817015,0.0000787314,0.00016925742,0.000059565198,0.00064106943,0.00079833396,0.0015442561,0.00005128092,0.000035874706],"category_scores_gemma":[0.00040904866,0.00006396003,0.00012876042,0.00009260722,0.00006175077,0.000451451,0.00014889528,0.000048378366,0.00007876692],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000051931587,0.00044077297,0.020711262,0.00022934025,0.00048057703,0.000026819122,0.0028181614,0.00018178465,0.0074919043,0.2614155,0.022912206,0.68323976],"study_design_scores_gemma":[0.0010242406,0.000086093205,0.04343843,0.00012390266,0.000054335138,0.0000047290373,0.00070154486,0.90290684,0.011402017,0.0056290706,0.03414469,0.00048413844],"about_ca_topic_score_codex":0.00014448162,"about_ca_topic_score_gemma":0.000032504027,"teacher_disagreement_score":0.90272504,"about_ca_system_score_codex":0.000043116546,"about_ca_system_score_gemma":0.000058410897,"threshold_uncertainty_score":0.7698353},"labels":[],"label_agreement":null},{"id":"W2775948036","doi":"10.1002/dac.3480","title":"Analysis of instantaneous availability of communication system based on the influence of support equipment","year":2017,"lang":"en","type":"article","venue":"International Journal of Communication Systems","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"National Natural Science Foundation of China","keywords":"Computer science; Queueing theory; Exponential distribution; Exponential function; Markov chain; Communications system; Transmission (telecommunications); Function (biology); Service (business); Process (computing); Reliability engineering; Real-time computing; Distributed computing; Telecommunications; Computer network; Machine learning","score_opus":0.03731038697784972,"score_gpt":0.33542971888903594,"score_spread":0.2981193319111862,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2775948036","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96781486,0.00035012772,0.027688991,0.0020936383,0.00013332239,0.0002475729,0.000040471714,0.000012307589,0.0016187058],"genre_scores_gemma":[0.99873054,0.00022055006,0.0009844932,0.000021419706,0.000009060267,0.000008105185,0.0000068611444,0.000004086407,0.0000148640365],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99539924,0.0010441259,0.0016581755,0.00013516944,0.0016628212,0.000100451056],"domain_scores_gemma":[0.9861024,0.002107534,0.0040110117,0.0035474973,0.0041769785,0.00005456794],"candidate_categories":["open_science"],"consensus_categories":[],"category_scores_codex":[0.005907154,0.00010607203,0.0005435089,0.00056882104,0.0001688731,0.00015909939,0.008181123,0.000064085114,0.000014245998],"category_scores_gemma":[0.00135042,0.000076174096,0.00036212528,0.0004445539,0.00039372305,0.00042734278,0.0005924685,0.00024830585,0.0000025264064],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0003709533,0.0012606407,0.21336196,0.00017839938,0.004449501,0.000010313804,0.0026070606,0.6771285,0.0029879466,0.09201719,0.00012216697,0.0055053886],"study_design_scores_gemma":[0.0010103646,0.000405485,0.22930111,0.0018918291,0.0004189579,0.00003174995,0.0011143694,0.7602311,0.0039966353,0.00042556427,0.0009554828,0.00021734863],"about_ca_topic_score_codex":0.0006180122,"about_ca_topic_score_gemma":0.000033317505,"teacher_disagreement_score":0.09159163,"about_ca_system_score_codex":0.00022732653,"about_ca_system_score_gemma":0.00028726883,"threshold_uncertainty_score":0.9971851},"labels":[],"label_agreement":null},{"id":"W2780467792","doi":"10.2495/safe-v8-n1-139-149","title":"Safety status: An innovative concept for maintaining the safety integrity level of operational safety systems","year":2018,"lang":"en","type":"article","venue":"International Journal of Safety and Security Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Risk analysis (engineering); System safety; Occupational safety and health; Reliability engineering; Engineering; Computer science; Business; Medicine","score_opus":0.03764960121178722,"score_gpt":0.3171622025678596,"score_spread":0.27951260135607237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2780467792","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.018074006,0.0003050285,0.9780832,0.0020491513,0.0008465619,0.00020318365,0.00031867958,0.000022373682,0.00009780137],"genre_scores_gemma":[0.9861147,0.00033016622,0.012778459,0.000103536906,0.00060817914,0.0000035187234,0.00003333854,0.000010483755,0.000017641065],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9973082,0.0001650158,0.0011223938,0.0002605513,0.0008499142,0.00029392884],"domain_scores_gemma":[0.9946288,0.0011913613,0.00038616848,0.0002449205,0.0034034292,0.00014531241],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0030639505,0.000185293,0.00040095914,0.00025497473,0.00023895972,0.00019578967,0.0011147851,0.00010560354,0.000023403276],"category_scores_gemma":[0.0012617212,0.00013889452,0.0001420317,0.00045744993,0.00023221447,0.001020076,0.00026121055,0.0005629204,8.2882883e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0023310278,0.00032884564,0.002667231,0.00009901654,0.0015877631,0.000022802993,0.025490787,0.11681992,0.0011840988,0.8118136,0.00015719543,0.037497714],"study_design_scores_gemma":[0.002377246,0.000683144,0.013634321,0.00042465905,0.000030974737,0.00019812245,0.001880642,0.93712777,0.0012046463,0.002377921,0.03969568,0.0003648467],"about_ca_topic_score_codex":0.000117966694,"about_ca_topic_score_gemma":0.00002534709,"teacher_disagreement_score":0.96804065,"about_ca_system_score_codex":0.0002670485,"about_ca_system_score_gemma":0.00036696278,"threshold_uncertainty_score":0.5663954},"labels":[],"label_agreement":null},{"id":"W2783654832","doi":"10.1109/iccons.2017.8250774","title":"Reliability assessment of component based software by using basis path testing","year":2017,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Horizon College and Seminary","funders":"","keywords":"Cyclomatic complexity; Computer science; Software construction; Software development; Reliability engineering; Software sizing; Software reliability testing; Software quality; Software engineering; Component (thermodynamics); Software metric; Verification and validation; Component-based software engineering; Software development process; Reliability (semiconductor); Package development process; Software measurement; Software; Engineering; Programming language","score_opus":0.06307970262978242,"score_gpt":0.3515389638541508,"score_spread":0.2884592612243684,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2783654832","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2279485,0.000015187904,0.7709763,0.00036894612,0.000055322125,0.00013442576,0.000009530302,0.00010883049,0.00038292413],"genre_scores_gemma":[0.61688596,0.0000014836472,0.38303003,0.000040595794,0.000009037297,0.0000048127586,0.0000021623136,0.000004630415,0.00002130465],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976628,0.000198237,0.00041951673,0.0005750465,0.0008020211,0.00034238695],"domain_scores_gemma":[0.9957638,0.0010123454,0.0002824665,0.002204294,0.00058487395,0.00015223492],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0017813613,0.00015178548,0.00033661947,0.000081844555,0.00061242323,0.0003405258,0.001770838,0.000070749775,0.00005363192],"category_scores_gemma":[0.0028676991,0.00012604817,0.00015947624,0.000260303,0.00020893267,0.0005174597,0.0006964677,0.0001943163,0.0000039048],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000043769837,0.0006371172,0.9193125,0.00018397119,0.000036263344,0.0000074647096,0.00003054312,0.0047351797,0.012179972,0.0004183344,0.00036356508,0.062090714],"study_design_scores_gemma":[0.00024370357,0.000060077986,0.12983328,0.00005569412,0.000010515933,8.8005453e-7,0.000006663534,0.86543274,0.0035942877,0.0004915493,0.000118894815,0.00015171921],"about_ca_topic_score_codex":0.0017066606,"about_ca_topic_score_gemma":0.000007750372,"teacher_disagreement_score":0.86069757,"about_ca_system_score_codex":0.00014581264,"about_ca_system_score_gemma":0.00034256617,"threshold_uncertainty_score":0.5140095},"labels":[],"label_agreement":null},{"id":"W2788598677","doi":"","title":"Mapping features to source code in dynamically configured avionics software","year":2012,"lang":"en","type":"article","venue":"PolyPublie (École Polytechnique de Montréal)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Natural Sciences and Engineering Research Council of Canada; Consortium de Recherche et d’innovation en Aérospatiale au Québec","keywords":"Computer science; Program comprehension; Static program analysis; Avionics; Software; Feature (linguistics); Source code; Software system; Software engineering; Avionics software; Code (set theory); Reverse engineering; Feature model; Software construction; Software development; Programming language; Engineering; Set (abstract data type)","score_opus":0.012271808558411597,"score_gpt":0.25550396404195547,"score_spread":0.24323215548354388,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2788598677","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09871822,0.00093249383,0.8929473,0.005717675,0.000095628646,0.0005703476,0.00001382234,0.00088283257,0.00012166689],"genre_scores_gemma":[0.7205695,0.00006192101,0.27576014,0.0024550473,0.000085363325,0.0002938459,0.000013855324,0.0000374184,0.000722935],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9962211,0.0003213906,0.00056547904,0.00067040755,0.0007624226,0.0014591954],"domain_scores_gemma":[0.9970411,0.00045460474,0.00012702853,0.0014411648,0.00018345013,0.0007526536],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018483956,0.00034700186,0.00047759624,0.0009531382,0.00025130928,0.00039592135,0.0018418147,0.0003451212,0.000024688636],"category_scores_gemma":[0.0013613676,0.00034575668,0.0002250023,0.0022103477,0.00008016353,0.00069879985,0.0008321119,0.00079606066,0.00007552662],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001037473,0.0013730199,0.5906146,0.00014712862,0.00013156443,0.00010061539,0.005669243,0.020648774,0.0094969375,0.05165038,0.0046166233,0.31544733],"study_design_scores_gemma":[0.0009049293,0.00018801018,0.62540764,0.0002375547,0.000019877509,0.0001806651,0.00035657294,0.34178957,0.0053274455,0.0043434612,0.0198344,0.0014098776],"about_ca_topic_score_codex":0.004119301,"about_ca_topic_score_gemma":0.0027850196,"teacher_disagreement_score":0.62185127,"about_ca_system_score_codex":0.0008159903,"about_ca_system_score_gemma":0.00028910197,"threshold_uncertainty_score":0.99989945},"labels":[],"label_agreement":null},{"id":"W2802046167","doi":"10.1007/s11334-018-0310-0","title":"Introduction to special issue on verification and evaluation of computer systems","year":2018,"lang":"en","type":"article","venue":"Innovations in Systems and Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Computer science; Computer Applications; Political science","score_opus":0.02269678944781473,"score_gpt":0.28615576391732606,"score_spread":0.26345897446951133,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2802046167","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.32556313,0.000093667724,0.6719282,0.0003205562,0.0016181326,0.00037748474,0.0000027095136,0.00006403188,0.0000320847],"genre_scores_gemma":[0.9830769,0.0000074183035,0.010944036,0.000012718194,0.0058568935,0.00006498594,0.000005579176,0.0000071700915,0.00002430779],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9987634,0.00006810937,0.00035204925,0.000302492,0.00039024866,0.00012374733],"domain_scores_gemma":[0.9988253,0.00012069607,0.000059432645,0.00031540781,0.00064366154,0.000035472436],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015698317,0.00008622801,0.00016483267,0.0005404182,0.00006310536,0.0001249722,0.00013138869,0.000057192097,0.0000046066425],"category_scores_gemma":[0.00050259795,0.00008594923,0.000010963209,0.0012759529,0.000027254522,0.00021278847,0.000058362715,0.00008659371,0.000008167104],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022163687,0.00017779219,0.0144589245,0.0006112969,0.00009572657,0.0000017651843,0.004363895,0.54943633,0.001553477,0.116689295,0.010648391,0.30194095],"study_design_scores_gemma":[0.0002902781,0.00017277586,0.033510733,0.00017172645,0.0000073444917,0.000009263086,0.00006762769,0.9322827,0.00022335199,0.00004191448,0.033045586,0.00017673534],"about_ca_topic_score_codex":0.000072331204,"about_ca_topic_score_gemma":0.0000032815217,"teacher_disagreement_score":0.66098416,"about_ca_system_score_codex":0.00009304309,"about_ca_system_score_gemma":0.000031835407,"threshold_uncertainty_score":0.35049078},"labels":[],"label_agreement":null},{"id":"W2805428613","doi":"10.1016/j.atherosclerosissup.2018.04.245","title":"Identifying the Genetic Basis of Vascular Cognitive Impairment Using a Custom Designed Next-generation Sequencing-based Gene Panel","year":2018,"lang":"en","type":"article","venue":"Atherosclerosis Supplements","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Western Hospital; London Health Sciences Centre; Hamilton Health Sciences; Hamilton General Hospital; Occupational Cancer Research Centre; Health Sciences Centre; Ottawa Hospital; University of Ottawa; Thunder Bay Regional Health Sciences Centre; Western University; University of Toronto; NOSM University; York University; Ontario Stroke Network; McMaster University; Robarts Clinical Trials; St. Michael's Hospital; Sunnybrook Health Science Centre","funders":"","keywords":"Exploit; Computer science; Vulnerability (computing); Software; Computer security; Threat model; Software security assurance; Field (mathematics); Risk analysis (engineering); Information security; Operating system","score_opus":0.19635004812176762,"score_gpt":0.3254698799899092,"score_spread":0.1291198318681416,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2805428613","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5320163,0.00006896935,0.46736258,0.000066327535,0.0000722316,0.00036315693,0.00001669031,0.000031707463,0.0000020867947],"genre_scores_gemma":[0.9451227,0.000055124245,0.054336518,0.00024121703,0.00011126744,0.00008192812,0.000022915716,0.000021743437,0.000006582813],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.99665314,0.0005690402,0.0005722107,0.00063778844,0.0010324001,0.0005353988],"domain_scores_gemma":[0.99804604,0.00021850657,0.00024081045,0.00084223214,0.0005340869,0.000118300166],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012459814,0.00023998167,0.0002900927,0.00018335159,0.00059873576,0.00033489763,0.0009500503,0.000079800426,0.00021515919],"category_scores_gemma":[0.00011809247,0.00018874937,0.00026534783,0.0009711634,0.00024157463,0.00038825706,0.0003179452,0.000118087286,0.000036942434],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019354185,0.0003622756,0.04008321,0.0000576854,0.00039465073,0.0000024418403,0.0027959037,0.0017302172,0.94490105,0.000050687726,0.00012443648,0.009478094],"study_design_scores_gemma":[0.0010795743,0.00031630477,0.016473785,0.000106338746,0.00014548078,0.0000016539585,0.00029565595,0.42902616,0.55214155,0.00012031355,0.000030326892,0.0002628837],"about_ca_topic_score_codex":0.0010176275,"about_ca_topic_score_gemma":0.000096905635,"teacher_disagreement_score":0.42729592,"about_ca_system_score_codex":0.00026919012,"about_ca_system_score_gemma":0.00031280713,"threshold_uncertainty_score":0.7696976},"labels":[],"label_agreement":null},{"id":"W2807804912","doi":"10.1287/mnsc.2017.3004","title":"Process Flexibility in Baseball: The Value of Positional Flexibility","year":2018,"lang":"en","type":"article","venue":"Management Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":18,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Flexibility (engineering); Computer science; Generalizability theory; Process (computing); Value (mathematics); Operations research; Operations management; Industrial engineering; Economics; Engineering; Mathematics; Statistics; Machine learning","score_opus":0.025949470535524443,"score_gpt":0.3425185213716851,"score_spread":0.31656905083616066,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2807804912","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.81505656,0.000031799398,0.15967242,0.0038566494,0.00017357287,0.0006221109,0.0000028627314,0.00013685686,0.020447172],"genre_scores_gemma":[0.99186015,0.0000040541863,0.007695838,0.00024191888,0.00001996355,0.000023805187,5.0932175e-7,0.0000019561712,0.00015178943],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.9969757,0.00014568867,0.00032940914,0.0007524425,0.0013923576,0.00040438992],"domain_scores_gemma":[0.9981339,0.00015532729,0.00007573475,0.0012645804,0.0002937106,0.00007672846],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006106271,0.00009830151,0.00013532318,0.00026404244,0.00033284834,0.00014120528,0.0031546396,0.000022241336,0.0000722899],"category_scores_gemma":[0.000319342,0.00006779592,0.00006261195,0.00407577,0.0019812072,0.0007000687,0.0009793368,0.000119343305,0.000037165762],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000079776655,0.0017005412,0.15252574,0.00042152478,0.000049170132,0.00001808204,0.0029599036,0.00804852,0.0012693556,0.7530495,0.0005044535,0.07937344],"study_design_scores_gemma":[0.00023394705,0.00011732939,0.49613705,0.00004328229,0.000006222681,0.0000016344006,0.00017274605,0.4246722,0.005112013,0.07308772,0.00025715533,0.00015870841],"about_ca_topic_score_codex":0.000120607336,"about_ca_topic_score_gemma":0.000033816483,"teacher_disagreement_score":0.6799618,"about_ca_system_score_codex":0.00012708851,"about_ca_system_score_gemma":0.00015168612,"threshold_uncertainty_score":0.72998405},"labels":[],"label_agreement":null},{"id":"W2808848181","doi":"10.1109/compsac.2018.10345","title":"Traceability in Systems Engineering: An Avionics Case Study","year":2018,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"University of Nottingham","keywords":"Traceability; Requirements traceability; TRACE (psycholinguistics); Computer science; Avionics; Software engineering; Semantics (computer science); Systems engineering; Requirements engineering; Programming language; Engineering; Software","score_opus":0.036023616891966745,"score_gpt":0.3288733087053243,"score_spread":0.2928496918133575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2808848181","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.89672273,0.000025064604,0.10251903,0.000050426657,0.00010635951,0.00026491773,2.9242563e-7,0.00024062848,0.00007052753],"genre_scores_gemma":[0.99650294,8.115613e-7,0.0033581057,0.000008060296,0.000044851182,0.000023011053,2.2513032e-7,0.000003883433,0.000058095902],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9985406,0.0002000799,0.00025986595,0.00044081215,0.00030514688,0.00025352795],"domain_scores_gemma":[0.9985944,0.00016770684,0.000018275043,0.0009386193,0.00015136538,0.00012965256],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016997667,0.000089581445,0.00016319288,0.00015038151,0.000074142365,0.00021458966,0.00056068593,0.000045661043,0.000018991339],"category_scores_gemma":[0.0001807686,0.000074039104,0.00003563707,0.0008541706,0.0000366653,0.00042981812,0.00015951353,0.00014849365,0.00003563359],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002496023,0.0112652695,0.8633582,0.00017202795,0.00012270907,0.00727353,0.025463851,0.019802386,0.00020076169,0.015088892,0.0001965486,0.05703085],"study_design_scores_gemma":[0.00020653609,0.00039548057,0.0094845025,0.0000037722739,0.0000025818204,0.00019987731,0.0014113577,0.987964,0.000030052423,0.00006135935,0.00012514058,0.00011537622],"about_ca_topic_score_codex":0.0031951545,"about_ca_topic_score_gemma":0.0013079259,"teacher_disagreement_score":0.9681616,"about_ca_system_score_codex":0.00007918058,"about_ca_system_score_gemma":0.00006175877,"threshold_uncertainty_score":0.4830139},"labels":[],"label_agreement":null},{"id":"W2887326578","doi":"","title":"Perspectives on Probabilistic Assessment of Systems and Software","year":2016,"lang":"en","type":"preprint","venue":"HAL (Le Centre pour la Communication Scientifique Directe)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Safran Electronics (Canada)","funders":"","keywords":"Probabilistic logic; Computer science; Software construction; Avionics software; Social software engineering; Software; Software development; Software system; Software metric; Relevance (law); Software engineering; Software quality; Artificial intelligence; Programming language","score_opus":0.01898831506008601,"score_gpt":0.27807434263464104,"score_spread":0.259086027574555,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2887326578","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.037990384,0.0016320406,0.94674927,0.005329568,0.00012592346,0.00057299284,0.00005002848,0.00023104262,0.0073187603],"genre_scores_gemma":[0.9269071,0.00069891027,0.070706956,0.000011310886,0.000016166301,0.0000995037,0.000024413532,0.000019371397,0.0015162639],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99247754,0.004923801,0.00049261807,0.0010500485,0.0007596155,0.00029638867],"domain_scores_gemma":[0.9895939,0.004102946,0.00042823332,0.002869184,0.0028337552,0.00017193313],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006015012,0.0002664379,0.00047973404,0.0002784592,0.00021023474,0.00048268406,0.0018264123,0.00020884618,0.000019797906],"category_scores_gemma":[0.0037656117,0.00022084724,0.00018517114,0.00036247892,0.00038921947,0.0001491462,0.0022411009,0.00050182204,0.0000075281137],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000006603951,0.0010290525,0.007195464,0.0009858392,0.00020120856,0.0000052601968,0.008723666,0.00039625127,0.00043958632,0.9339176,0.00022471226,0.046874776],"study_design_scores_gemma":[0.0024086675,0.000013398181,0.051771335,0.023540497,0.0002069942,0.000036889764,0.0014465479,0.73449194,0.008427606,0.17025425,0.0047890297,0.0026128318],"about_ca_topic_score_codex":0.00037985545,"about_ca_topic_score_gemma":0.00005520232,"teacher_disagreement_score":0.88891673,"about_ca_system_score_codex":0.0002062097,"about_ca_system_score_gemma":0.0004960603,"threshold_uncertainty_score":0.9005889},"labels":[],"label_agreement":null},{"id":"W2889305181","doi":"","title":"Design and Analysis of Control Strategies for a Cyber Physical System","year":2017,"lang":"en","type":"article","venue":"Computer Systems: Science & Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal; École de Technologie Supérieure; Concordia University","funders":"","keywords":"Reliability (semiconductor); Computer science; Reliability engineering; Cyber-physical system; Abstraction; Process (computing); Control system; Industrial control system; Complex system; Distributed computing; Control (management); Engineering; Artificial intelligence","score_opus":0.02217073667263983,"score_gpt":0.277944441651588,"score_spread":0.25577370497894814,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2889305181","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07189601,0.00010119689,0.92713594,0.000028638407,0.0003651449,0.000341184,0.0000038712788,0.00011385027,0.000014148204],"genre_scores_gemma":[0.9491384,0.0000015560607,0.05068204,0.0000029490486,0.0001129799,0.000052326177,3.61459e-7,0.0000056864565,0.0000037115715],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99812186,0.000052271607,0.00030672748,0.00057325925,0.00053619425,0.0004096702],"domain_scores_gemma":[0.997679,0.0005105358,0.00018103165,0.0010994275,0.0003718826,0.00015814995],"candidate_categories":["scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0018637269,0.00016199186,0.0005814511,0.00058811886,0.00042564323,0.0017178424,0.0019631775,0.000040730025,1.3496825e-7],"category_scores_gemma":[0.00010543781,0.00013642084,0.00016118334,0.00091335725,0.0002477799,0.0012335728,0.00032888172,0.00008047312,0.0000015432175],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000037006098,0.000026449403,0.0011363071,0.00018303556,0.00028430505,0.000004011208,0.0005913316,0.892023,0.0029790595,0.10100625,0.000012648297,0.001749872],"study_design_scores_gemma":[0.0002167131,0.000052177744,0.004573733,0.00006938386,0.00008604283,0.000004650585,0.000050121762,0.9945077,0.00024337828,0.0000253377,0.000029901074,0.00014091129],"about_ca_topic_score_codex":0.000061728664,"about_ca_topic_score_gemma":7.3952754e-7,"teacher_disagreement_score":0.8772424,"about_ca_system_score_codex":0.0000811783,"about_ca_system_score_gemma":0.00017334284,"threshold_uncertainty_score":0.9993185},"labels":[],"label_agreement":null},{"id":"W2900428710","doi":"10.1016/j.ress.2018.11.004","title":"A comparison of fault trees and the Dynamic Flowgraph Methodology for the analysis of FPGA-based safety systems part 2: Theoretical investigations","year":2018,"lang":"en","type":"article","venue":"Reliability Engineering & System Safety","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":12,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ontario Institute of Technology; Canadian Nuclear Safety Commission","funders":"","keywords":"Fault tree analysis; Design for manufacturability; Reliability engineering; Computer science; Binary decision diagram; Field-programmable gate array; Implicant; Reliability (semiconductor); Consistency (knowledge bases); Boolean function; Algorithm; Embedded system; Engineering","score_opus":0.028783955137388524,"score_gpt":0.31855396106694017,"score_spread":0.28977000592955166,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2900428710","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.032430265,0.0006879738,0.96431816,0.0011491764,0.00025721054,0.00093227095,0.00008687051,0.000116101684,0.000021980797],"genre_scores_gemma":[0.9800248,0.000029489798,0.01972872,0.00001049914,0.00003788594,0.00013833298,0.000013156838,0.000011750522,0.0000053447407],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99578846,0.0013685366,0.0013765123,0.0005516211,0.00055455795,0.00036031904],"domain_scores_gemma":[0.9768971,0.020288736,0.00032096327,0.0016883888,0.0006874047,0.000117392316],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01038265,0.00023552313,0.0012624625,0.00031767707,0.00029786475,0.00006369135,0.0010736777,0.00015959144,0.0000049536425],"category_scores_gemma":[0.004149821,0.00013558268,0.0005861507,0.0023924308,0.0017067321,0.00008772936,0.00015574637,0.00023657162,9.1390166e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002073865,0.00007939902,0.004234884,0.0008408096,0.0011978504,1.6061016e-7,0.0010495499,0.6943967,0.00023624489,0.2960288,0.000012652795,0.001715606],"study_design_scores_gemma":[0.0005513462,0.00008919796,0.009107971,0.00012284271,0.000751265,0.0000014916924,0.00028589534,0.9878976,0.00039828196,0.00035487732,0.00031017198,0.00012906278],"about_ca_topic_score_codex":0.00026238707,"about_ca_topic_score_gemma":0.00007065822,"teacher_disagreement_score":0.9475946,"about_ca_system_score_codex":0.00012837573,"about_ca_system_score_gemma":0.0001325166,"threshold_uncertainty_score":0.62885255},"labels":[],"label_agreement":null},{"id":"W2912010915","doi":"","title":"Theory and Applications of Satisfiability Testing: 7th International Conference, SAT 2004, Vancouver, BC, Canada, May 10-13, 2004, Revised Selected Papers (Lecture Notes in Computer Science)","year":2005,"lang":"en","type":"article","venue":"Theory and Applications of Satisfiability Testing","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Satisfiability; Computer science; Mathematics education; Theoretical computer science; Library science; Mathematics","score_opus":0.013809007036564742,"score_gpt":0.265525670005974,"score_spread":0.25171666296940926,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2912010915","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.22192778,0.00042088173,0.7746386,0.00044123133,0.00003180078,0.0013732406,0.00009993062,0.00013150151,0.0009350285],"genre_scores_gemma":[0.9345054,0.00002062845,0.06508345,0.00007159125,0.000073313175,0.00018916011,0.000015335754,0.000011198457,0.000029924151],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99622315,0.000625267,0.0009590121,0.0010516075,0.00068241294,0.00045856123],"domain_scores_gemma":[0.98755777,0.009204621,0.00041534496,0.0011039006,0.0014935688,0.00022479199],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.006230524,0.0002921986,0.000518175,0.0003336956,0.00043359742,0.00011837754,0.0011373359,0.000117523094,0.00011638523],"category_scores_gemma":[0.0059459493,0.00027557416,0.000055606037,0.0032215524,0.0016841722,0.0004271364,0.0004823722,0.00036334505,0.0000019599397],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000041469528,0.00029201218,0.16932334,0.0002519099,0.000037377547,3.8570852e-7,0.0003644046,0.0057649086,0.0030301218,0.026411071,0.000021690335,0.7944613],"study_design_scores_gemma":[0.0017887284,0.00027692493,0.35580942,0.00044031735,0.00011994576,0.000027008862,0.0003726513,0.20945194,0.0061374903,0.42256275,0.0017176161,0.0012952201],"about_ca_topic_score_codex":0.011120538,"about_ca_topic_score_gemma":0.021895345,"teacher_disagreement_score":0.7931661,"about_ca_system_score_codex":0.000260376,"about_ca_system_score_gemma":0.0014261623,"threshold_uncertainty_score":0.99996966},"labels":[],"label_agreement":null},{"id":"W2921598390","doi":"","title":"An Introduction to Fast Time Simulations for RPAS Collision Avoidance System Evaluation","year":2018,"lang":"en","type":"preprint","venue":"HAL (Le Centre pour la Communication Scientifique Directe)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Safran Electronics (Canada)","funders":"","keywords":"Collision avoidance; Computer science; Collision; Collision avoidance system; Computer security","score_opus":0.020463929259943486,"score_gpt":0.2941480854491307,"score_spread":0.2736841561891872,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2921598390","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.066352636,0.00012300207,0.91599995,0.014306431,0.00035940815,0.0016338781,0.000093602495,0.00044099428,0.00069011364],"genre_scores_gemma":[0.76105267,0.000020390518,0.23548132,0.000048277092,0.00027637556,0.00037703203,0.0009829004,0.000036998114,0.0017240286],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9894873,0.006769851,0.0006607082,0.001557548,0.0011120816,0.0004125439],"domain_scores_gemma":[0.9833893,0.0017078674,0.0004209935,0.0041819722,0.0100095,0.00029036595],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0152610345,0.00030779358,0.0004148713,0.00045319812,0.0007812449,0.0011559412,0.0024955028,0.000293675,0.00006430289],"category_scores_gemma":[0.0047110827,0.00033015865,0.0002216399,0.0010798833,0.00013495887,0.0004816988,0.0012530185,0.00037550417,0.00013367346],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013310424,0.003271048,0.0013840061,0.0014945323,0.00046856553,0.0000026594946,0.028670197,0.22897695,0.03658459,0.1772509,0.019206582,0.50255686],"study_design_scores_gemma":[0.0002871114,0.0000024124531,0.0004909551,0.00049813197,0.000047243993,0.000002202364,0.000041294064,0.97116506,0.017834216,0.0028483672,0.006463271,0.00031972385],"about_ca_topic_score_codex":0.0002440896,"about_ca_topic_score_gemma":0.0003279125,"teacher_disagreement_score":0.7421881,"about_ca_system_score_codex":0.00059926155,"about_ca_system_score_gemma":0.0005881003,"threshold_uncertainty_score":0.99991506},"labels":[],"label_agreement":null},{"id":"W2942747183","doi":"10.1109/latw.2019.8704589","title":"Probabilistic High-Level Estimation of Vulnerability and Fault Mitigation of Critical Systems Using Fault-Mitigation Trees (FMTs)","year":2019,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Computer science; Probabilistic logic; Fault tree analysis; Reliability engineering; Redundancy (engineering); Implementation; Scalability; Vulnerability (computing); Vulnerability assessment; Markov process; Engineering; Software engineering; Artificial intelligence; Computer security; Mathematics","score_opus":0.033452535358615,"score_gpt":0.3193314998570797,"score_spread":0.2858789644984647,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2942747183","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.59622043,0.000035949208,0.40317354,0.00016308077,0.000062330255,0.0002607433,0.0000082985825,0.000033153952,0.0000424832],"genre_scores_gemma":[0.96069586,0.0000030119586,0.03921502,0.000007048482,0.000014634636,0.000011917316,0.000011185375,0.000005315437,0.000036010282],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99775213,0.00033132193,0.00061426725,0.00044673868,0.00065439934,0.0002011211],"domain_scores_gemma":[0.9974114,0.0011991396,0.00014902398,0.0005525182,0.00059884996,0.00008909414],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0011720882,0.0001278933,0.0003598677,0.00016182844,0.00007611278,0.000100561985,0.00028966117,0.00011215659,0.000022141481],"category_scores_gemma":[0.0017251812,0.00010943055,0.000071988965,0.0004591671,0.0002810095,0.0007032125,0.00012659191,0.00012071716,0.0000062528693],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000047412188,0.0009586676,0.03804236,0.0051737847,0.000107155734,0.0000018799683,0.0018236182,0.5415545,0.052507155,0.3153912,0.000025856796,0.044366363],"study_design_scores_gemma":[0.00022934543,0.00010449538,0.016478088,0.00011614794,0.000017357488,0.0000040040245,0.00011527665,0.959695,0.009835951,0.013290999,0.0000015140658,0.000111841895],"about_ca_topic_score_codex":0.00077328255,"about_ca_topic_score_gemma":0.000022005543,"teacher_disagreement_score":0.41814044,"about_ca_system_score_codex":0.00008605846,"about_ca_system_score_gemma":0.000121631856,"threshold_uncertainty_score":0.4462448},"labels":[],"label_agreement":null},{"id":"W2975056157","doi":"10.1002/j.2334-5837.2019.00592.x","title":"Implementing systems engineering and project management processes in a Canadian company – Overview and Results Achieved","year":2019,"lang":"en","type":"article","venue":"INCOSE International Symposium","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Lachine Hospital; École de Technologie Supérieure","funders":"","keywords":"Capability Maturity Model Integration; Maturity (psychological); Audit; Process (computing); Engineering management; Engineering; Systems engineering; Information security management system; Computer science; Software development process; Software development; Software; Business","score_opus":0.01551138298554937,"score_gpt":0.2827738985106589,"score_spread":0.26726251552510955,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2975056157","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9885346,0.0013588369,0.0016684723,0.0030404902,0.00037879022,0.0009775494,0.000042298707,0.00009669941,0.0039022171],"genre_scores_gemma":[0.9984097,0.00036979036,0.00091534085,0.000043563145,0.000026515596,0.00003911476,0.00001169919,0.0000053321896,0.00017893813],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99867517,0.000031034113,0.00030287646,0.00035874487,0.00030360982,0.0003285583],"domain_scores_gemma":[0.99943966,0.00013660788,0.000056528857,0.00018197652,0.00009515957,0.00009007176],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008850753,0.00009921295,0.00013618516,0.00030121827,0.00004927739,0.00035595582,0.00041318696,0.000030548497,0.0000023223636],"category_scores_gemma":[0.0000717297,0.000094572795,0.000017926663,0.00037573467,0.000011732816,0.00033800522,0.0003794979,0.00009789481,0.0000072967587],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015674657,0.0004992,0.79416424,0.0070775007,0.0010940345,0.0005492798,0.010020386,0.019339455,0.0030137761,0.12638603,0.00064179435,0.03705757],"study_design_scores_gemma":[0.0016590987,0.00008578055,0.06531837,0.000915359,0.0000123719465,0.0000695878,0.0002280973,0.7882802,0.00014716295,0.00017873493,0.14265107,0.00045417264],"about_ca_topic_score_codex":0.07476114,"about_ca_topic_score_gemma":0.026723597,"teacher_disagreement_score":0.76894075,"about_ca_system_score_codex":0.00018385233,"about_ca_system_score_gemma":0.000097822514,"threshold_uncertainty_score":0.9910362},"labels":[],"label_agreement":null},{"id":"W2977953926","doi":"10.1007/978-3-030-32079-9_7","title":"Accelerated Learning of Predictive Runtime Monitors for Rare Failure","year":2019,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Event (particle physics); Inference; Markov chain; Exploit; Rare events; Sample (material); Sampling (signal processing); Construct (python library); Importance sampling; Artificial intelligence; Machine learning; Data mining; Programming language; Statistics; Mathematics; Monte Carlo method","score_opus":0.01997430481071661,"score_gpt":0.2742073740594166,"score_spread":0.2542330692487,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2977953926","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000508533,0.00027135655,0.9970335,0.00036504332,0.00051454693,0.0007122677,0.000013384751,0.00013233897,0.0004490598],"genre_scores_gemma":[0.6985162,0.000075346165,0.2990749,0.00018459342,0.00047141383,0.000031862088,0.00003170121,0.00006264604,0.0015513673],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9959604,0.00006802559,0.0005858885,0.0015046692,0.0012320142,0.00064899493],"domain_scores_gemma":[0.99576396,0.0015893787,0.00036449204,0.001148751,0.0009905626,0.00014284677],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001184741,0.0004340134,0.00076988625,0.001008915,0.0002460269,0.0003789807,0.0033806595,0.000417992,0.000027295897],"category_scores_gemma":[0.0004925326,0.00038299244,0.00029484762,0.0010609922,0.0006083583,0.0006135743,0.0012305806,0.0010231668,0.000025284082],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004389558,0.0000953551,0.0028410428,0.0004109698,0.00013923066,0.000023773624,0.0016756502,0.41360778,0.00070379395,0.0067337584,0.00007924749,0.57364553],"study_design_scores_gemma":[0.00038259363,0.00054802746,0.00033748776,0.0004691296,0.000018293877,0.000008538829,9.931728e-7,0.9685136,0.0041035344,0.02363367,0.0014836626,0.0005004875],"about_ca_topic_score_codex":0.000023831259,"about_ca_topic_score_gemma":0.000010361835,"teacher_disagreement_score":0.69800764,"about_ca_system_score_codex":0.00024047206,"about_ca_system_score_gemma":0.001049725,"threshold_uncertainty_score":0.9998622},"labels":[],"label_agreement":null},{"id":"W2978699951","doi":"10.48550/arxiv.1910.02452","title":"Can we rely on smartphone applications?","year":2019,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Android (operating system); Computer science; Weibull distribution; Reliability (semiconductor); Mobile device; Smartphone application; Software; Software quality; Reliability engineering; World Wide Web; Multimedia; Engineering; Operating system; Software development","score_opus":0.07682296760132028,"score_gpt":0.21153609054327727,"score_spread":0.13471312294195698,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2978699951","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.050661374,0.00013690708,0.94143355,0.0026206688,0.00026576387,0.0007019909,0.000029017625,0.00042547478,0.0037252624],"genre_scores_gemma":[0.9925088,0.0006997246,0.00086219196,0.00018909462,0.00007019988,0.0000054931934,0.000022665889,0.000014406139,0.00562739],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976203,0.00018811318,0.00018960785,0.0014222983,0.00020339983,0.0003763034],"domain_scores_gemma":[0.9965052,0.00032397747,0.00015846819,0.0025996396,0.00021226176,0.00020045853],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0003461597,0.0002624179,0.0003528088,0.0003958904,0.00016625704,0.00016225367,0.0026625278,0.00028821247,0.00004458521],"category_scores_gemma":[0.000049273225,0.0002862456,0.00030116542,0.0011088947,0.00011061722,0.00014149256,0.0019293291,0.00083804096,0.0010245657],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006379173,0.0006390776,0.010092628,0.00044091,0.00036138215,0.00018162558,0.00033670545,0.5535435,0.000039133687,0.40953726,0.0030671272,0.021696867],"study_design_scores_gemma":[0.0008421499,0.00017863445,0.0033643655,0.0002517079,0.000102223166,0.000003994025,0.00007147426,0.6779783,0.00039134087,0.28627715,0.02934854,0.0011901701],"about_ca_topic_score_codex":0.0003356018,"about_ca_topic_score_gemma":0.00004897443,"teacher_disagreement_score":0.94184744,"about_ca_system_score_codex":0.0003444745,"about_ca_system_score_gemma":0.00037545085,"threshold_uncertainty_score":0.999959},"labels":[],"label_agreement":null},{"id":"W2979674497","doi":"10.1109/qrs-c.2019.00017","title":"Security Vulnerability Metrics for Connected Vehicles","year":2019,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"","keywords":"Computer science; Computer security; Autopilot; Software; Vulnerability (computing); Software security assurance; Software development; Engineering; Security service; Information security","score_opus":0.022562175702930218,"score_gpt":0.3027522454864915,"score_spread":0.28019006978356126,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2979674497","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.50035673,0.00006556391,0.49624714,0.0011692598,0.000118432,0.00039238232,0.000004529672,0.00025538434,0.0013905569],"genre_scores_gemma":[0.9830968,0.0000071475074,0.016177794,0.00017869321,0.000022593567,0.000023105738,0.0000037919278,0.0000040002087,0.00048608263],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99856716,0.00011812909,0.0002022982,0.0004694304,0.0003511909,0.00029176983],"domain_scores_gemma":[0.9969215,0.0018060684,0.000034671288,0.00077958417,0.00035582492,0.00010235001],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012631641,0.00008972445,0.00020157381,0.00014178314,0.000090494104,0.0001514219,0.0007908905,0.00007211505,0.00014115009],"category_scores_gemma":[0.0012114067,0.00007145178,0.00015411111,0.0011533457,0.000036664907,0.00030900803,0.00022322984,0.00013263678,0.00015901717],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006727061,0.0010536183,0.33210522,0.0004850646,0.00021240454,0.0000032793707,0.00074176525,0.00058564526,0.003246122,0.5409527,0.0070164055,0.11353054],"study_design_scores_gemma":[0.00093351514,0.0002562504,0.012414679,0.000007348524,0.000010039151,0.0000019485192,0.00006706018,0.84534764,0.014041282,0.10925033,0.017328693,0.00034121447],"about_ca_topic_score_codex":0.00006309029,"about_ca_topic_score_gemma":0.000017325812,"teacher_disagreement_score":0.844762,"about_ca_system_score_codex":0.000052475378,"about_ca_system_score_gemma":0.000080657686,"threshold_uncertainty_score":0.29137188},"labels":[],"label_agreement":null},{"id":"W2983522883","doi":"10.7554/elife.29744.023","title":"Author response: Sculpting ion channel functional expression with engineered ubiquitin ligases","year":2017,"lang":"en","type":"peer-review","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Columbia College","funders":"","keywords":"Ubiquitin; Ubiquitin ligase; Ion channel; Cell biology; Chemistry; Ubiquitin-Protein Ligases; Protein degradation; Computational biology; Biology; Biochemistry; Gene","score_opus":0.07922055864056461,"score_gpt":0.34758716078211055,"score_spread":0.26836660214154595,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2983522883","genre_codex":"methods","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00030876652,0.017777814,0.79445904,0.18160081,0.0020164012,0.0010845049,0.00009300056,0.0009458406,0.0017138278],"genre_scores_gemma":[0.005120006,0.0014952419,0.023379466,0.0011832326,0.001438317,0.00043764905,0.0004887565,0.00008537055,0.96637195],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99501544,0.000528834,0.00054177735,0.0012753052,0.001995505,0.0006431269],"domain_scores_gemma":[0.99486434,0.0011947475,0.00038094466,0.0023142977,0.0009379775,0.00030767292],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.003694044,0.0004911461,0.0008417554,0.0004405824,0.00067331863,0.00061184727,0.0021667993,0.00035866743,0.00041929437],"category_scores_gemma":[0.0038229937,0.0003401336,0.00038129833,0.0005193676,0.000119203076,0.00054224057,0.0009499604,0.0008786336,0.00018634743],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010908759,0.00009854152,0.000008769583,0.0016028294,0.00008865414,0.000053247706,0.00004568407,0.0002848781,0.00028139073,0.000110377136,0.98567104,0.011645475],"study_design_scores_gemma":[0.00042323492,0.000289896,0.0002956383,0.013932188,0.00009069354,0.00006515404,0.000021241858,0.012195042,0.003787034,0.00019447997,0.967833,0.00087242236],"about_ca_topic_score_codex":0.00011592398,"about_ca_topic_score_gemma":0.000026161348,"teacher_disagreement_score":0.96465814,"about_ca_system_score_codex":0.00016532317,"about_ca_system_score_gemma":0.0008523477,"threshold_uncertainty_score":0.99990505},"labels":[],"label_agreement":null},{"id":"W3043308735","doi":"10.22215/etd/2014-10596","title":"UML Model to Fault Tree Model Transformation for Dependability Analysis","year":2014,"lang":"en","type":"dissertation","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"","keywords":"Dependability; Fault tree analysis; Sequence diagram; Computer science; Unified Modeling Language; Model transformation; Transformation (genetics); Programming language; Applications of UML; Class diagram; Systems Modeling Language; Activity diagram; Software engineering; Theoretical computer science; Reliability engineering; Software; Artificial intelligence; Engineering; Consistency (knowledge bases)","score_opus":0.031472994412580405,"score_gpt":0.3402905023223898,"score_spread":0.3088175079098094,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3043308735","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016179465,0.000016768141,0.9772078,0.0006423382,0.000057340967,0.0008351333,0.00004738976,0.00023413183,0.004779637],"genre_scores_gemma":[0.80643743,0.000021041633,0.18324338,0.00022983336,0.00003376698,0.0004900124,0.0013026852,0.000025002368,0.008216869],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99668103,0.00009831633,0.0007721896,0.0010884717,0.00084378663,0.0005161796],"domain_scores_gemma":[0.99704427,0.0002775138,0.00014087836,0.0013466565,0.0008976486,0.00029306326],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0014338316,0.00036112528,0.00080107286,0.0010236525,0.0002470963,0.00035139621,0.0016279152,0.0003975018,0.000024236226],"category_scores_gemma":[0.00035414356,0.0003154809,0.0011053736,0.0019008421,0.000017651842,0.0005723585,0.000056655284,0.00027839275,0.000041746203],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006134029,0.00011128656,0.000023465745,0.0002499173,0.00036558582,1.2541503e-7,0.0014396041,0.82948554,0.00010017723,0.013008094,0.00034980735,0.15480506],"study_design_scores_gemma":[0.00019489824,0.00005248615,0.0001212071,0.000011390387,0.0003877835,8.4904734e-8,0.000060476043,0.9777074,0.0010465652,0.019956894,0.000098964345,0.00036188198],"about_ca_topic_score_codex":0.00021761392,"about_ca_topic_score_gemma":0.0072921696,"teacher_disagreement_score":0.7939644,"about_ca_system_score_codex":0.00017542166,"about_ca_system_score_gemma":0.00034677377,"threshold_uncertainty_score":0.9999297},"labels":[],"label_agreement":null},{"id":"W3080084182","doi":"10.1007/978-3-030-57993-7_8","title":"Bayesian Modelling for Product Testing and Release","year":2020,"lang":"en","type":"book-chapter","venue":"IFIP advances in information and communication technology","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Western University","funders":"","keywords":"Product (mathematics); Reliability (semiconductor); Computer science; Reliability engineering; Bayesian probability; Process (computing); New product development; Product design; Risk analysis (engineering); Engineering; Business; Mathematics; Artificial intelligence; Marketing","score_opus":0.02564438308093037,"score_gpt":0.2752641842245442,"score_spread":0.24961980114361382,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3080084182","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000019666088,0.012840272,0.9604743,0.0052303555,0.000024019924,0.00054794975,0.000007084738,0.0002965403,0.020559816],"genre_scores_gemma":[0.13837197,0.065472424,0.7933764,0.00063932675,0.000032887023,0.00035188338,0.00010605081,0.000028389766,0.0016206682],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9990069,0.00001872066,0.0004758128,0.00023279317,0.0001318478,0.00013392458],"domain_scores_gemma":[0.9983573,0.0003745157,0.00026413787,0.000730544,0.00022901426,0.000044505225],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003200233,0.00014384407,0.00024042076,0.00049213815,0.00017669144,0.00010692367,0.000707421,0.00018042455,0.000001691623],"category_scores_gemma":[0.00054622116,0.00014859006,0.000026768152,0.00024944017,0.00020685174,0.0012832416,0.00046687564,0.0004407761,0.000005353912],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000004201521,0.0000033672732,0.000043900076,0.00010867077,0.0000044325943,1.4238394e-7,0.00010689855,0.00088213256,6.072233e-7,0.4418936,0.000009140989,0.5569429],"study_design_scores_gemma":[0.00016206625,0.000042175787,0.000001883684,0.000114795366,0.0000044945896,0.0000063187863,0.00004201885,0.4067894,0.0000263188,0.44874653,0.1439219,0.00014211932],"about_ca_topic_score_codex":0.000005941495,"about_ca_topic_score_gemma":0.000006944139,"teacher_disagreement_score":0.5568008,"about_ca_system_score_codex":0.00003967572,"about_ca_system_score_gemma":0.000060438204,"threshold_uncertainty_score":0.60593265},"labels":[],"label_agreement":null},{"id":"W3082189468","doi":"10.1109/mcse.2020.3019770","title":"Raising the Bar: Assurance Cases for Scientific Software","year":2020,"lang":"en","type":"article","venue":"Computing in Science & Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Documentation; Software engineering; Correctness; Software; Software quality analyst; Software system; Software construction; Programming language","score_opus":0.057996320672263915,"score_gpt":0.3051431806933281,"score_spread":0.24714686002106417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3082189468","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.18835056,0.00021361273,0.8091589,0.0014941989,0.00035875652,0.00018674207,9.876817e-7,0.00022616112,0.000010088024],"genre_scores_gemma":[0.9254303,0.000002256969,0.07430765,0.00014253224,0.00009559845,0.000008506994,5.084517e-7,0.00000774479,0.0000048861],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975417,0.000033423115,0.00029957615,0.00076423696,0.00068121514,0.0006798017],"domain_scores_gemma":[0.99714696,0.0018185097,0.00005723183,0.0005669846,0.0002341533,0.00017615796],"candidate_categories":["scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0027897072,0.0001407689,0.00019055247,0.00021878124,0.0007512874,0.001238235,0.0025582814,0.00003036293,0.0000017994331],"category_scores_gemma":[0.00801029,0.00011269672,0.00009363823,0.0038961591,0.00034132643,0.0006576708,0.00064801174,0.00024212938,0.00001130315],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000033802085,0.000027982964,0.007739831,0.00008838269,0.000006462523,0.000021045944,0.0027933621,0.92268467,0.0098786075,0.009980674,0.00016731348,0.04660828],"study_design_scores_gemma":[0.00011799354,0.000026046127,0.004287351,0.00006586967,0.0000016928828,0.000009762298,0.000028900604,0.99023587,0.0025798383,0.00015980004,0.0023320012,0.00015488731],"about_ca_topic_score_codex":0.000009528972,"about_ca_topic_score_gemma":0.0000023088626,"teacher_disagreement_score":0.73707974,"about_ca_system_score_codex":0.00009662545,"about_ca_system_score_gemma":0.00031126195,"threshold_uncertainty_score":0.9997986},"labels":[],"label_agreement":null},{"id":"W3114029247","doi":"10.20982/tqmp.16.4.p424","title":"An Equivalence Testing Approach for Evaluating Substantial Mediation","year":2020,"lang":"en","type":"article","venue":"The Quantitative Methods for Psychology","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University","funders":"","keywords":"Equivalence (formal languages); Mediation; Computer science; Psychology; Mathematics; Sociology; Discrete mathematics; Social science","score_opus":0.5316020390732338,"score_gpt":0.605027051904734,"score_spread":0.07342501283150027,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3114029247","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.004714053,0.00024242111,0.99055606,0.003149069,0.00015552506,0.00088192115,0.0000096072035,0.00013241453,0.00015890873],"genre_scores_gemma":[0.038546696,0.0000071396375,0.9603379,0.00065767457,0.00012246665,0.00029455798,0.000014148852,0.000014823976,0.000004625073],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99578875,0.0024278213,0.00038572337,0.0007484355,0.00024078845,0.0004084637],"domain_scores_gemma":[0.9876364,0.01088653,0.00021250192,0.0005528039,0.0005980489,0.00011370084],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0102687795,0.00014866289,0.00030448067,0.000082265615,0.00038305702,0.00010102424,0.001466699,0.000079734775,0.000005235495],"category_scores_gemma":[0.014825597,0.00010760361,0.00014404042,0.0010371237,0.00021497245,0.00031043982,0.00009476959,0.00020376896,0.000004810833],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032023303,0.00016656559,0.00035217678,0.00012924119,0.000107835505,4.2982083e-7,0.0071516577,0.0025684666,0.14173584,0.14269483,0.00010080137,0.7046719],"study_design_scores_gemma":[0.00048823634,0.0017605975,0.00027595222,0.0000034322113,0.000024395888,0.0000027219069,0.000268263,0.92501277,0.0019040618,0.06996982,0.00015898292,0.00013075759],"about_ca_topic_score_codex":0.000006066856,"about_ca_topic_score_gemma":4.5930656e-7,"teacher_disagreement_score":0.92244434,"about_ca_system_score_codex":0.000019372772,"about_ca_system_score_gemma":0.00007813017,"threshold_uncertainty_score":0.99347293},"labels":[],"label_agreement":null},{"id":"W3121645392","doi":"10.7287/peerj.preprints.2601","title":"Anatomy of a crash repository","year":2016,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Crash; Computer science; Parsing; Debugging; Metadata; Data science; Programming language; World Wide Web","score_opus":0.010626706529955912,"score_gpt":0.2863821910409885,"score_spread":0.2757554845110326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3121645392","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0809125,0.000044829558,0.9049967,0.0018987572,0.00004801594,0.00003878649,2.6578175e-7,0.000106284824,0.0119539],"genre_scores_gemma":[0.98780745,0.000011002117,0.009232475,0.000027642373,0.000011572786,0.000002292223,2.6351668e-8,0.0000012240278,0.0029063374],"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.9993345,0.000043499687,0.00012503413,0.00016263529,0.00022844794,0.0001058875],"domain_scores_gemma":[0.999159,0.00020938524,0.000024862704,0.00045794647,0.000100795405,0.00004805937],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00021029553,0.000032973447,0.000077929864,0.00006432996,0.000025335303,0.000018354936,0.0004919702,0.000021970036,0.00005032238],"category_scores_gemma":[0.00009931914,0.000017688133,0.00006127959,0.00025829455,0.000054628956,0.00017532929,0.00014610453,0.000022178385,0.000039722516],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000061142596,0.0002549504,0.19535248,0.000043554504,0.00008850536,0.000029480809,0.00018633188,0.0000049455984,0.07676595,0.21267062,0.006063477,0.5085336],"study_design_scores_gemma":[0.001131892,0.00028962476,0.10289922,0.00012389357,0.000013401733,0.00003402248,0.000038999024,0.012034323,0.80073845,0.047569714,0.03464899,0.00047744001],"about_ca_topic_score_codex":0.000033946497,"about_ca_topic_score_gemma":0.0000021687906,"teacher_disagreement_score":0.9068949,"about_ca_system_score_codex":0.000014319805,"about_ca_system_score_gemma":0.00004738933,"threshold_uncertainty_score":0.09142113},"labels":[],"label_agreement":null},{"id":"W3127412380","doi":"10.1016/b978-0-12-819582-6.00012-5","title":"Reliability, MTTF, and sensitivity evaluation of a computer network system connected in star topology","year":2021,"lang":"en","type":"book-chapter","venue":"Elsevier eBooks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Mean time between failures; Reliability (semiconductor); Network topology; Topology (electrical circuits); Sensitivity (control systems); Markov process; Laplace transform; Star network; Transformation (genetics); Reliability engineering; Markov chain; Star (game theory); Process (computing); Computer science; Engineering; Failure rate; Mathematics; Electronic engineering; Computer network; Electrical engineering; Physics; Operating system; Astrophysics; Statistics","score_opus":0.026368043499604654,"score_gpt":0.27770534371400357,"score_spread":0.25133730021439893,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127412380","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.06336831,0.009535777,0.028371938,0.000759156,0.0020329268,0.0053658416,0.00007791719,0.0005323289,0.8899558],"genre_scores_gemma":[0.867864,0.00031785716,0.032010645,0.00032514767,0.0009311119,0.00013267931,0.000099410056,0.00011344706,0.09820571],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99550974,0.0013125966,0.0007873261,0.0009937743,0.0010152678,0.0003813094],"domain_scores_gemma":[0.9961478,0.0011362553,0.00029868112,0.0011833287,0.0011136681,0.00012027654],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.005823671,0.00032333456,0.0010475874,0.00024561322,0.00008404345,0.00007970867,0.00030682932,0.00042617833,0.00002894014],"category_scores_gemma":[0.00021859538,0.00031427314,0.0002159111,0.000118400465,0.0002893154,0.00008590722,0.0008343835,0.00057329645,0.00000849271],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000013867447,0.00002817021,0.0002915371,0.00051658123,0.00013575002,0.000112458045,0.0002702115,0.0008956449,0.000017350847,0.041768562,0.0000220612,0.9559278],"study_design_scores_gemma":[0.002214606,0.00041495077,0.0061596134,0.0044608545,0.00051583105,0.00027665048,0.000056900033,0.8891412,0.000092327675,0.053519838,0.041666534,0.0014806961],"about_ca_topic_score_codex":0.000021253194,"about_ca_topic_score_gemma":0.00035170285,"teacher_disagreement_score":0.9544471,"about_ca_system_score_codex":0.00031948567,"about_ca_system_score_gemma":0.00053496857,"threshold_uncertainty_score":0.9999309},"labels":[],"label_agreement":null},{"id":"W3127476678","doi":"10.1016/b978-0-12-804288-5.00015-5","title":"Some extended testing procedures and optimal design of tests","year":2021,"lang":"en","type":"book-chapter","venue":"Elsevier eBooks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Reliability engineering; Acceptance testing; Engineering; Software engineering","score_opus":0.04090456269504005,"score_gpt":0.280212640240276,"score_spread":0.23930807754523592,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3127476678","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0004702516,0.040262327,0.0361954,0.00040309073,0.00034261344,0.0023156551,0.000028827397,0.0006328181,0.919349],"genre_scores_gemma":[0.002740919,0.00048239526,0.17335787,0.00013277252,0.0002555298,0.00004828948,0.0000049851483,0.0000607822,0.82291645],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9976596,0.00008385352,0.000495232,0.000788059,0.0006370591,0.0003361532],"domain_scores_gemma":[0.9974471,0.00089524494,0.00023940252,0.0008139984,0.00044272759,0.00016157726],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00065278134,0.00032772473,0.00063737325,0.00023461223,0.00012369548,0.00016925605,0.00076757814,0.00023778508,0.000017267783],"category_scores_gemma":[0.0007876319,0.0002981438,0.00016102708,0.00006918728,0.00025225908,0.00014409571,0.0007077851,0.0004398504,0.000011542484],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000004236463,0.000018646086,0.000019898253,0.0003458441,0.000098519704,0.0000694178,0.0001298658,0.00005865547,0.00033185893,0.010701226,0.000018332801,0.9882035],"study_design_scores_gemma":[0.0031516072,0.002949196,0.0019872757,0.016262224,0.0011135945,0.0012097782,0.00009239378,0.08496723,0.00880896,0.6785384,0.19369775,0.007221616],"about_ca_topic_score_codex":8.299882e-7,"about_ca_topic_score_gemma":0.0000018364802,"teacher_disagreement_score":0.9809819,"about_ca_system_score_codex":0.000043616405,"about_ca_system_score_gemma":0.00073241437,"threshold_uncertainty_score":0.9999471},"labels":[],"label_agreement":null},{"id":"W3130185535","doi":"10.1109/icmla51294.2020.00123","title":"Forecasting Attrition from the Canadian Armed Forces using Multivariate LSTM","year":2020,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Department of National Defence","funders":"","keywords":"Univariate; Multivariate statistics; Attrition; Computer science; Volume (thermodynamics); Artificial intelligence; Multivariate analysis; Econometrics; Machine learning; Mathematics","score_opus":0.17508625985266524,"score_gpt":0.306841124187358,"score_spread":0.13175486433469275,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3130185535","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.16715674,0.000087464876,0.8101008,0.021534126,0.00007550906,0.0001916972,0.000010861384,0.000121783596,0.0007210054],"genre_scores_gemma":[0.9604774,0.0000023263262,0.037669063,0.0016992028,0.00012287694,0.0000033639549,0.000005505681,0.0000039590495,0.00001626182],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9988914,0.00011032811,0.00015470306,0.00029369793,0.0002713715,0.00027849505],"domain_scores_gemma":[0.9989013,0.0004920955,0.000033481912,0.0002482318,0.00009300862,0.00023188973],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00030395068,0.000074494405,0.00010654944,0.000038126407,0.0004662973,0.00037219663,0.00067015877,0.000044189434,0.00007877212],"category_scores_gemma":[0.0006262028,0.000050242583,0.00007248869,0.00056842034,0.00004050957,0.00028305958,0.00015146226,0.00015028282,0.00002988826],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007559306,0.00022430367,0.26359344,0.00015836573,0.00089887914,0.00032227507,0.04045001,0.050396938,0.015505112,0.03644124,0.008634614,0.5832992],"study_design_scores_gemma":[0.00010901026,0.000011403717,0.0018189304,0.0000103342245,0.0000055005385,9.614384e-7,0.00009151076,0.9947342,0.0005401329,0.0014795013,0.0011242714,0.00007424462],"about_ca_topic_score_codex":0.46436152,"about_ca_topic_score_gemma":0.17274866,"teacher_disagreement_score":0.94433725,"about_ca_system_score_codex":0.00008327677,"about_ca_system_score_gemma":0.00023303289,"threshold_uncertainty_score":0.84234655},"labels":[],"label_agreement":null},{"id":"W3130519220","doi":"10.1109/trustcom50675.2020.00067","title":"Evaluating the Soundness of Security Metrics from Vulnerability Scoring Frameworks","year":2020,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Soundness; Computer science; Software security assurance; Vulnerability (computing); Secure coding; Context (archaeology); Vulnerability assessment; Computer security; Vulnerability management; Software; Security information and event management; Information security; Risk analysis (engineering); Security service; Cloud computing security; Business","score_opus":0.12866832481398738,"score_gpt":0.3999609885309979,"score_spread":0.2712926637170105,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3130519220","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.42538902,0.00015989345,0.5710055,0.0030318713,0.00005784696,0.00009664866,0.0000020586824,0.0000747097,0.00018240993],"genre_scores_gemma":[0.96477216,0.000011751995,0.034658425,0.00045760596,0.000081900565,0.000007823875,0.0000012020333,0.0000037906755,0.000005317498],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975107,0.0004904002,0.00035527922,0.00048139092,0.0009356845,0.00022659119],"domain_scores_gemma":[0.99502796,0.0035311931,0.000097482836,0.0008909615,0.00032753506,0.00012484922],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0022284805,0.00010181912,0.00024368777,0.000048528644,0.00017965984,0.00018368794,0.0015667201,0.00012007473,0.00013319575],"category_scores_gemma":[0.007077342,0.000065965076,0.00015627408,0.0021214138,0.000115795396,0.00027349044,0.0007812519,0.00075144507,0.000020682539],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000085103406,0.0006596968,0.26150715,0.00032586508,0.0004460372,0.000008621971,0.03003822,0.022771256,0.003874553,0.048940975,0.000429805,0.6309127],"study_design_scores_gemma":[0.00011936376,0.000091244365,0.0028842646,0.000011244176,0.000013738853,1.8870448e-7,0.00024827765,0.93992835,0.0034119114,0.053131096,0.00006064116,0.000099652505],"about_ca_topic_score_codex":0.0005862297,"about_ca_topic_score_gemma":0.000013508392,"teacher_disagreement_score":0.9171571,"about_ca_system_score_codex":0.00003475013,"about_ca_system_score_gemma":0.000117921045,"threshold_uncertainty_score":0.8472751},"labels":[],"label_agreement":null},{"id":"W3137652218","doi":"10.33915/etd.1146","title":"Software tool for reliability estimation","year":2001,"lang":"en","type":"dissertation","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"McGill University","keywords":"Software construction; Software reliability testing; Computer science; Software sizing; Software; Avionics software; Reliability (semiconductor); Reliability engineering; Software quality; Verification and validation; Software system; Component (thermodynamics); Component-based software engineering; Software development; Embedded system; Software engineering; Engineering; Operating system; Power (physics)","score_opus":0.015958597063741524,"score_gpt":0.31891557404561055,"score_spread":0.302956976981869,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3137652218","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0053567467,0.00010130026,0.9907282,0.00028916114,0.00040970225,0.0007544296,0.000010048872,0.00053930585,0.0018111046],"genre_scores_gemma":[0.010137592,0.00014403212,0.87350696,0.00016674829,0.00018609189,0.0007586192,0.002027816,0.000050494007,0.11302162],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99750775,0.00007731912,0.00049668504,0.0008871864,0.0006375897,0.0003934691],"domain_scores_gemma":[0.9969533,0.0008909783,0.00015220777,0.0011828981,0.00072632066,0.000094298724],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009170456,0.00025958978,0.00039167542,0.00022517519,0.00023260091,0.0003380832,0.0011620796,0.00036590404,0.00015308116],"category_scores_gemma":[0.0029919534,0.00023046338,0.0003823127,0.00066526345,0.00002488353,0.00050151756,0.00007280148,0.00027616182,0.000123458],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006178977,0.00024621785,0.0007292581,0.0009038097,0.0000731365,0.0000036489791,0.00032058413,0.0017417424,0.000018533901,0.010904955,0.0091511775,0.97584516],"study_design_scores_gemma":[0.00096675195,0.0003576909,0.006697414,0.0002603357,0.00013482974,0.000006955326,0.00012781327,0.6237844,0.0025565084,0.32578984,0.03770879,0.0016086624],"about_ca_topic_score_codex":0.00006551341,"about_ca_topic_score_gemma":0.000057103785,"teacher_disagreement_score":0.9742365,"about_ca_system_score_codex":0.00014969338,"about_ca_system_score_gemma":0.00035638438,"threshold_uncertainty_score":0.9398024},"labels":[],"label_agreement":null},{"id":"W3143573151","doi":"10.1109/icse.2012.6227068","title":"Locating features in dynamically configured avionics software","year":2012,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Avionics software; Avionics; Computer science; Program comprehension; Software construction; Software; Software system; Static program analysis; Software engineering; Software development; Verification and validation; Software framework; Business process reengineering; Embedded system; Software sizing; Programming language; Engineering","score_opus":0.01235885215419218,"score_gpt":0.2799967756042588,"score_spread":0.2676379234500666,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3143573151","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.050293926,0.00062109466,0.94426686,0.0013277854,0.00013139605,0.00014289653,8.4062515e-7,0.00031141008,0.002903779],"genre_scores_gemma":[0.9215818,0.000011304231,0.07754698,0.00024923394,0.00003231549,0.0000108765,0.0000034089355,0.000005035422,0.0005590747],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9985956,0.00011009723,0.00020595016,0.00025152243,0.00036295265,0.0004738909],"domain_scores_gemma":[0.9988896,0.00035866685,0.00003147482,0.00048031737,0.000090208014,0.00014971489],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007603414,0.00009725155,0.0001515109,0.0001329568,0.000075785196,0.000121125384,0.00069712877,0.00008648454,0.00006859945],"category_scores_gemma":[0.0006587149,0.000079547855,0.00006897,0.00070411584,0.000044706183,0.00045367962,0.00023996804,0.00026287933,0.00010678415],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011903866,0.0007610752,0.374054,0.00010817059,0.000043786837,0.000018385885,0.0021297894,0.0011086412,0.00076777843,0.1857417,0.0011546473,0.43410012],"study_design_scores_gemma":[0.0012981569,0.00013637087,0.69827497,0.000115880124,0.000018512847,0.000064587715,0.00044599478,0.27447096,0.0038100446,0.01505969,0.0050931405,0.0012116863],"about_ca_topic_score_codex":0.00017374702,"about_ca_topic_score_gemma":0.00010311238,"teacher_disagreement_score":0.8712878,"about_ca_system_score_codex":0.000079161066,"about_ca_system_score_gemma":0.000076186734,"threshold_uncertainty_score":0.32438675},"labels":[],"label_agreement":null},{"id":"W3173415420","doi":"10.1109/tse.2021.3092692","title":"An Experience Report on Producing Verifiable Builds for Large-Scale Commercial Systems","year":2021,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University; Huawei Technologies (Canada)","funders":"","keywords":"Toolchain; Verifiable secret sharing; Computer science; Software engineering; Process (computing); TRACE (psycholinguistics); Code (set theory); Software; Notation; Programming language; Theoretical computer science; Set (abstract data type)","score_opus":0.019071393867009387,"score_gpt":0.28494894857246433,"score_spread":0.26587755470545493,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3173415420","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05915898,0.000060774262,0.93842775,0.000073327596,0.0013230665,0.00027084074,0.000026978321,0.0006509307,0.0000073836877],"genre_scores_gemma":[0.90162075,0.000015912863,0.0974132,0.000062502775,0.00015425977,0.00036560913,0.000012034584,0.00003645412,0.00031926183],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976586,0.00006207681,0.00038043922,0.00085953326,0.00051011593,0.0005292224],"domain_scores_gemma":[0.9977643,0.00036937968,0.000051150102,0.0013295211,0.00028172965,0.0002039018],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005666261,0.00022132446,0.00031161096,0.00019808105,0.0003926388,0.00030399364,0.00052733644,0.000120863515,0.000017394144],"category_scores_gemma":[0.00018586287,0.00023584127,0.00019002768,0.0008251245,0.000018126677,0.00054047967,0.000006147263,0.0003509243,0.000012699002],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019329163,0.0005260607,0.00029147405,0.00016430265,0.000058126872,0.000104801,0.0013755418,0.98758614,0.0018183114,0.00021363556,0.00010313382,0.007739155],"study_design_scores_gemma":[0.00063729857,0.00031577662,0.0006341088,0.0002492036,0.00003064666,0.00017146066,0.00023877833,0.8860283,0.100975886,0.000018326666,0.010042803,0.0006574086],"about_ca_topic_score_codex":0.000045231296,"about_ca_topic_score_gemma":0.000013098543,"teacher_disagreement_score":0.84246176,"about_ca_system_score_codex":0.00015679712,"about_ca_system_score_gemma":0.00014785078,"threshold_uncertainty_score":0.9617328},"labels":[],"label_agreement":null},{"id":"W3175914189","doi":"10.1109/newcas50681.2021.9462743","title":"Dynamic Fault Tree Analysis and Risk Mitigation Strategies of Data Communication System via Statistical Model Checking","year":2021,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Fault tree analysis; Computer science; Redundancy (engineering); Modular design; Reliability engineering; Triple modular redundancy; Model checking; Reliability (semiconductor); Critical system; Life-critical system; Fault tolerance; Fault detection and isolation; Distributed computing; Engineering; Theoretical computer science; Software engineering; Artificial intelligence","score_opus":0.029958012693513603,"score_gpt":0.33033529552875507,"score_spread":0.30037728283524145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3175914189","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05069373,0.00021264153,0.94849813,0.00011442785,0.000003960657,0.00004485369,0.000051396386,0.000053187192,0.00032767665],"genre_scores_gemma":[0.68562263,0.00021281371,0.31396037,0.0000023985167,0.0000011161368,0.0000022435584,0.00017805303,0.0000019259392,0.000018463074],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9984557,0.00033573536,0.00031838246,0.00039796374,0.0003655289,0.00012666863],"domain_scores_gemma":[0.9973959,0.0005072956,0.00010594211,0.0016573396,0.0002752815,0.000058286467],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008469498,0.000075594566,0.00024537215,0.0001425909,0.00012325235,0.00023217987,0.00080781995,0.000048959453,0.0000075998537],"category_scores_gemma":[0.00016111712,0.000067537396,0.00004870429,0.001071178,0.000097720054,0.0007644251,0.00071000244,0.0001261403,0.0000018577093],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015658929,0.0004001828,0.043575235,0.00071721495,0.0029264493,0.000016306722,0.0016377226,0.22242197,0.0036155616,0.2949958,0.000055452467,0.4296224],"study_design_scores_gemma":[0.00008475557,0.0000071884933,0.009425812,0.000014367948,0.00020917994,0.0000017841818,0.00064291223,0.98483175,0.0002240637,0.004494649,0.0000015327122,0.00006197932],"about_ca_topic_score_codex":0.0007608691,"about_ca_topic_score_gemma":0.0020911037,"teacher_disagreement_score":0.7624098,"about_ca_system_score_codex":0.00004032557,"about_ca_system_score_gemma":0.0001351077,"threshold_uncertainty_score":0.27540952},"labels":[],"label_agreement":null},{"id":"W3198768126","doi":"10.1142/s0218539321400076","title":"White-Box and Black-Box Reliability Modeling Framework: Integration Through Analytical Model and User Profile Validation via Deep Learning — A Practitioner’s Approach","year":2021,"lang":"en","type":"article","venue":"International Journal of Reliability Quality and Safety Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Black box; Computer science; White box; Reliability (semiconductor); Component (thermodynamics); Reliability engineering; Software; Artificial intelligence; Machine learning; Data mining; Process (computing); Artificial neural network; Software quality; Software development; Engineering; Programming language","score_opus":0.03229639751757927,"score_gpt":0.3235113993504981,"score_spread":0.2912150018329188,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3198768126","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.21727644,0.00020884974,0.7785529,0.0036357264,0.00009413908,0.00010776593,0.0000050592125,0.000045262317,0.00007388208],"genre_scores_gemma":[0.6967559,0.0008371075,0.3021945,0.00007041787,0.000084538915,0.000005224165,0.000016611375,0.000009916468,0.000025812258],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9964365,0.00041670311,0.0011709625,0.000626063,0.0010796749,0.00027010613],"domain_scores_gemma":[0.99609447,0.0012460822,0.0003151661,0.0003663973,0.0017772019,0.00020065803],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0044254786,0.00023774928,0.00048187916,0.0001756978,0.00017057921,0.00044042376,0.00037707158,0.0002485516,0.000013886614],"category_scores_gemma":[0.005928209,0.00022062534,0.00017575952,0.0003749358,0.00014416662,0.0019301242,0.0003872516,0.0011907965,9.974327e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001112631,0.00021410715,0.0066129663,0.00019336266,0.0001243924,0.000008673357,0.0021216727,0.95711887,0.0002888226,0.025922144,0.0000044388553,0.0072792666],"study_design_scores_gemma":[0.00038999386,0.000049908704,0.0040305457,0.00012866409,0.00003559154,0.00009670761,0.0003809336,0.9753371,0.0002235024,0.019016702,0.00009218485,0.00021818082],"about_ca_topic_score_codex":0.000053414,"about_ca_topic_score_gemma":0.000004143896,"teacher_disagreement_score":0.47947943,"about_ca_system_score_codex":0.0002327781,"about_ca_system_score_gemma":0.0001605582,"threshold_uncertainty_score":0.899684},"labels":[],"label_agreement":null},{"id":"W3203682716","doi":"10.1109/icas49788.2021.9551199","title":"Fault Tree Analysis And Risk Mitigation Strategies For Autonomous Systems Via Statistical Model Checking","year":2021,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Fault tree analysis; Computer science; Reliability (semiconductor); Reliability engineering; Tree (set theory); Fault tolerance; Distributed computing; Engineering","score_opus":0.02069252906400945,"score_gpt":0.2979113082215689,"score_spread":0.27721877915755944,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3203682716","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.033715695,0.00016509523,0.96544707,0.0001431902,0.000022122662,0.000106950545,0.00002436777,0.000093131115,0.00028235948],"genre_scores_gemma":[0.8218113,0.00005173441,0.17777547,0.000015733951,0.000016919046,0.00003127382,0.000028551074,0.000003830273,0.0002651482],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9985757,0.00012952082,0.0002711573,0.0004853277,0.00030046873,0.00023786203],"domain_scores_gemma":[0.9985562,0.0005849448,0.000060032795,0.0003770663,0.00031147254,0.00011031484],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005433529,0.00010120453,0.00027250085,0.00015366418,0.00018588558,0.0007583352,0.00021973782,0.000064962755,0.000009434001],"category_scores_gemma":[0.0002069937,0.00008779746,0.00011736388,0.0007943479,0.000055280187,0.00039084692,0.00011104051,0.00010037702,0.0000030800334],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010900944,0.00016857559,0.016707778,0.00021725119,0.0015542598,0.000018316614,0.00096729374,0.54784954,0.0010223753,0.28907675,0.00017090175,0.14223605],"study_design_scores_gemma":[0.00013435008,0.000021200844,0.0022442394,0.0000039242746,0.00018556876,0.0000023161208,0.00022365752,0.9764227,0.00033358415,0.02029137,0.000029800416,0.000107254855],"about_ca_topic_score_codex":0.0004919496,"about_ca_topic_score_gemma":0.00047332648,"teacher_disagreement_score":0.78809565,"about_ca_system_score_codex":0.000046917954,"about_ca_system_score_gemma":0.00022101984,"threshold_uncertainty_score":0.7312644},"labels":[],"label_agreement":null},{"id":"W3212387882","doi":"10.32920/ryerson.14663400.v1","title":"Development Framework for Performance-Based Output Specifications to Encourage Innovation in Public-Private Partnerships","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Procurement; Business; Quality (philosophy); Product (mathematics); Private sector; Hierarchy; Scale (ratio); Process management; Public sector; Engineering management; Engineering; Marketing; Economics; Economic growth; Mathematics","score_opus":0.2814960032416046,"score_gpt":0.3488015418627655,"score_spread":0.0673055386211609,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3212387882","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13909647,0.000017865766,0.84812635,0.01144707,0.00020417628,0.0007086043,0.0000028088095,0.00014960428,0.00024702057],"genre_scores_gemma":[0.4831935,0.000012250382,0.51551473,0.0003597501,0.000047679234,0.00060082367,0.00011656687,0.000009129422,0.0001455677],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99658424,0.00012468253,0.0009458169,0.0010670371,0.00072458,0.00055362395],"domain_scores_gemma":[0.99664456,0.00055903353,0.00018609487,0.001502848,0.0009554627,0.00015202326],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0024431688,0.00026206236,0.00038613836,0.001219043,0.00019893948,0.0010282381,0.001836381,0.00032902244,0.00006581946],"category_scores_gemma":[0.0019744132,0.00025658868,0.00012015925,0.0035724044,0.000034702858,0.00029427838,0.0012175321,0.00076018006,0.00004701398],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000031455646,0.0010792481,0.031437635,0.0010149628,0.00015914669,0.000008912783,0.0040974477,0.023288516,0.00015702762,0.52914244,0.00084295037,0.40874025],"study_design_scores_gemma":[0.0010407146,0.00014022616,0.08505572,0.0016167756,0.00002261609,0.0000017541945,0.0008234671,0.6414283,0.024757424,0.04000722,0.20252068,0.0025851184],"about_ca_topic_score_codex":0.000007916033,"about_ca_topic_score_gemma":0.00006709454,"teacher_disagreement_score":0.61813974,"about_ca_system_score_codex":0.0004970912,"about_ca_system_score_gemma":0.0018262258,"threshold_uncertainty_score":0.9999886},"labels":[],"label_agreement":null},{"id":"W4205180101","doi":"10.1007/978-94-007-0753-5_2464","title":"Reliability Generalization","year":2014,"lang":"en","type":"book-chapter","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Generalization; Reliability engineering; Reliability (semiconductor); Computer science; Psychology; Mathematics; Engineering; Physics; Mathematical analysis; Thermodynamics","score_opus":0.021005725289992012,"score_gpt":0.26163174890882995,"score_spread":0.24062602361883795,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4205180101","genre_codex":"methods","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000025852923,0.000062229024,0.5869742,0.00038565064,0.00008826651,0.0000882738,0.0000010691862,0.0002182888,0.41217944],"genre_scores_gemma":[0.0018672437,0.0001886877,0.01950789,0.00041276898,0.00017981902,0.0000066655516,0.000027248898,0.000020635824,0.97778904],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99801236,0.000051964496,0.0003468871,0.0007279405,0.00064134767,0.00021951443],"domain_scores_gemma":[0.9976396,0.00019542444,0.00010073689,0.0015925175,0.00034265226,0.00012906714],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00070502015,0.000218176,0.0003306849,0.00016887869,0.000096521006,0.00017137628,0.0011062027,0.0003000135,0.0008502967],"category_scores_gemma":[0.00017043279,0.00018065434,0.00025117001,0.00008242471,0.00008488456,0.00013122725,0.0004110518,0.00027621572,0.0010773323],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[7.704676e-7,0.000007107377,0.000019924333,0.000032001255,0.000018139273,0.0000018339964,0.000011097982,0.00015099705,0.0000057519064,0.95244575,0.007848467,0.039458185],"study_design_scores_gemma":[0.000066973196,0.000036793306,0.00003774917,0.000023919518,0.00001186507,0.0000022756558,1.8107036e-7,0.04320819,0.000043912427,0.2228332,0.7334493,0.00028568256],"about_ca_topic_score_codex":0.000030363572,"about_ca_topic_score_gemma":0.000011570706,"teacher_disagreement_score":0.7296125,"about_ca_system_score_codex":0.00009777729,"about_ca_system_score_gemma":0.00010997012,"threshold_uncertainty_score":0.9997004},"labels":[],"label_agreement":null},{"id":"W4205741845","doi":"10.1109/icisat54145.2021.9678411","title":"Checklist-based Software Quality Evaluation of Tango Controls","year":2021,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary","funders":"","keywords":"Computer science; Checklist; Software quality; Quality (philosophy); Software; Software engineering; Software development; Operating system; Psychology","score_opus":0.07073781121475682,"score_gpt":0.3805900324477471,"score_spread":0.3098522212329903,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4205741845","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11975422,0.00032223633,0.876678,0.0014287558,0.000052588148,0.00014363718,0.0000050887206,0.00011014254,0.0015053359],"genre_scores_gemma":[0.9523533,0.000004142263,0.046982385,0.0002817671,0.000016920882,0.000023217446,0.000012569811,0.0000032342205,0.00032243927],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9970238,0.0007423926,0.00033899237,0.0003520681,0.0013621683,0.00018057137],"domain_scores_gemma":[0.99631727,0.0007566016,0.00008501393,0.00085409946,0.0019103647,0.00007665244],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0038009514,0.00007614454,0.00023059818,0.000067091896,0.00005866029,0.00007985131,0.00044863884,0.00005793406,0.0004750508],"category_scores_gemma":[0.0045290026,0.00006605917,0.00017181726,0.0007458316,0.00005582241,0.0001796274,0.00011983478,0.000089083136,0.000030148996],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000045128832,0.0017699266,0.12954362,0.00035452223,0.0002828079,0.000018126315,0.00048808864,0.014348236,0.022703312,0.09565843,0.0019506637,0.73283714],"study_design_scores_gemma":[0.002587744,0.00007955527,0.090736024,0.000040752315,0.00007125154,0.0000024606272,0.0001001342,0.7509965,0.13165778,0.02170837,0.0016790141,0.00034044747],"about_ca_topic_score_codex":0.00013541366,"about_ca_topic_score_gemma":0.00007854487,"teacher_disagreement_score":0.8325991,"about_ca_system_score_codex":0.00007474181,"about_ca_system_score_gemma":0.0008120828,"threshold_uncertainty_score":0.54219663},"labels":[],"label_agreement":null},{"id":"W4206812656","doi":"10.1002/9780470382844.ch15","title":"Software Reliability","year":2008,"lang":"en","type":"other","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Reliability (semiconductor); Reliability engineering; Software quality; Software; Software engineering; Operating system; Engineering; Software development; Physics","score_opus":0.016021009266382738,"score_gpt":0.2653716589717222,"score_spread":0.2493506497053395,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4206812656","genre_codex":"methods","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000012324776,0.0005074609,0.50957704,0.00028118028,0.00014614327,0.00014448303,0.000005053181,0.0017683106,0.4875691],"genre_scores_gemma":[0.000027719243,0.00071907544,0.14641576,0.00015913017,0.00013364971,0.000019016232,0.000008910395,0.00012360679,0.85239315],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99776775,0.000118605356,0.00022338728,0.0008414761,0.0006768878,0.0003718742],"domain_scores_gemma":[0.9973898,0.00019882861,0.00008150274,0.0020806899,0.00008177422,0.00016739228],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00026078394,0.0002407533,0.00038490727,0.00031197374,0.000073672614,0.00008481463,0.0018042357,0.00033620506,0.0035256213],"category_scores_gemma":[0.00046742026,0.0001924388,0.0002513571,0.00065908197,0.00014492415,0.00010081292,0.00054328796,0.0003301153,0.002355322],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[4.4891584e-7,0.00006107983,0.00055523927,0.000044439723,0.000026231348,0.00001855256,0.000014848261,0.000003777523,2.553566e-7,0.0008869784,0.9745737,0.023814451],"study_design_scores_gemma":[0.00007911531,0.000019432802,0.000105849205,0.000030189614,0.0000039552033,0.000008103309,0.0000010674353,0.0006051289,0.000012442758,0.00096233806,0.99792045,0.0002519478],"about_ca_topic_score_codex":0.000641436,"about_ca_topic_score_gemma":0.000056848614,"teacher_disagreement_score":0.36482403,"about_ca_system_score_codex":0.00006127336,"about_ca_system_score_gemma":0.00021291732,"threshold_uncertainty_score":0.99842143},"labels":[],"label_agreement":null},{"id":"W4211099021","doi":"10.1007/978-3-642-39512-3_3","title":"Software Models","year":2013,"lang":"en","type":"book-chapter","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"","keywords":"Unified Modeling Language; Computer science; Applications of UML; Object Constraint Language; UML tool; Systems Modeling Language; Modeling language; Class diagram; Software engineering; Programming language; Dependability; Software","score_opus":0.03946716358122084,"score_gpt":0.24950368954122515,"score_spread":0.2100365259600043,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4211099021","genre_codex":"methods","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[3.432135e-7,0.00027156225,0.5182314,0.00033246147,0.00005206744,0.00010739009,0.0000020539976,0.00031548744,0.48068723],"genre_scores_gemma":[0.00013882994,0.00027558947,0.08206408,0.00029482073,0.000077735895,0.00001412862,0.000009351362,0.000024497796,0.91710097],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99790514,0.00001701903,0.0002982507,0.0007136539,0.00074587576,0.0003200526],"domain_scores_gemma":[0.9977328,0.0002345416,0.0000773571,0.0014487796,0.0003117039,0.00019478184],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00021491741,0.00027410357,0.00037573135,0.0002254182,0.000094518786,0.0003119729,0.001738996,0.00033332518,0.0034239525],"category_scores_gemma":[0.00004597118,0.00022144697,0.0003207837,0.00006734854,0.00008189745,0.00053078734,0.0007724574,0.00044093383,0.00636791],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[3.5284017e-7,0.0000086263335,0.0000013795185,0.000022580769,0.000050139235,0.0000087154185,0.000019565283,0.00014007889,4.1126864e-7,0.8554909,0.021922192,0.122335054],"study_design_scores_gemma":[0.000052921398,0.000025364901,0.0000014540361,0.000037891336,0.0000093637655,0.00000470257,6.156337e-7,0.047575556,0.000007733344,0.77643436,0.17553915,0.00031089634],"about_ca_topic_score_codex":0.000056377157,"about_ca_topic_score_gemma":0.000006484478,"teacher_disagreement_score":0.43641374,"about_ca_system_score_codex":0.000073144474,"about_ca_system_score_gemma":0.000160762,"threshold_uncertainty_score":0.99748707},"labels":[],"label_agreement":null},{"id":"W4223509913","doi":"10.1142/s0218194022500231","title":"The Safety and Performance of Prominent Programming Languages","year":2022,"lang":"en","type":"article","venue":"International Journal of Software Engineering and Knowledge Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Lakehead University","funders":"","keywords":"Computer science; Benchmarking; Python (programming language); Java; Memory safety; Rust (programming language); Programming language; Software; Software engineering","score_opus":0.005247951322101604,"score_gpt":0.23780677742766604,"score_spread":0.23255882610556444,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4223509913","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6974987,0.009890066,0.29133445,0.00029262103,0.00079703727,0.00009476099,0.00000399082,0.00007791643,0.000010442349],"genre_scores_gemma":[0.9827968,0.00051367877,0.016532194,0.0000025450522,0.00009617062,0.0000066202333,6.548583e-7,0.000008439204,0.000042904383],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99906427,0.000020486883,0.0003166081,0.000103822014,0.00035413567,0.00014068074],"domain_scores_gemma":[0.9990428,0.000480726,0.00009714731,0.00010600978,0.00021001988,0.000063263],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008162387,0.000088216,0.00014589497,0.00020190404,0.00010147467,0.00007801487,0.0005489874,0.000017617082,0.000001978742],"category_scores_gemma":[0.00049162965,0.000069368514,0.000060024406,0.00020243559,0.000023218978,0.0001783283,0.00038584182,0.00027807063,2.0521145e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000118733755,0.0001243876,0.024445152,0.00033125776,0.00062324567,0.000061636245,0.003178506,0.15155682,0.0017196604,0.005033625,0.00005742347,0.81274956],"study_design_scores_gemma":[0.0008502208,0.00051119603,0.023332935,0.00029463097,0.000027177652,0.0007149259,0.00023394178,0.9143453,0.0018657374,0.000024140152,0.057487946,0.00031179533],"about_ca_topic_score_codex":0.0000026588746,"about_ca_topic_score_gemma":3.1449434e-7,"teacher_disagreement_score":0.8124378,"about_ca_system_score_codex":0.00006961483,"about_ca_system_score_gemma":0.000050754305,"threshold_uncertainty_score":0.28287658},"labels":[],"label_agreement":null},{"id":"W4225646930","doi":"10.3390/software1020007","title":"Dependability Modeling of Software Systems with UML and DAM: A Guide for Real-Time Practitioners","year":2022,"lang":"en","type":"article","venue":"Software","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"Ministerio de Ciencia e Innovación","keywords":"Dependability; Unified Modeling Language; Computer science; Applications of UML; Software engineering; Systems engineering; Software; Reliability engineering; Engineering; Programming language","score_opus":0.02013970520117941,"score_gpt":0.28012021978310075,"score_spread":0.25998051458192134,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4225646930","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.057713192,0.00028561198,0.9405174,0.0002463397,0.00006497735,0.00065315363,0.00014841552,0.0003246726,0.00004620669],"genre_scores_gemma":[0.70686495,0.000057725807,0.29179412,0.000056016837,0.00004514454,0.00059432356,0.000102035374,0.000037117337,0.00044853726],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9974715,0.00025388005,0.00043566967,0.000694356,0.0007730829,0.00037152827],"domain_scores_gemma":[0.9974159,0.00093067955,0.00017668756,0.0008473392,0.00048840273,0.00014101715],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0015229412,0.00017495661,0.00038796384,0.00017457666,0.00052645034,0.000119811964,0.00068574375,0.000057412835,0.00002911108],"category_scores_gemma":[0.0009781597,0.0001559387,0.00011915086,0.0006413937,0.00009829529,0.0004525405,0.000537962,0.00021809508,0.000002747506],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00056699803,0.00097310764,0.05963037,0.0025065558,0.00062022166,0.000057047608,0.003546385,0.8635695,0.00051973586,0.003999621,0.0034496458,0.060560796],"study_design_scores_gemma":[0.000984906,0.00067101826,0.00042028667,0.000082080216,0.000062931496,0.00007866244,0.0006191163,0.9904917,0.000074083866,0.0035983631,0.0024759595,0.00044091995],"about_ca_topic_score_codex":0.00096253847,"about_ca_topic_score_gemma":0.000025305426,"teacher_disagreement_score":0.6491518,"about_ca_system_score_codex":0.000172384,"about_ca_system_score_gemma":0.00040334527,"threshold_uncertainty_score":0.63589954},"labels":[],"label_agreement":null},{"id":"W4231610476","doi":"10.7287/peerj.preprints.2670","title":"Signature-based detection of behavioural deviations in flight simulators - Experiments on FlightGear and JSBSim","year":2016,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Crew; Computer science; Flight simulator; Signature (topology); Simulation; Event (particle physics); Reliability engineering; Real-time computing; Engineering; Aeronautics","score_opus":0.042923548740766024,"score_gpt":0.3111923053098991,"score_spread":0.26826875656913307,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4231610476","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6443916,0.00023955984,0.35376793,0.000547393,0.00022108277,0.00045061408,0.000009801165,0.00011600855,0.00025598373],"genre_scores_gemma":[0.9979614,0.000015626445,0.0018034067,0.00005366038,0.000019964895,0.00004721541,0.0000054374846,0.00000925016,0.000084048],"study_design_codex":"observational","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.997923,0.00020752297,0.0004077226,0.00065677974,0.000560107,0.00024487727],"domain_scores_gemma":[0.9984429,0.00040386483,0.00013906686,0.0007545068,0.00015621223,0.00010345423],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00042214862,0.00021175838,0.00034413725,0.0006004298,0.00006916957,0.00009629537,0.00059480313,0.00031366776,0.000029134919],"category_scores_gemma":[0.00015698567,0.00015655845,0.00014876093,0.00037678666,0.00007288166,0.00014570825,0.00049357873,0.00041364468,0.000009822042],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007132984,0.007934731,0.44664934,0.0017340468,0.0010237837,0.00014624183,0.0053063976,0.06890222,0.14837576,0.039359633,0.00043528405,0.27941924],"study_design_scores_gemma":[0.0016579394,0.00036087196,0.05084544,0.00055320543,0.000035761957,8.497944e-7,0.00002847234,0.2942387,0.6424075,0.008905714,0.00020387534,0.00076168554],"about_ca_topic_score_codex":0.00017026236,"about_ca_topic_score_gemma":0.000053495092,"teacher_disagreement_score":0.49403173,"about_ca_system_score_codex":0.0001670721,"about_ca_system_score_gemma":0.00014353404,"threshold_uncertainty_score":0.63842684},"labels":[],"label_agreement":null},{"id":"W4237331462","doi":"10.1145/974044.974074","title":"Computing the performability of layered distributed systems with a management architecture","year":2004,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University","funders":"","keywords":"Computer science; Fault tree analysis; Queueing theory; Reliability (semiconductor); Distributed computing; Fault management; Architecture; Software architecture; Software; Computation; Reliability engineering; Embedded system; Operating system; Computer network; Engineering; Programming language","score_opus":0.008679833706736436,"score_gpt":0.23187418168311755,"score_spread":0.2231943479763811,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4237331462","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.12287871,0.000041456842,0.8750582,0.0010733162,0.000021721553,0.0002611255,0.0000024592223,0.00008363414,0.0005793809],"genre_scores_gemma":[0.9853857,0.000003870458,0.014531552,0.000027232056,0.000011371389,0.000007978066,0.0000028204256,0.0000026993143,0.000026737258],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99865067,0.00009825017,0.0002295672,0.0002719533,0.00050813216,0.00024140786],"domain_scores_gemma":[0.9988098,0.00015947789,0.00006394313,0.0007838164,0.00012720349,0.00005577251],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007332624,0.00009638904,0.00016907329,0.000054506916,0.00016059748,0.000112362104,0.000903287,0.000026262884,0.0000042134407],"category_scores_gemma":[0.000029611008,0.000047847854,0.00007202666,0.0008032791,0.00012735081,0.000089534784,0.00032019094,0.00015651125,0.0000055472674],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000027888675,0.00025567895,0.008587399,0.00034119518,0.00020345948,0.000012596017,0.0010356173,0.9011499,0.00002849113,0.06559324,0.000035877478,0.022728598],"study_design_scores_gemma":[0.0029401095,0.0006837646,0.08735193,0.0003872227,0.0000722631,0.00010490407,0.0023324827,0.88991,0.0016624762,0.011923079,0.0019763652,0.0006554456],"about_ca_topic_score_codex":0.00028751497,"about_ca_topic_score_gemma":0.000022499222,"teacher_disagreement_score":0.86250705,"about_ca_system_score_codex":0.000062960346,"about_ca_system_score_gemma":0.000054482247,"threshold_uncertainty_score":0.19511789},"labels":[],"label_agreement":null},{"id":"W4238757451","doi":"10.32920/ryerson.14665977","title":"A Study On Automatic Software Quality And Reliability Analysis","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Computer science; Software quality assurance; Software quality; Software engineering; Reliability (semiconductor); Software; Debugging; Reliability engineering; Quality (philosophy); Static program analysis; Product (mathematics); Software construction; Software quality analyst; Set (abstract data type); Verification and validation; Software metric; Software development; Programming language; Engineering","score_opus":0.06847448746783201,"score_gpt":0.38188484865565037,"score_spread":0.31341036118781834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4238757451","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6050357,0.00008321649,0.39290917,0.00076930824,0.00008186376,0.00048673767,0.000007100382,0.00049262965,0.00013424513],"genre_scores_gemma":[0.93792886,0.000031093874,0.061459307,0.00019703172,0.000025414254,0.00013810926,0.000026828804,0.000011631852,0.0001817338],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99241024,0.002093073,0.0010091272,0.0024395515,0.0015932706,0.00045476147],"domain_scores_gemma":[0.9919781,0.0020708835,0.00025921047,0.0048207133,0.0005699317,0.00030116874],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.004585597,0.00044600718,0.0014387209,0.0006667383,0.00022671447,0.0013379792,0.0018114774,0.00030949683,0.00014522512],"category_scores_gemma":[0.0035185847,0.0003695729,0.0007766697,0.0025971807,0.00013061943,0.000228429,0.005372187,0.0011057004,0.000023108276],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012162982,0.0046890113,0.91616946,0.00077383517,0.0042678625,0.00014171556,0.005522924,0.012702405,0.0000028097177,0.0006203543,0.00012050617,0.054976918],"study_design_scores_gemma":[0.0002611222,0.00015624745,0.90151036,0.00004954046,0.0005705068,0.0000013068793,0.00081832183,0.09202459,0.000026212165,0.004022091,0.000011461275,0.0005482576],"about_ca_topic_score_codex":0.0014831618,"about_ca_topic_score_gemma":0.00023023091,"teacher_disagreement_score":0.33289313,"about_ca_system_score_codex":0.00021010806,"about_ca_system_score_gemma":0.00041043898,"threshold_uncertainty_score":0.9998756},"labels":[],"label_agreement":null},{"id":"W4239150821","doi":"10.32920/ryerson.14663400","title":"Development Framework for Performance-Based Output Specifications to Encourage Innovation in Public-Private Partnerships","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Procurement; Business; Quality (philosophy); Product (mathematics); Private sector; Scale (ratio); Hierarchy; Process management; Public sector; Set (abstract data type); Engineering management; Risk analysis (engineering); Computer science; Engineering; Marketing; Economics; Mathematics","score_opus":0.2814960032416046,"score_gpt":0.3488015418627655,"score_spread":0.0673055386211609,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4239150821","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13909647,0.000017865766,0.84812635,0.01144707,0.00020417628,0.0007086043,0.0000028088095,0.00014960428,0.00024702057],"genre_scores_gemma":[0.4831935,0.000012250382,0.51551473,0.0003597501,0.000047679234,0.00060082367,0.00011656687,0.000009129422,0.0001455677],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99658424,0.00012468253,0.0009458169,0.0010670371,0.00072458,0.00055362395],"domain_scores_gemma":[0.99664456,0.00055903353,0.00018609487,0.001502848,0.0009554627,0.00015202326],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0024431688,0.00026206236,0.00038613836,0.001219043,0.00019893948,0.0010282381,0.001836381,0.00032902244,0.00006581946],"category_scores_gemma":[0.0019744132,0.00025658868,0.00012015925,0.0035724044,0.000034702858,0.00029427838,0.0012175321,0.00076018006,0.00004701398],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000031455646,0.0010792481,0.031437635,0.0010149628,0.00015914669,0.000008912783,0.0040974477,0.023288516,0.00015702762,0.52914244,0.00084295037,0.40874025],"study_design_scores_gemma":[0.0010407146,0.00014022616,0.08505572,0.0016167756,0.00002261609,0.0000017541945,0.0008234671,0.6414283,0.024757424,0.04000722,0.20252068,0.0025851184],"about_ca_topic_score_codex":0.000007916033,"about_ca_topic_score_gemma":0.00006709454,"teacher_disagreement_score":0.61813974,"about_ca_system_score_codex":0.0004970912,"about_ca_system_score_gemma":0.0018262258,"threshold_uncertainty_score":0.9999886},"labels":[],"label_agreement":null},{"id":"W4241857059","doi":"10.32920/ryerson.14664417","title":"Probabilistic Program Execution is a Viable Way to Find Domains from Software","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Computer science; Probabilistic logic; Program comprehension; Software; Set (abstract data type); Program analysis; Probabilistic analysis of algorithms; Theoretical computer science; Algorithm; Software system; Programming language; Artificial intelligence","score_opus":0.02626209323586308,"score_gpt":0.305504360580541,"score_spread":0.2792422673446779,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4241857059","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11233107,0.00027489098,0.8815434,0.0024281954,0.00036598512,0.0015480538,0.000054025626,0.0010084443,0.00044592223],"genre_scores_gemma":[0.1521882,0.00008536104,0.8414651,0.0009896772,0.00023684383,0.0014014702,0.00033268725,0.00004064428,0.0032600188],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.995049,0.0002751205,0.0005893321,0.0021352912,0.0012291276,0.0007221167],"domain_scores_gemma":[0.9953221,0.0005030154,0.00012479638,0.0028963992,0.0007149428,0.00043874764],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0006606952,0.0004294222,0.0006860899,0.00026039314,0.00021875612,0.0018740749,0.0024642716,0.0004694515,0.0007216643],"category_scores_gemma":[0.0011770028,0.00038519636,0.00046950983,0.0012341033,0.000090420624,0.00028293816,0.0052940273,0.00087208813,0.00031365812],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004382404,0.0034652906,0.005645659,0.0012806639,0.0009349994,0.0002175455,0.007262376,0.018580137,0.0004030439,0.0029636326,0.01939273,0.9398101],"study_design_scores_gemma":[0.0014050646,0.0009465652,0.031552605,0.0022345996,0.00036773708,0.000021853963,0.00055477227,0.70120996,0.0070272526,0.19660524,0.053460646,0.0046136845],"about_ca_topic_score_codex":0.0027156947,"about_ca_topic_score_gemma":0.0002727956,"teacher_disagreement_score":0.9351964,"about_ca_system_score_codex":0.0003799638,"about_ca_system_score_gemma":0.0006748224,"threshold_uncertainty_score":0.99986},"labels":[],"label_agreement":null},{"id":"W42434049","doi":"10.1142/9789812562425_0006","title":"Properties of Software Systems Synthesized from Components","year":2004,"lang":"en","type":"book-chapter","venue":"Series on component-based software development","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":22,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Computer science","score_opus":0.03779400953434549,"score_gpt":0.22166050908704446,"score_spread":0.18386649955269896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W42434049","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04845212,0.015397079,0.90420884,0.0021885198,0.0064620064,0.008997356,0.0016858244,0.007913558,0.004694681],"genre_scores_gemma":[0.43217653,0.00079016894,0.50235814,0.0010958062,0.0005614012,0.0011339203,0.0035765166,0.0009166959,0.05739082],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9920125,0.00023924514,0.001947208,0.0018427006,0.0030520956,0.00090625783],"domain_scores_gemma":[0.9943401,0.0008946594,0.0009852381,0.0024312192,0.0008982987,0.00045046612],"candidate_categories":["metaepi_narrow"],"consensus_categories":["metaepi_narrow"],"category_scores_codex":[0.00072671217,0.0012756421,0.0020805846,0.00087202585,0.0005385946,0.0003847722,0.003055331,0.0007107069,0.00020815451],"category_scores_gemma":[0.00039718766,0.0011510706,0.0006324731,0.0002660857,0.0005497834,0.00039170953,0.00087908737,0.00093414134,0.0004883775],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.014218669,0.014627051,0.007978832,0.041402455,0.028915582,0.0047064363,0.01589338,0.11017341,0.016571427,0.26288155,0.006797059,0.47583413],"study_design_scores_gemma":[0.019659635,0.0036725667,0.00716634,0.10977607,0.0012816276,0.000111594374,0.00042423318,0.005956172,0.2295985,0.04116467,0.5581422,0.023046408],"about_ca_topic_score_codex":0.00039826054,"about_ca_topic_score_gemma":0.000035477784,"teacher_disagreement_score":0.5513451,"about_ca_system_score_codex":0.0012143549,"about_ca_system_score_gemma":0.0023774097,"threshold_uncertainty_score":0.9999995},"labels":[],"label_agreement":null},{"id":"W4245349988","doi":"10.1109/icse.2015.274","title":"Understanding the Software Fault Introduction Process","year":2015,"lang":"en","type":"article","venue":"2015 IEEE/ACM 37th IEEE International Conference on Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo","funders":"","keywords":"Debugging; Computer science; Software engineering; Process (computing); Software bug; Fault (geology); Meaning (existential); Programming language; Algorithmic program debugging; Software; Epistemology","score_opus":0.1682799662528167,"score_gpt":0.34883778124163134,"score_spread":0.18055781498881465,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4245349988","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009420894,0.000046683614,0.9803864,0.0051361118,0.0037617765,0.00024912896,0.000015369342,0.0007713224,0.00021235558],"genre_scores_gemma":[0.97399825,0.000042100186,0.023630723,0.00021818893,0.0013818659,0.000101131845,0.000026399486,0.000041969004,0.00055934384],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9963878,0.00009966546,0.0004699945,0.0007929552,0.0016966169,0.0005529758],"domain_scores_gemma":[0.99682426,0.0004826176,0.00016190471,0.0011632951,0.0010572335,0.00031067058],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011385807,0.000360304,0.00030864633,0.00040920696,0.00020721805,0.0006571236,0.0030730204,0.0001577468,0.00007289371],"category_scores_gemma":[0.0041220794,0.0002888697,0.00015078748,0.0007766498,0.00010579539,0.0009789821,0.00024792392,0.0007325542,0.00022058183],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018957294,0.00052112236,0.009852723,0.00021828423,0.00079671905,0.00011051045,0.0050260914,0.7886999,0.00075789436,0.122340746,0.055464935,0.016021485],"study_design_scores_gemma":[0.0015547624,0.0004007753,0.0007183193,0.00034371743,0.0000499502,0.00013416068,0.00090545625,0.94222546,0.0043543675,0.034758374,0.013092365,0.0014622904],"about_ca_topic_score_codex":0.000029189288,"about_ca_topic_score_gemma":0.000007922021,"teacher_disagreement_score":0.9645774,"about_ca_system_score_codex":0.00073379057,"about_ca_system_score_gemma":0.0003828702,"threshold_uncertainty_score":0.99995637},"labels":[],"label_agreement":null},{"id":"W4245656272","doi":"10.1007/978-3-662-53120-4_300114","title":"Condition Monitoring","year":2019,"lang":"en","type":"book-chapter","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières","funders":"","keywords":"Environmental science; Computer science","score_opus":0.030826940864362384,"score_gpt":0.29133977371139197,"score_spread":0.2605128328470296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4245656272","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000012206902,0.00021444733,0.12713061,0.00023140138,0.00032730735,0.0001121478,0.000002148968,0.00021077847,0.87175894],"genre_scores_gemma":[0.0038530773,0.00018859502,0.003988186,0.000039609677,0.00017188123,0.0000024314613,0.000008260294,0.000012498382,0.99173546],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99875295,0.000010968019,0.00017114461,0.00041300865,0.00048683616,0.00016508144],"domain_scores_gemma":[0.9988175,0.00016712441,0.000057139856,0.00076005165,0.00013042387,0.00006773168],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00019000372,0.00014175531,0.00021416867,0.00016290869,0.000051196035,0.00015418485,0.00071967556,0.00018615328,0.0006039287],"category_scores_gemma":[0.000022315213,0.00012430469,0.00016940082,0.00004086164,0.000031020834,0.00020466147,0.00027940623,0.00028863596,0.0038151199],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[7.850644e-7,0.00000745699,0.00017047224,0.000041440228,0.000059370024,0.000011409715,0.000019889847,0.0000209788,0.00002126298,0.9458434,0.0017299646,0.05207357],"study_design_scores_gemma":[0.00026817544,0.00013179764,0.0004091692,0.00031244452,0.00003569821,0.000012892061,0.0000057515526,0.006225838,0.0009770335,0.25673229,0.7340149,0.0008739727],"about_ca_topic_score_codex":0.000009501969,"about_ca_topic_score_gemma":8.2234675e-7,"teacher_disagreement_score":0.73228496,"about_ca_system_score_codex":0.00007355206,"about_ca_system_score_gemma":0.00011675747,"threshold_uncertainty_score":0.9969605},"labels":[],"label_agreement":null},{"id":"W4246549423","doi":"10.4018/978-1-5225-0196-1.ch044","title":"An Innovative Approach to the Development of Project Management Processes for Small-Scale Projects in a Large Engineering Company","year":2016,"lang":"en","type":"book-chapter","venue":"IGI Global eBooks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Tetra Tech (Canada); École de Technologie Supérieure","funders":"","keywords":"Engineering management; Project planning; Project management; Project management triangle; Scale (ratio); OPM3; Program management; Engineering; Process management; Extreme project management; Process (computing); Systems engineering; Business; Computer science","score_opus":0.04092949145808811,"score_gpt":0.28767419842577874,"score_spread":0.24674470696769063,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4246549423","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00062359637,0.00008117682,0.76945716,0.000057848814,0.00005687388,0.00338339,0.0000766523,0.00011830287,0.226145],"genre_scores_gemma":[0.35412565,0.000006157595,0.6276873,0.00021967306,0.00017990501,0.002949503,0.000022265549,0.000083520084,0.014726034],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9977311,0.000027356145,0.00055192254,0.0007334382,0.00047414383,0.00048207756],"domain_scores_gemma":[0.99848014,0.00009039323,0.00015330077,0.00069041405,0.0005132063,0.000072569586],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00081869424,0.00031993623,0.0004576653,0.00031519204,0.000093781244,0.00010749011,0.0015608909,0.00013621642,4.3845273e-7],"category_scores_gemma":[0.000056034864,0.00021364061,0.00008128639,0.0003736913,0.000038956365,0.00007882678,0.00057641696,0.00016206877,0.0000048718553],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004761497,0.00017841259,0.00013710362,0.0014957921,0.00021627358,0.0000030574327,0.0039231726,0.00010238112,0.000014993562,0.9539337,0.000081211336,0.039866257],"study_design_scores_gemma":[0.015223635,0.0040262034,0.010129227,0.020770123,0.0005083119,0.00008311731,0.0032798038,0.06834416,0.008471341,0.23691976,0.6188645,0.013379865],"about_ca_topic_score_codex":0.000010012342,"about_ca_topic_score_gemma":0.0001869074,"teacher_disagreement_score":0.71701396,"about_ca_system_score_codex":0.00028817804,"about_ca_system_score_gemma":0.0007915965,"threshold_uncertainty_score":0.8712011},"labels":[],"label_agreement":null},{"id":"W4246765009","doi":"10.1016/s1571-0661(05)82541-1","title":"Preface","year":2003,"lang":"en","type":"article","venue":"Electronic Notes in Theoretical Computer Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"IBM; Library science; Model checking; Computer science; Operations research; Programming language; Mathematics","score_opus":0.007605294203265922,"score_gpt":0.27226888465223936,"score_spread":0.2646635904489734,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4246765009","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03860809,0.00018334245,0.958521,0.00079669623,0.0001370414,0.00013437645,1.345218e-7,0.00012142387,0.0014979065],"genre_scores_gemma":[0.92851037,0.000024226669,0.07119757,0.00022169134,0.000023591341,0.000010286973,1.5265806e-7,0.0000053109584,0.0000068224645],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9962894,0.00031618547,0.00028561073,0.00092365046,0.0008806955,0.0013044782],"domain_scores_gemma":[0.9976775,0.0009138655,0.000036649155,0.0010281503,0.00013783519,0.00020601222],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00413469,0.0001649673,0.0002149751,0.0003195283,0.00025975858,0.00037335418,0.0026433584,0.000060646482,0.000073940035],"category_scores_gemma":[0.0012656813,0.00013849926,0.00008114592,0.0033094536,0.0016726472,0.00059957197,0.00051717844,0.0004968,0.00009371113],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000001936938,0.00007763943,0.0007786238,0.000003352988,0.000002629138,0.000005691634,0.00011089875,0.00060643157,0.00020656943,0.96474844,0.0000025913826,0.03345519],"study_design_scores_gemma":[0.00020215528,0.00017824666,0.0005218411,0.00001286412,0.0000014790205,0.000025180518,0.0000016689247,0.3280863,0.014283994,0.6560471,0.0004303106,0.00020883877],"about_ca_topic_score_codex":0.000005621223,"about_ca_topic_score_gemma":0.000004541214,"teacher_disagreement_score":0.88990223,"about_ca_system_score_codex":0.00029295724,"about_ca_system_score_gemma":0.0006595947,"threshold_uncertainty_score":0.61629385},"labels":[],"label_agreement":null},{"id":"W4251584098","doi":"10.32920/ryerson.14665977.v1","title":"A Study On Automatic Software Quality And Reliability Analysis","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Computer science; Software quality assurance; Software quality; Software engineering; Reliability (semiconductor); Debugging; Software; Reliability engineering; Quality (philosophy); Static program analysis; Product (mathematics); Software quality analyst; Software construction; Verification and validation; Set (abstract data type); Software metric; Software development; Programming language; Engineering","score_opus":0.06847448746783201,"score_gpt":0.38188484865565037,"score_spread":0.31341036118781834,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4251584098","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6050357,0.00008321649,0.39290917,0.00076930824,0.00008186376,0.00048673767,0.000007100382,0.00049262965,0.00013424513],"genre_scores_gemma":[0.93792886,0.000031093874,0.061459307,0.00019703172,0.000025414254,0.00013810926,0.000026828804,0.000011631852,0.0001817338],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99241024,0.002093073,0.0010091272,0.0024395515,0.0015932706,0.00045476147],"domain_scores_gemma":[0.9919781,0.0020708835,0.00025921047,0.0048207133,0.0005699317,0.00030116874],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.004585597,0.00044600718,0.0014387209,0.0006667383,0.00022671447,0.0013379792,0.0018114774,0.00030949683,0.00014522512],"category_scores_gemma":[0.0035185847,0.0003695729,0.0007766697,0.0025971807,0.00013061943,0.000228429,0.005372187,0.0011057004,0.000023108276],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012162982,0.0046890113,0.91616946,0.00077383517,0.0042678625,0.00014171556,0.005522924,0.012702405,0.0000028097177,0.0006203543,0.00012050617,0.054976918],"study_design_scores_gemma":[0.0002611222,0.00015624745,0.90151036,0.00004954046,0.0005705068,0.0000013068793,0.00081832183,0.09202459,0.000026212165,0.004022091,0.000011461275,0.0005482576],"about_ca_topic_score_codex":0.0014831618,"about_ca_topic_score_gemma":0.00023023091,"teacher_disagreement_score":0.33289313,"about_ca_system_score_codex":0.00021010806,"about_ca_system_score_gemma":0.00041043898,"threshold_uncertainty_score":0.9998756},"labels":[],"label_agreement":null},{"id":"W4252315572","doi":"10.4018/978-1-4666-9737-9.ch007","title":"An Innovative Approach to the Development of Project Management Processes for Small-Scale Projects in a Large Engineering Company","year":2015,"lang":"en","type":"book-chapter","venue":"Advances in IT standards and standardization research (AISSR) book series/Advances in IT standards and standardization research series","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Tetra Tech (Canada); École de Technologie Supérieure","funders":"","keywords":"Engineering management; Project planning; Project management triangle; Project management; Scale (ratio); OPM3; Program management; Engineering; Process management; Process (computing); Extreme project management; Systems engineering; Business; Computer science","score_opus":0.06364662283684819,"score_gpt":0.4113230214019805,"score_spread":0.3476763985651323,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4252315572","genre_codex":"methods","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00019859965,0.03858311,0.8970494,0.0011857919,0.00019927813,0.009480352,0.007620525,0.00013828733,0.04554464],"genre_scores_gemma":[0.058216088,0.55093575,0.34703085,0.00027515084,0.00064534997,0.012174241,0.0022414112,0.000828812,0.027652374],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.9843382,0.0009929201,0.0024311726,0.0024293168,0.007721726,0.0020866692],"domain_scores_gemma":[0.9808374,0.001162561,0.00046115788,0.0013221953,0.01576119,0.00045548318],"candidate_categories":["metaresearch","metaepi_narrow","sts","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0316569,0.0009763525,0.0018881785,0.0040304065,0.0013784512,0.0011296241,0.0017933841,0.0004855984,0.000034341618],"category_scores_gemma":[0.00434826,0.0008259072,0.00011719742,0.0057666977,0.0014514106,0.0058451355,0.0015095937,0.0017579673,5.005368e-7],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.016192028,0.0011854275,0.0022836197,0.023367986,0.00050069706,0.00012140353,0.07659831,0.06817308,0.000043808912,0.6572548,0.005353239,0.1489256],"study_design_scores_gemma":[0.0025653432,0.0017168409,0.00007915558,0.0029541273,0.000014640485,0.000010068781,0.010100469,0.002386006,0.00030030712,0.010561424,0.9683891,0.0009225104],"about_ca_topic_score_codex":0.000024399773,"about_ca_topic_score_gemma":0.005798353,"teacher_disagreement_score":0.9630359,"about_ca_system_score_codex":0.0030886945,"about_ca_system_score_gemma":0.008303471,"threshold_uncertainty_score":0.9999216},"labels":[],"label_agreement":null},{"id":"W4252446866","doi":"10.7873/date.2015.0817","title":"Towards An Accurate Reliability, Availability and Maintainability Analysis Approach for Satellite Systems Based on Probabilistic Model Checking","year":2015,"lang":"en","type":"article","venue":"Design, Automation & Test in Europe Conference & Exhibition (DATE), 2015","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal; Concordia University","funders":"","keywords":"Maintainability; Reliability engineering; Computer science; Reliability (semiconductor); Probabilistic logic; Satellite; Model checking; Artificial intelligence; Software engineering; Engineering; Programming language","score_opus":0.15304814912744188,"score_gpt":0.34656849167650194,"score_spread":0.19352034254906006,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4252446866","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.065651074,0.00006423754,0.9292713,0.00072665006,0.00008497149,0.0027521504,0.00013607905,0.0006882171,0.00062533404],"genre_scores_gemma":[0.76669127,0.000031284966,0.23106004,0.00012324374,0.00004481826,0.0005888483,0.0011285695,0.00004104839,0.00029088324],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9900532,0.003250216,0.0017525051,0.002436497,0.0015561234,0.0009514684],"domain_scores_gemma":[0.9891443,0.0018275337,0.00064286665,0.0030713708,0.004574369,0.0007395529],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.01501258,0.00070009707,0.001131878,0.0010656458,0.0003590517,0.001553606,0.0013930365,0.00036550994,0.00003927356],"category_scores_gemma":[0.016060939,0.00064789044,0.00023575482,0.004021054,0.00049017224,0.0018575514,0.00036758403,0.00057682925,0.0001198413],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002335673,0.0021141535,0.014232084,0.000774054,0.00007178714,0.000002429088,0.001791458,0.9699075,0.0005719727,0.0055060764,0.00054298766,0.004251883],"study_design_scores_gemma":[0.0009206691,0.00023084994,0.009773816,0.00009185028,0.00012436055,0.0000054036286,0.000059470003,0.97821784,0.000060774197,0.008046969,0.0017688577,0.00069913804],"about_ca_topic_score_codex":0.00037881298,"about_ca_topic_score_gemma":0.00024581605,"teacher_disagreement_score":0.7010402,"about_ca_system_score_codex":0.00064364757,"about_ca_system_score_gemma":0.0015641557,"threshold_uncertainty_score":0.99959725},"labels":[],"label_agreement":null},{"id":"W4252513877","doi":"10.32920/ryerson.14664417.v1","title":"Probabilistic Program Execution is a Viable Way to Find Domains from Software","year":2021,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"","keywords":"Computer science; Probabilistic logic; Software; Set (abstract data type); Program comprehension; Probabilistic analysis of algorithms; Algorithm; Theoretical computer science; Software system; Programming language; Artificial intelligence","score_opus":0.02626209323586308,"score_gpt":0.305504360580541,"score_spread":0.2792422673446779,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4252513877","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11233107,0.00027489098,0.8815434,0.0024281954,0.00036598512,0.0015480538,0.000054025626,0.0010084443,0.00044592223],"genre_scores_gemma":[0.1521882,0.00008536104,0.8414651,0.0009896772,0.00023684383,0.0014014702,0.00033268725,0.00004064428,0.0032600188],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.995049,0.0002751205,0.0005893321,0.0021352912,0.0012291276,0.0007221167],"domain_scores_gemma":[0.9953221,0.0005030154,0.00012479638,0.0028963992,0.0007149428,0.00043874764],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0006606952,0.0004294222,0.0006860899,0.00026039314,0.00021875612,0.0018740749,0.0024642716,0.0004694515,0.0007216643],"category_scores_gemma":[0.0011770028,0.00038519636,0.00046950983,0.0012341033,0.000090420624,0.00028293816,0.0052940273,0.00087208813,0.00031365812],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004382404,0.0034652906,0.005645659,0.0012806639,0.0009349994,0.0002175455,0.007262376,0.018580137,0.0004030439,0.0029636326,0.01939273,0.9398101],"study_design_scores_gemma":[0.0014050646,0.0009465652,0.031552605,0.0022345996,0.00036773708,0.000021853963,0.00055477227,0.70120996,0.0070272526,0.19660524,0.053460646,0.0046136845],"about_ca_topic_score_codex":0.0027156947,"about_ca_topic_score_gemma":0.0002727956,"teacher_disagreement_score":0.9351964,"about_ca_system_score_codex":0.0003799638,"about_ca_system_score_gemma":0.0006748224,"threshold_uncertainty_score":0.99986},"labels":[],"label_agreement":null},{"id":"W4254342849","doi":"10.1515/iupac.68.0042","title":"Collision Efficiency","year":2016,"lang":"en","type":"dataset","venue":"IUPAC Standards Online","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Glossary; Terminology; Field (mathematics); Computer science; Linguistics; Philosophy; Mathematics","score_opus":0.015344527016647475,"score_gpt":0.4131786912496758,"score_spread":0.39783416423302836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4254342849","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00003350464,0.00059125206,0.101255044,0.0010471868,0.00046891053,0.00021023427,0.8962217,0.00014960802,0.000022589811],"genre_scores_gemma":[0.000022611881,0.0014335199,0.0006229325,0.00017809635,0.0003998033,0.000014363258,0.9969404,0.000015713531,0.0003725823],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9947586,0.00024031439,0.0005496775,0.0010171199,0.0027825248,0.0006517766],"domain_scores_gemma":[0.9957107,0.00050635525,0.00020752453,0.0023353435,0.0009543983,0.00028570797],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0016504544,0.00038631307,0.0006457695,0.00047460117,0.00024802316,0.0003200259,0.003008274,0.00038231307,0.00063661503],"category_scores_gemma":[0.0019798945,0.00026448286,0.00032131778,0.0010707165,0.00019679527,0.00026332235,0.001090888,0.00057267566,0.000018542829],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000016014508,0.00025063573,0.0000036021286,0.00006952691,0.000037886366,0.00006325342,0.000007479327,0.000013851729,0.0000028439094,0.000062203755,0.98784435,0.011628325],"study_design_scores_gemma":[0.00047408842,0.00019813246,0.000007473243,0.00022639945,0.00002374065,0.000009755808,0.0000029128028,0.0008035925,0.000033011198,0.0005732722,0.9973004,0.00034723757],"about_ca_topic_score_codex":0.00013069871,"about_ca_topic_score_gemma":0.00012551642,"teacher_disagreement_score":0.10071871,"about_ca_system_score_codex":0.0007200297,"about_ca_system_score_gemma":0.0020308292,"threshold_uncertainty_score":0.99998075},"labels":[],"label_agreement":null},{"id":"W4281479810","doi":"10.4414/saez.2022.20803","title":"E-Rettungswagen voraussichtlich bereits 2023 im Einsatz","year":2022,"lang":"de","type":"article","venue":"Schweizerische Ärztezeitung","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"World Federation of Science Journalists","funders":"","keywords":"Chemistry","score_opus":0.022928764173301457,"score_gpt":0.2771925968441741,"score_spread":0.2542638326708726,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4281479810","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4391672,0.23981892,0.125058,0.07364163,0.035587143,0.008706537,0.0022269813,0.0049585593,0.070835054],"genre_scores_gemma":[0.94957656,0.0014848419,0.009104279,0.0018674055,0.0025457244,0.0007292583,0.00029499925,0.0002283297,0.0341686],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.985931,0.0027546145,0.0016370446,0.002840173,0.004253315,0.0025838553],"domain_scores_gemma":[0.9931391,0.0009183063,0.00060780585,0.0037530845,0.0005128462,0.0010688455],"candidate_categories":["metaepi_narrow","sts","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["open_science","insufficient_payload"],"category_scores_codex":[0.005583811,0.0009959831,0.001358,0.0010715126,0.003888601,0.0014823527,0.006109845,0.00036760716,0.009606279],"category_scores_gemma":[0.00097248005,0.001045848,0.0011276361,0.006364539,0.00057475985,0.0011463995,0.008164824,0.0035319817,0.006921464],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0011227285,0.007499815,0.04260037,0.0017570724,0.009329997,0.0058186664,0.018712493,0.027626278,0.018370925,0.06963136,0.42298606,0.37454423],"study_design_scores_gemma":[0.0023359242,0.0010543624,0.0065861004,0.0001291392,0.0006598047,0.00011945188,0.001277986,0.10886732,0.00198978,0.0031351664,0.8712098,0.0026351481],"about_ca_topic_score_codex":0.0009063484,"about_ca_topic_score_gemma":0.00003717097,"teacher_disagreement_score":0.51040936,"about_ca_system_score_codex":0.0006161223,"about_ca_system_score_gemma":0.0015733769,"threshold_uncertainty_score":0.99985695},"labels":[],"label_agreement":null},{"id":"W4286488146","doi":"10.1145/3550270","title":"Parameter Coverage for Testing of Autonomous Driving Systems under Uncertainty","year":2022,"lang":"en","type":"article","venue":"ACM Transactions on Software Engineering and Methodology","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":33,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Japan Society for the Promotion of Science; JST-Mirai Program; European Regional Development Fund; Japan Science and Technology Agency; University of Waterloo; Exploratory Research for Advanced Technology; Science Foundation Ireland","keywords":"Computer science; Correctness; Heuristics; Trustworthiness; Process (computing); Cover (algebra); Operations research; Reliability engineering; Mathematical optimization; Risk analysis (engineering); Algorithm; Computer security","score_opus":0.10734772623091261,"score_gpt":0.3267423154097948,"score_spread":0.21939458917888222,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4286488146","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.047853246,0.00017561267,0.9510717,0.00010569228,0.00036723408,0.00018923823,0.000019913878,0.00021616883,0.0000012189192],"genre_scores_gemma":[0.4478139,0.000011368639,0.5519018,0.000031937805,0.000012545471,0.0001633597,0.000002100307,0.000010981441,0.000051997653],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9985334,0.00034369595,0.0002794736,0.00037564998,0.00018400639,0.00028372632],"domain_scores_gemma":[0.98333883,0.01590977,0.00006571726,0.0005229283,0.00008742582,0.00007530549],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016103655,0.00013078765,0.00032977303,0.0002630632,0.00026543846,0.00003567729,0.00047375078,0.00006436862,0.000008000967],"category_scores_gemma":[0.0029137349,0.00013388325,0.0001142335,0.0004825775,0.000036562713,0.000088345856,0.00005402951,0.000313057,4.2202475e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000009086564,0.0000411814,0.0002271737,0.00007705827,0.00007115202,0.0000015448672,0.00016700807,0.9334105,0.00046621225,0.00074416667,0.0000039006786,0.06478102],"study_design_scores_gemma":[0.0006212847,0.0006940136,0.0009816682,0.000040280964,0.000064758555,0.00010126102,0.00016523007,0.98951334,0.0010351263,0.0035688258,0.0028710295,0.00034319537],"about_ca_topic_score_codex":0.000105493404,"about_ca_topic_score_gemma":0.0000016576896,"teacher_disagreement_score":0.39996067,"about_ca_system_score_codex":0.000086542015,"about_ca_system_score_gemma":0.00007531814,"threshold_uncertainty_score":0.54596007},"labels":[],"label_agreement":null},{"id":"W4286717422","doi":"10.4028/p-u6g3p6","title":"Failure Prediction of Automotive Components Utilizing a Path Independent Forming Limit Criterion","year":2022,"lang":"en","type":"article","venue":"Key engineering materials","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Volvo (Canada)","funders":"Blekinge Tekniska Högskola; VINNOVA","keywords":"Automotive industry; Limit (mathematics); Transformation (genetics); Path (computing); Fuel efficiency; Component (thermodynamics); Production (economics); Automotive engineering; Computer science; Engineering; Reliability engineering; Mathematics; Mathematical analysis","score_opus":0.019820243207558083,"score_gpt":0.2267748636339353,"score_spread":0.2069546204263772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4286717422","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8808836,0.000045082634,0.117937826,0.00011207008,0.00050526205,0.00018335905,0.00006817308,0.00024859782,0.000016077018],"genre_scores_gemma":[0.99268645,0.0000087697945,0.007128329,0.000012583222,0.000042412223,0.00006395676,0.00002999544,0.000012891866,0.000014611664],"study_design_codex":"bench_or_experimental","study_design_gemma":"bench_or_experimental","domain_scores_codex":[0.99834025,0.00011941975,0.0003782969,0.00030486286,0.0005803748,0.00027679588],"domain_scores_gemma":[0.999285,0.00009271605,0.0000995786,0.00036680466,0.0000917951,0.0000641178],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00089320494,0.00012914375,0.0002642955,0.00024123445,0.00014561227,0.00010886373,0.0005568888,0.000044632772,0.00010586533],"category_scores_gemma":[0.00015024316,0.00013561924,0.00007517611,0.00037190854,0.000013770588,0.0003285973,0.0006865286,0.00014897864,0.000005086294],"study_design_candidate":"bench_or_experimental","study_design_consensus":"bench_or_experimental","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000015192322,0.00012080947,0.00059264747,0.0002307119,0.0000724221,0.000019238407,0.0016798283,0.023146074,0.9714889,0.0015033166,0.00008417595,0.0010467031],"study_design_scores_gemma":[0.0012461848,0.0005781584,0.04221768,0.00033364721,0.000053031723,0.000102004225,0.0005086685,0.41058967,0.5380396,0.00069265545,0.0049893013,0.00064936833],"about_ca_topic_score_codex":0.00006489848,"about_ca_topic_score_gemma":3.8447257e-7,"teacher_disagreement_score":0.43344924,"about_ca_system_score_codex":0.00014564993,"about_ca_system_score_gemma":0.000034683144,"threshold_uncertainty_score":0.55303925},"labels":[],"label_agreement":null},{"id":"W4287218407","doi":"10.5539/cis.v15n3p47","title":"Homogenous Multiple Classifier System for Software Quality Assessment Based on Support Vector Machine","year":2022,"lang":"en","type":"article","venue":"Computer and Information Science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"Tertiary Education Trust Fund","keywords":"Support vector machine; Computer science; Quadratic classifier; Machine learning; Artificial intelligence; Software; AdaBoost; Classifier (UML); Linear discriminant analysis; Software quality; Data mining; Margin classifier; Confusion matrix; Random forest; Pattern recognition (psychology); Software development","score_opus":0.03609607299771823,"score_gpt":0.31457991063425106,"score_spread":0.27848383763653284,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4287218407","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006179289,0.0000061525325,0.99193037,0.00051615114,0.000424283,0.0003862309,0.00006655808,0.00018321678,0.00030775703],"genre_scores_gemma":[0.9170816,0.0000014151648,0.08169329,0.001019237,0.000026253834,0.00012335884,0.000039160655,0.0000028027773,0.000012870078],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9975651,0.00012111774,0.0004239932,0.00035659166,0.0011917178,0.00034148202],"domain_scores_gemma":[0.9982102,0.0005343203,0.00015306813,0.000595034,0.00034018228,0.00016720344],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.0032748622,0.00012193375,0.00019098085,0.0004096391,0.0014126346,0.000630418,0.0011390235,0.000024285668,0.000017111064],"category_scores_gemma":[0.00015485857,0.00010612135,0.00008799626,0.0010932054,0.00014233492,0.0025017173,0.00062944595,0.00017651978,0.000011226674],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011281658,0.0005348315,0.0293537,0.00076912955,0.00003699966,0.000006257676,0.0033447545,0.10831348,0.00021556819,0.0838155,0.0021703572,0.7713266],"study_design_scores_gemma":[0.0005220093,0.00035694853,0.019813301,0.000006761423,0.0000022261493,0.0000072885796,0.0000664541,0.96801615,0.00011301421,0.000053504802,0.010898673,0.00014364846],"about_ca_topic_score_codex":0.000024337176,"about_ca_topic_score_gemma":0.0000010046375,"teacher_disagreement_score":0.9109023,"about_ca_system_score_codex":0.00031735792,"about_ca_system_score_gemma":0.0005469031,"threshold_uncertainty_score":0.9998874},"labels":[],"label_agreement":null},{"id":"W4289438365","doi":"10.48550/arxiv.1810.01577","title":"Moment-Sum-Of-Squares Approach For Fast Risk Estimation In Uncertain\\n Environments","year":2018,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Probabilistic logic; Bounded function; Mathematics; Moment (physics); Chebyshev filter; Mathematical optimization; Explained sum of squares; Probability distribution; Chebyshev nodes; Polynomial; Applied mathematics; Statistics; Mathematical analysis","score_opus":0.06879104267056715,"score_gpt":0.22613713485952425,"score_spread":0.1573460921889571,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4289438365","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.1516757,0.000029613624,0.8475025,0.000016523685,0.000058757436,0.0004493504,0.00003967459,0.000037900238,0.0001899736],"genre_scores_gemma":[0.959333,0.00019694504,0.040005196,0.000006805092,0.000021535041,0.000007424793,0.00005078809,0.000010471032,0.00036784995],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9979213,0.00023015923,0.00028524382,0.0010400068,0.00018684832,0.00033647908],"domain_scores_gemma":[0.99815387,0.00024483976,0.0003023174,0.001108362,0.00008624683,0.000104344916],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00071584655,0.00022598219,0.00036933113,0.00043747298,0.00011329939,0.000065270855,0.0015702982,0.00024892954,0.000014676026],"category_scores_gemma":[0.00012982643,0.0002497092,0.0002444001,0.0006159866,0.00021099419,0.00027212454,0.0012357295,0.00034425542,0.000013769952],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039080773,0.00031662863,0.023569021,0.00019503821,0.00010851442,0.000006902845,0.0003281613,0.9654557,0.00001917581,0.0056116274,0.000056537712,0.0042936206],"study_design_scores_gemma":[0.0004207546,0.000077490105,0.0018442422,0.00004399938,0.000043672106,2.3643418e-7,0.000065352186,0.9726654,0.00034572207,0.024175856,0.000097038734,0.00022020617],"about_ca_topic_score_codex":0.00036919225,"about_ca_topic_score_gemma":0.000015047321,"teacher_disagreement_score":0.8076573,"about_ca_system_score_codex":0.00032579055,"about_ca_system_score_gemma":0.00013121235,"threshold_uncertainty_score":0.9999955},"labels":[],"label_agreement":null},{"id":"W4293098027","doi":"10.1145/3478432.3499102","title":"Universal Design of Interactive Mathematical Notebooks on Programming","year":2022,"lang":"en","type":"article","venue":"Proceedings of the 53rd ACM Technical Symposium on Computer Science Education V. 2","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"","keywords":"Computer science; Mathematical proof; Compiler; Programming language; Software engineering; Graphics; USable; Multimedia; Computer graphics (images)","score_opus":0.021648178604182434,"score_gpt":0.29136559805032547,"score_spread":0.269717419446143,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4293098027","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2299559,0.000016112825,0.73593843,0.025752615,0.0015295482,0.002706612,0.000004570837,0.0005007128,0.0035954795],"genre_scores_gemma":[0.8578522,0.000001708938,0.14173444,0.00020466697,0.000046316924,0.000075117845,2.156602e-7,0.0000071462596,0.000078146346],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99689776,0.0000613644,0.00042051027,0.0006603752,0.0016167464,0.000343247],"domain_scores_gemma":[0.99750113,0.000635143,0.0003652347,0.0007566194,0.0006126911,0.00012920059],"candidate_categories":["open_science"],"consensus_categories":[],"category_scores_codex":[0.0020948693,0.00016000798,0.00025456224,0.00045214762,0.00051217945,0.00019071678,0.0064753443,0.00004449249,0.000013464342],"category_scores_gemma":[0.00077065075,0.000119437784,0.00016179222,0.0023675559,0.0006356072,0.00048125122,0.0032054738,0.00049214135,0.0000060716043],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00021347713,0.0068241246,0.0015018054,0.00022394108,0.0000610509,0.0000010325858,0.0047336523,0.007624846,0.15093175,0.6250284,0.002677853,0.20017809],"study_design_scores_gemma":[0.0010952931,0.0072349464,0.0044502793,0.0010062247,0.000083362844,0.00013974559,0.0011987941,0.45132035,0.3803066,0.14821897,0.0036661525,0.0012792714],"about_ca_topic_score_codex":0.000010158295,"about_ca_topic_score_gemma":7.838464e-8,"teacher_disagreement_score":0.62789637,"about_ca_system_score_codex":0.00042475673,"about_ca_system_score_gemma":0.0008421152,"threshold_uncertainty_score":0.9989001},"labels":[],"label_agreement":null},{"id":"W4293370349","doi":"10.21428/594757db.b85e6625","title":"Low Level Source Code Vulnerability Detection Using Advanced BERT Language Model","year":2022,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":10,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ontario Institute of Technology","funders":"","keywords":"Computer science; Code (set theory); Vulnerability (computing); Programming language; Computer security","score_opus":0.04765795215647902,"score_gpt":0.32095578367162914,"score_spread":0.27329783151515014,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4293370349","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.42190385,0.00002169704,0.5775728,0.00012986946,0.000036850754,0.00008003741,0.0000048853994,0.00014883584,0.00010114993],"genre_scores_gemma":[0.96122724,0.0000017049873,0.037379533,0.00018360333,0.000016771453,0.00003152287,0.000002581752,0.000008421487,0.0011486476],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99806815,0.00024759842,0.00021491118,0.000526125,0.00062850845,0.0003147155],"domain_scores_gemma":[0.99885136,0.00017210977,0.00004858989,0.00074630545,0.000086043474,0.00009557115],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00094505196,0.00010955644,0.0001629576,0.00012488008,0.0006086295,0.0000910714,0.0007387624,0.000033104137,0.00011896479],"category_scores_gemma":[0.00019002413,0.000105612904,0.00012622353,0.00077836344,0.000044357224,0.00034868086,0.0007721233,0.00034001196,0.000010653158],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012402748,0.00014080612,0.00028908998,0.000014440111,0.000012285473,0.00000304059,0.000718743,0.7426342,0.02262778,0.00016660367,0.000014199589,0.23336644],"study_design_scores_gemma":[0.00017598642,0.00003734721,0.00012246697,0.0000018236951,0.0000036639947,0.0000080001355,0.00021732977,0.9821252,0.015983256,0.001007508,0.00018440268,0.00013297827],"about_ca_topic_score_codex":0.00039995756,"about_ca_topic_score_gemma":0.00014419883,"teacher_disagreement_score":0.5401933,"about_ca_system_score_codex":0.00028722605,"about_ca_system_score_gemma":0.00011775333,"threshold_uncertainty_score":0.46811476},"labels":[],"label_agreement":null},{"id":"W4295267415","doi":"","title":"Scenario Based Analysis Procedure Guide V4.","year":2020,"lang":"en","type":"paratext","venue":"OSTI OAI (U.S. Department of Energy Office of Scientific and Technical Information)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Safran Electronics (Canada)","funders":"","keywords":"Computer science","score_opus":0.015496463385359113,"score_gpt":0.26378741758580043,"score_spread":0.2482909542004413,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4295267415","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0019310747,0.0010790824,0.83682084,0.0024440791,0.0007241282,0.0006730442,0.0008307701,0.00024798515,0.15524901],"genre_scores_gemma":[0.939006,0.00037049476,0.037271846,0.00058423646,0.00007036656,0.00011889287,0.0135753695,0.000021347721,0.008981471],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9956349,0.00009965863,0.0014628212,0.0007195904,0.0016554827,0.00042752572],"domain_scores_gemma":[0.9962843,0.00036884265,0.00088425505,0.0011366745,0.001001289,0.0003246357],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009849495,0.00037504177,0.0009689055,0.0013061005,0.00028699174,0.0004982014,0.0016881487,0.0003748063,0.0003924903],"category_scores_gemma":[0.00049694994,0.0003254064,0.0005637012,0.004696449,0.00072191213,0.0010798057,0.0007054942,0.00033554077,0.00023099853],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002217946,0.0014832792,0.0020178198,0.0028294453,0.0025877766,0.000010704255,0.000031322106,0.02742688,0.00044173357,0.038334545,0.9103388,0.014275904],"study_design_scores_gemma":[0.0007547031,0.00043907738,0.006916764,0.00028091276,0.00077245425,0.0000071106406,0.0000051385714,0.01671102,0.0036437237,0.00021264459,0.9693942,0.0008622548],"about_ca_topic_score_codex":0.000088185836,"about_ca_topic_score_gemma":0.000039402767,"teacher_disagreement_score":0.9370749,"about_ca_system_score_codex":0.00009918218,"about_ca_system_score_gemma":0.0009943454,"threshold_uncertainty_score":0.9999198},"labels":[],"label_agreement":null},{"id":"W4298445647","doi":"10.1142/5760","title":"Computational Intelligence in Software Quality Assurance","year":2005,"lang":"en","type":"book","venue":"Series in machine perception and artificial intelligence","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":9,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Computer science; Quality assurance; Software quality; Software quality assurance; Quality (philosophy); Software; Software engineering; Engineering; Software development; Operations management; Operating system; Physics","score_opus":0.04175911062487599,"score_gpt":0.35027692794193893,"score_spread":0.30851781731706296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4298445647","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00079821533,0.00072901195,0.9906786,0.0019270922,0.00024629507,0.00041546914,0.000045066878,0.0001762322,0.0049840286],"genre_scores_gemma":[0.40664697,0.013037227,0.45383143,0.0019610915,0.0012038145,0.0002779705,0.0007210711,0.00017490385,0.122145504],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9958845,0.00038129455,0.001336066,0.0011305166,0.00077242684,0.00049523276],"domain_scores_gemma":[0.99798954,0.000801545,0.00022112473,0.0006160013,0.00022254192,0.00014923794],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0016788744,0.00043252853,0.0006646056,0.00065077236,0.00017975956,0.00034436432,0.0010803862,0.0003566866,0.0005146151],"category_scores_gemma":[0.00070942024,0.00044626274,0.00016356645,0.00073759863,0.0005530011,0.0007910121,0.00048588493,0.0011230921,0.00023170744],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000055616816,0.00018345557,0.0015371586,0.00018965255,0.000013553074,0.000023482955,0.00184669,0.044041786,0.000012574057,0.09997348,0.00013887821,0.85198367],"study_design_scores_gemma":[0.000056539175,0.0001516965,0.0079393685,0.00047635136,0.000008495284,0.00003502098,0.0004143345,0.3310557,0.000037332553,0.65156156,0.0071999677,0.001063633],"about_ca_topic_score_codex":0.0003101127,"about_ca_topic_score_gemma":0.00431025,"teacher_disagreement_score":0.85092,"about_ca_system_score_codex":0.00040436725,"about_ca_system_score_gemma":0.0004173583,"threshold_uncertainty_score":0.9997989},"labels":[],"label_agreement":null},{"id":"W4300945743","doi":"10.48550/arxiv.1409.3736","title":"A Linear Programming Approach to Error Bounds for Random Walks in the\\n Quarter-plane","year":2014,"lang":"","type":"preprint","venue":"arXiv (Cornell University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Random walk; Approximation error; Markov chain; Mathematics; Plane (geometry); Work (physics); Space (punctuation); Product (mathematics); State space; Markov process; Quarter (Canadian coin); Distribution (mathematics); Stationary distribution; Linear programming; Statistical physics; Applied mathematics; Mathematical analysis; Mathematical optimization; Computer science; Statistics; Physics; Geometry","score_opus":0.0925070305915379,"score_gpt":0.24049137503301804,"score_spread":0.14798434444148012,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4300945743","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11318244,0.00005540975,0.8816853,0.00067755184,0.000272548,0.0032605138,0.000020729976,0.00012261597,0.0007228663],"genre_scores_gemma":[0.9843615,0.00006421156,0.014038507,0.00023515732,0.000269358,0.00007429515,0.00007211638,0.000033632525,0.00085118366],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9926435,0.0014671359,0.00080324087,0.0030661244,0.00051309983,0.0015069479],"domain_scores_gemma":[0.993498,0.0020858035,0.00037267068,0.0029691376,0.0005689912,0.0005053588],"candidate_categories":["metaepi_narrow","open_science"],"consensus_categories":[],"category_scores_codex":[0.0057413597,0.0007465991,0.00124441,0.00096076203,0.0007251446,0.0008352323,0.0060705193,0.0006280607,0.000018274157],"category_scores_gemma":[0.0006570357,0.00068286783,0.0010070517,0.0032149507,0.00041149952,0.00042243916,0.0016038658,0.0015362191,0.00014062236],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0024870576,0.004114988,0.008376287,0.0022754662,0.0007678989,0.00025476472,0.015363438,0.8846875,0.000025336396,0.048183557,0.0008565419,0.032607153],"study_design_scores_gemma":[0.003275873,0.0005238215,0.00054132426,0.00019602811,0.00016826238,0.000012456776,0.0013027347,0.97404134,0.000010404717,0.0064040856,0.0126840845,0.0008395857],"about_ca_topic_score_codex":0.0008476643,"about_ca_topic_score_gemma":0.0002550949,"teacher_disagreement_score":0.8711791,"about_ca_system_score_codex":0.0003927724,"about_ca_system_score_gemma":0.0005187796,"threshold_uncertainty_score":0.99956226},"labels":[],"label_agreement":null},{"id":"W4317000347","doi":"10.1145/3573074.3573080","title":"The Java Pathfinder Workshop 2022","year":2023,"lang":"en","type":"article","venue":"ACM SIGSOFT Software Engineering Notes","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University","funders":"","keywords":"Computer science; Programming language; Java; Bytecode; Symbolic execution; Java bytecode; Model checking; Program analysis; Java Modeling Language; Pathfinder; Java annotation; Software engineering; Software; Real time Java; World Wide Web","score_opus":0.023949993973193247,"score_gpt":0.2714183101211945,"score_spread":0.24746831614800124,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4317000347","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20422222,0.003576881,0.7676269,0.012421445,0.00212716,0.0005944969,0.000018449777,0.009402543,0.000009901505],"genre_scores_gemma":[0.97951293,0.00048656037,0.018659808,0.00015091294,0.00023676334,0.00013058915,0.000019768579,0.00005363194,0.0007490456],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9974196,0.00007037934,0.00032974064,0.00057602924,0.0007659416,0.00083833904],"domain_scores_gemma":[0.86200726,0.13501321,0.00006658223,0.0024889586,0.00019189049,0.00023207246],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0011969745,0.0002527951,0.00023739248,0.00025005633,0.0005019355,0.00046850473,0.00292751,0.0001338077,0.000026371554],"category_scores_gemma":[0.21386069,0.0001927926,0.00021358512,0.0026922524,0.00006855947,0.0003503736,0.0014273629,0.0005004752,0.0005851457],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002099851,0.00015966939,0.3547589,0.00025815502,0.0004373842,0.00028575488,0.002225194,0.028583601,0.0024287847,0.010670736,0.027260922,0.5729099],"study_design_scores_gemma":[0.0013437232,0.00023761023,0.60590124,0.00037996465,0.00007814713,0.000074364994,0.00024729015,0.18005654,0.005622025,0.0064617367,0.19652712,0.0030702518],"about_ca_topic_score_codex":0.000020588855,"about_ca_topic_score_gemma":0.000007177954,"teacher_disagreement_score":0.7752907,"about_ca_system_score_codex":0.00007000548,"about_ca_system_score_gemma":0.00009416448,"threshold_uncertainty_score":0.7927613},"labels":[],"label_agreement":null},{"id":"W4318220122","doi":"10.1007/978-3-031-11085-6_7","title":"Reliability and Quality Assurance of M&S","year":2023,"lang":"en","type":"book-chapter","venue":"Simulation foundations, methods and applications","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Section (typography); Reliability (semiconductor); Reliability engineering; Quality assurance; Computer science; Engineering; Physics; Operations management; Thermodynamics; Operating system","score_opus":0.14281700986293003,"score_gpt":0.4720701631890869,"score_spread":0.3292531533261569,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4318220122","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000057092297,0.00034322593,0.98571175,0.0004524945,0.000041684023,0.0006098817,0.00005645885,0.00017187906,0.012555524],"genre_scores_gemma":[0.012080021,0.0019745706,0.88766026,0.0000764575,0.00014458183,0.00026949117,0.0002499665,0.000050494837,0.097494125],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9975877,0.0002796069,0.00080477627,0.00077042676,0.00039780268,0.00015971657],"domain_scores_gemma":[0.9916814,0.00593964,0.00043091207,0.001176514,0.00065407663,0.00011743907],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003106509,0.00021486446,0.0004677888,0.000264799,0.0003507572,0.0001542972,0.00036793126,0.0002354142,0.000044321903],"category_scores_gemma":[0.0011334391,0.00021501504,0.00013558853,0.00035280545,0.0003381949,0.0002811546,0.00029400946,0.00027619427,0.00004222635],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000026764073,0.000023838762,0.00021714107,0.00019184941,0.00004192391,3.870346e-8,0.00006186013,0.0026697784,0.000063283354,0.76095766,0.000024964327,0.23574498],"study_design_scores_gemma":[0.0001529098,0.000016285023,0.0037406706,0.000051556843,0.000051145933,6.198038e-7,0.000007934447,0.04144681,0.000019860845,0.6405546,0.3136669,0.00029070728],"about_ca_topic_score_codex":0.000041814506,"about_ca_topic_score_gemma":0.00002706276,"teacher_disagreement_score":0.31364194,"about_ca_system_score_codex":0.000046335525,"about_ca_system_score_gemma":0.00014043532,"threshold_uncertainty_score":0.8768059},"labels":[],"label_agreement":null},{"id":"W4322717938","doi":"10.1145/3624742","title":"Stress Testing Control Loops in Cyber-physical Systems","year":2023,"lang":"en","type":"article","venue":"ACM Transactions on Software Engineering and Methodology","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Natural Sciences and Engineering Research Council of Canada; Knut och Alice Wallenbergs Stiftelse; Lunds Universitet; European Commission","keywords":"Leverage (statistics); Computer science; Cyber-physical system; Software; Stress testing (software); Control (management); Control system; Control engineering; Systems engineering; Engineering; Artificial intelligence","score_opus":0.10009437022392521,"score_gpt":0.3358311838909224,"score_spread":0.23573681366699722,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4322717938","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.101771615,0.00009513474,0.89693165,0.0001809663,0.00025154228,0.00011743543,0.000010036407,0.0006388652,0.0000027418662],"genre_scores_gemma":[0.7863253,0.000043412172,0.21337035,0.000025509748,0.00005256109,0.00010006243,0.000002091106,0.000016694737,0.00006406198],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99835205,0.00037484383,0.00022370981,0.0004389648,0.00019804141,0.00041240363],"domain_scores_gemma":[0.9878367,0.011419573,0.000027414915,0.0005344105,0.00006667048,0.00011519821],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012278004,0.00015540652,0.0003622176,0.0004962876,0.00009993202,0.00006444022,0.00045776353,0.000111377485,0.0000016236617],"category_scores_gemma":[0.0036574192,0.00014842462,0.00007371672,0.0014382623,0.000041405587,0.00014072342,0.000026170486,0.00042981398,0.000020287252],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010806136,0.00008388864,0.0036418024,0.00012754613,0.00007250603,0.000051887957,0.0005379163,0.85163724,0.0006825081,0.0007784195,0.000007522193,0.14236796],"study_design_scores_gemma":[0.0007655548,0.00019332719,0.017492691,0.00012307477,0.000028392413,0.000043580094,0.00012294063,0.9788823,0.0007459168,0.0009903074,0.00028097304,0.0003309411],"about_ca_topic_score_codex":0.00016456771,"about_ca_topic_score_gemma":0.0000069184584,"teacher_disagreement_score":0.6845536,"about_ca_system_score_codex":0.000040118353,"about_ca_system_score_gemma":0.000038279715,"threshold_uncertainty_score":0.60525805},"labels":[],"label_agreement":null},{"id":"W4362647280","doi":"10.1109/rams51473.2023.10088275","title":"Integrating Reliability Engineering with Model Based Systems Engineering","year":2023,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Lockheed Martin (Canada)","funders":"","keywords":"Computer science; Scripting language; Table (database); Reliability (semiconductor); MATLAB; Reliability engineering; Software engineering; Programming language; Data mining; Engineering","score_opus":0.013541026391122547,"score_gpt":0.23379596813023742,"score_spread":0.2202549417391149,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4362647280","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.038297955,0.000015511296,0.95946825,0.00032750223,0.00006127099,0.0001395419,0.0000011332185,0.001527806,0.00016104599],"genre_scores_gemma":[0.87175804,0.0000024689755,0.12793149,0.00001780815,0.00002222804,0.00005506443,0.0000031058821,0.000013644285,0.00019614985],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99834687,0.000030717296,0.0002398294,0.0004563516,0.00050615065,0.00042008428],"domain_scores_gemma":[0.998442,0.000507574,0.000026648499,0.0007313422,0.00015605337,0.00013638854],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010227655,0.00015466368,0.00021029233,0.0002783745,0.00007642831,0.00026067146,0.0006454851,0.000058139478,0.0000040927175],"category_scores_gemma":[0.0005094465,0.00011411284,0.00007716969,0.0018380646,0.000016326789,0.00034744074,0.00016366343,0.00023994574,0.000046412828],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000012717823,0.000012584608,0.00074262085,0.00009125784,0.000009874753,0.0000065640147,0.00005584862,0.99217355,0.00064289995,0.0056738127,0.000053823747,0.0005358844],"study_design_scores_gemma":[0.00010406444,0.000026913802,0.00025924697,0.00005973162,0.0000029379833,0.0000017341359,0.000020045389,0.99845403,0.00073025154,0.000042600004,0.00014342458,0.00015503468],"about_ca_topic_score_codex":0.00009621485,"about_ca_topic_score_gemma":0.0000032737723,"teacher_disagreement_score":0.8334601,"about_ca_system_score_codex":0.00009239982,"about_ca_system_score_gemma":0.00011544146,"threshold_uncertainty_score":0.46533865},"labels":[],"label_agreement":null},{"id":"W4368340833","doi":"10.2172/1972169","title":"MCNP® Site Support Newsletter First Quarter 2023","year":2023,"lang":"en","type":"report","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"Los Alamos National Laboratory; National Nuclear Security Administration; U.S. Department of Energy","keywords":"Benchmark (surveying); Computer science; Code (set theory); Quarter (Canadian coin); Programming language; Archaeology; Geology; Geography","score_opus":0.08141183429092612,"score_gpt":0.3524968970234773,"score_spread":0.2710850627325512,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4368340833","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00043775217,0.0005250038,0.44446906,0.06637371,0.009806629,0.0015602802,0.00020439891,0.0050622877,0.47156087],"genre_scores_gemma":[0.0030065372,0.003477508,0.0121628465,0.0030565413,0.0035817726,0.00034059776,0.0007832561,0.00020601346,0.9733849],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9933784,0.00010331449,0.0008578349,0.0015346326,0.0031459194,0.0009798838],"domain_scores_gemma":[0.99524117,0.00072507403,0.00021443448,0.0025910302,0.0008965588,0.00033176228],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0023752407,0.00048011966,0.00083146326,0.00073228637,0.00024797762,0.0008214398,0.0023340464,0.0005682269,0.0027812705],"category_scores_gemma":[0.00067844125,0.0003933963,0.0007902122,0.0018191264,0.000109124885,0.00047576783,0.0012924074,0.00095897383,0.025590643],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000013186631,0.000047343634,0.0021007836,0.00019287485,0.0001377996,0.00029194137,0.0000907742,0.000019362767,0.0000037170203,0.00006704478,0.9884594,0.008587652],"study_design_scores_gemma":[0.00012389821,0.00008388133,0.0023354935,0.00007720905,0.000037281738,0.000047177946,0.000011047622,0.0036005606,0.000020275671,0.00037194713,0.99273413,0.00055711844],"about_ca_topic_score_codex":0.0025910526,"about_ca_topic_score_gemma":0.0013651712,"teacher_disagreement_score":0.5018241,"about_ca_system_score_codex":0.00034264685,"about_ca_system_score_gemma":0.0013709876,"threshold_uncertainty_score":0.99985176},"labels":[],"label_agreement":null},{"id":"W4382795996","doi":"10.1177/1748006x231184289","title":"A model-based safety analysis approach for airborne systems using state traversals","year":2023,"lang":"en","type":"article","venue":"Proceedings of the Institution of Mechanical Engineers Part O Journal of Risk and Reliability","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Manitoba","funders":"Aeronautical Science Foundation of China; National Natural Science Foundation of China","keywords":"Fault tree analysis; Computer science; Reliability engineering; State space; Model checking; Markov chain; Certification; Markov model; Reduction (mathematics); State (computer science); Simulation; Engineering; Algorithm; Machine learning; Mathematics","score_opus":0.024536700723139366,"score_gpt":0.2639670540791575,"score_spread":0.23943035335601812,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4382795996","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2900131,0.00007798464,0.7094197,0.00010657309,0.00011496313,0.00020598937,0.000035800753,0.000020697242,0.0000051629636],"genre_scores_gemma":[0.9532802,0.00024476866,0.046426963,0.0000041197495,0.000027412758,0.000004984283,0.0000013031514,0.000005053903,0.000005196214],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99777454,0.000054085496,0.0009547859,0.00027830031,0.0007030191,0.00023528739],"domain_scores_gemma":[0.99738204,0.00038044434,0.00068237976,0.00022560701,0.0011848805,0.00014467005],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0056406446,0.00014451402,0.0006400314,0.00038516225,0.00016096233,0.000045456247,0.000687769,0.00011426385,5.2090417e-7],"category_scores_gemma":[0.0021101,0.0000939592,0.00060589216,0.0020666355,0.00019946828,0.00031299485,0.00013053838,0.00029789648,8.324133e-8],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00011939045,0.00011407274,0.0007854684,0.00034426406,0.00020436365,1.4110972e-7,0.00014719438,0.9922257,0.00074542686,0.004622753,0.000021904598,0.00066929375],"study_design_scores_gemma":[0.0005277711,0.00009829485,0.00034577696,0.00008005248,0.00038418995,0.000002346522,0.00009719914,0.99206537,0.0035871165,0.002650702,0.0000684738,0.00009270248],"about_ca_topic_score_codex":0.000030889783,"about_ca_topic_score_gemma":3.662057e-7,"teacher_disagreement_score":0.6632671,"about_ca_system_score_codex":0.000099690275,"about_ca_system_score_gemma":0.00020963217,"threshold_uncertainty_score":0.3831545},"labels":[],"label_agreement":null},{"id":"W4385078510","doi":"10.18280/isi.280304","title":"Experimental Investigations to Fault Reduction System for Software Applications","year":2023,"lang":"en","type":"article","venue":"Ingénierie des systèmes d information","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Reduction (mathematics); Computer science; Software; Reliability engineering; Fault (geology); Engineering; Geology; Operating system; Seismology; Mathematics; Geometry","score_opus":0.028264115964766065,"score_gpt":0.2905615456978224,"score_spread":0.2622974297330563,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385078510","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.029798416,0.00002528124,0.9669397,0.0002843006,0.00019602942,0.0010986651,0.000032033233,0.0012974029,0.0003281797],"genre_scores_gemma":[0.91780514,0.0000033698134,0.078488305,0.0000755036,0.000101161386,0.003162484,0.00024205318,0.000009863563,0.000112149006],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99859816,0.000047318827,0.00046424865,0.00022115294,0.00035762662,0.0003114705],"domain_scores_gemma":[0.9986734,0.00013364869,0.00012049002,0.00046911888,0.00043739355,0.00016599082],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00056873146,0.00012336795,0.0001526715,0.00054977264,0.0006555661,0.00046420252,0.0005096457,0.00007902926,0.0000032475612],"category_scores_gemma":[0.00038158716,0.00012603712,0.0000970502,0.0021982063,0.000071181734,0.0024016448,0.00016184857,0.000073286996,0.0005852595],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005187733,0.0001657148,0.0011831631,0.002766768,0.0001989304,0.0000018807596,0.038620926,0.061389755,0.0109191295,0.24916217,0.019239228,0.61630046],"study_design_scores_gemma":[0.0018848871,0.0007321959,0.0045566475,0.0007542873,0.0000675215,0.00013724098,0.023843706,0.60744154,0.17906512,0.035997774,0.14363882,0.0018802736],"about_ca_topic_score_codex":0.00004018674,"about_ca_topic_score_gemma":0.0000013542732,"teacher_disagreement_score":0.8884514,"about_ca_system_score_codex":0.0004446859,"about_ca_system_score_gemma":0.00012925161,"threshold_uncertainty_score":0.7522522},"labels":[],"label_agreement":null},{"id":"W4385480910","doi":"10.1093/jrsssa/qnad090","title":"Statistical Methods for Mediation, Confounding and Moderation Analysis Using R and SAS","year":2023,"lang":"en","type":"article","venue":"Journal of the Royal Statistical Society Series A (Statistics in Society)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Prioris.ai (Canada)","funders":"","keywords":"Moderation; Mediation; Confounding; Statistical analysis; Statistics; Psychology; Mathematics; Sociology; Social science","score_opus":0.035637607176933525,"score_gpt":0.3953241508230078,"score_spread":0.3596865436460743,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385480910","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.006021086,0.0002003777,0.9923503,0.0008308029,0.00016991359,0.00018614753,0.00021643689,0.0000204462,0.000004511848],"genre_scores_gemma":[0.040731728,0.00041649744,0.95851076,0.0001447781,0.00008161997,0.000009305833,0.000020119809,0.000013316653,0.000071898074],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9974476,0.00042202318,0.00078718486,0.00033625565,0.0005996033,0.00040730616],"domain_scores_gemma":[0.9918202,0.006986356,0.00032133394,0.00021857079,0.00046341735,0.00019012694],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003454872,0.00018351304,0.0005558824,0.000088134046,0.00052365375,0.00042912524,0.00037662385,0.00013735429,0.000015900796],"category_scores_gemma":[0.0031320713,0.00014042176,0.00026206023,0.001338019,0.00045052712,0.00027463943,0.00031016118,0.0004177726,4.0582583e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00013315321,0.00021922828,0.045459203,0.0009050095,0.0034715817,0.000025266852,0.013730216,0.070775196,0.00041123637,0.75925195,0.014513238,0.09110474],"study_design_scores_gemma":[0.00036958052,0.000075502336,0.016115932,0.0000162514,0.0002755041,0.000006476938,0.0006610392,0.86057025,0.0000166992,0.12136148,0.00039290515,0.00013835712],"about_ca_topic_score_codex":0.00005224752,"about_ca_topic_score_gemma":0.000028742837,"teacher_disagreement_score":0.7897951,"about_ca_system_score_codex":0.00021721481,"about_ca_system_score_gemma":0.0002513235,"threshold_uncertainty_score":0.5726233},"labels":[],"label_agreement":null},{"id":"W4385624965","doi":"10.1109/newcas57931.2023.10198045","title":"Statistical Model Checking based Analysis of Fault Trees and Power Consumption to Enhance Autonomous Systems Reliability","year":2023,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Fault tree analysis; Reliability engineering; Redundancy (engineering); Computer science; Power consumption; Reliability (semiconductor); Component (thermodynamics); Model checking; Probability model; Statistical model; Power (physics); Engineering; Artificial intelligence; Statistics; Mathematics; Algorithm","score_opus":0.029494019488069945,"score_gpt":0.34331321006660714,"score_spread":0.3138191905785372,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4385624965","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.37127396,0.000017706921,0.6281071,0.00024518085,0.000018153507,0.00011155843,0.000021422595,0.00013413603,0.00007080053],"genre_scores_gemma":[0.96448183,0.000013640025,0.035247315,0.000044429875,0.000003217297,0.000026870599,0.000011983657,0.0000043457744,0.00016637653],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9979718,0.00016342736,0.00042367875,0.0006129427,0.0005387563,0.0002894179],"domain_scores_gemma":[0.9978361,0.0009799785,0.00006259677,0.0007041561,0.00024195242,0.00017518843],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014714383,0.00011685171,0.00041291298,0.00058793474,0.00008470014,0.00013505714,0.00038959662,0.0000673602,0.00003688293],"category_scores_gemma":[0.0006682544,0.000099423254,0.00010679443,0.0023415152,0.000103005725,0.00017036955,0.00021760954,0.0000994289,0.000033414424],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000021425265,0.00013371221,0.051495302,0.00013572506,0.00024940344,0.0000053775198,0.000516587,0.925479,0.0016898114,0.013551452,0.0002514812,0.006470712],"study_design_scores_gemma":[0.00005834289,0.0000400631,0.050184727,0.000012087331,0.000062221545,1.6474783e-7,0.000026174279,0.94866335,0.0004768711,0.00034593433,0.000024730829,0.00010530561],"about_ca_topic_score_codex":0.0003106017,"about_ca_topic_score_gemma":0.000076752614,"teacher_disagreement_score":0.5932079,"about_ca_system_score_codex":0.000070770926,"about_ca_system_score_gemma":0.00009770084,"threshold_uncertainty_score":0.40543628},"labels":[],"label_agreement":null},{"id":"W4386609685","doi":"10.1007/978-3-031-40923-3","title":"Computer Safety, Reliability, and Security","year":2023,"lang":"en","type":"book","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Austrian Institute of Technology; Österreichische Forschungsförderungsgesellschaft; Centre National de la Recherche Scientifique; KU Leuven; Mälardalens högskola; Politechnika Gdańska; Ames Research Center; Università degli Studi di Firenze; European Research Consortium for Informatics and Mathematics; Universidade de Lisboa; Bundesministerium für Verkehr, Innovation und Technologie; European Commission; University of Toronto; National Aeronautics and Space Administration","keywords":"Computer science; Reliability (semiconductor); Computer security; Reliability engineering; Engineering","score_opus":0.015329498636984667,"score_gpt":0.2790020788586862,"score_spread":0.26367258022170154,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386609685","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00029981553,0.00033276106,0.99386805,0.002505817,0.0013848516,0.0005007663,0.000012571623,0.0005260337,0.0005693294],"genre_scores_gemma":[0.08145567,0.0016508629,0.90252566,0.0045657596,0.0039911573,0.000088969515,0.00008469746,0.00022837566,0.005408825],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99280024,0.00025992846,0.0008205183,0.0029461836,0.0019945155,0.001178607],"domain_scores_gemma":[0.99326664,0.0031098705,0.00021662845,0.0025277934,0.0004764753,0.00040258587],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.004120884,0.0006291769,0.0009411347,0.0013162411,0.00048553705,0.0012173681,0.0048728925,0.00052289973,0.0000148595245],"category_scores_gemma":[0.000593982,0.0005645637,0.00023891333,0.0032022963,0.0017490921,0.00078268844,0.0066226847,0.0016497938,0.00014516905],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001198677,0.00011427235,0.00093204255,0.00035300382,0.000034064895,0.0001910015,0.0015417845,0.038562883,0.000008582994,0.0033694422,0.00092449645,0.9539564],"study_design_scores_gemma":[0.00022675058,0.00015052066,0.0006832096,0.0002711,0.00000782207,0.000036691195,1.3622103e-7,0.70848936,0.00005231302,0.28718692,0.002325047,0.0005701012],"about_ca_topic_score_codex":0.00007872117,"about_ca_topic_score_gemma":0.00014240276,"teacher_disagreement_score":0.95338637,"about_ca_system_score_codex":0.00072679453,"about_ca_system_score_gemma":0.001636938,"threshold_uncertainty_score":0.99981946},"labels":[],"label_agreement":null},{"id":"W4387331371","doi":"10.1109/tnsm.2023.3318406","title":"EV Charging Infrastructure Discovery to Contextualize Its Deployment Security","year":2023,"lang":"en","type":"article","venue":"IEEE Transactions on Network and Service Management","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":17,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Hydro-Québec; Concordia University","funders":"Concordia University","keywords":"Software deployment; Computer science; Computer security; Computer network; Telecommunications","score_opus":0.01603467153867251,"score_gpt":0.26332560933928556,"score_spread":0.24729093780061306,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387331371","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.042981815,0.000070519396,0.94818556,0.006733737,0.0004634923,0.0005468664,0.000009695099,0.0004316778,0.00057663437],"genre_scores_gemma":[0.99350107,0.0012354256,0.0012359796,0.003082947,0.000079064,0.00014483358,0.0000039737206,0.000016116986,0.0007006084],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99813735,0.00010344749,0.00023332343,0.000566924,0.000470757,0.00048819568],"domain_scores_gemma":[0.99906623,0.00013521696,0.000032516517,0.00051777833,0.000065315966,0.00018292916],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004397156,0.00018827831,0.00020646476,0.00024099695,0.00041058377,0.0003406351,0.0005426616,0.00005420305,0.000029097158],"category_scores_gemma":[0.0000020308314,0.00017283423,0.000087286346,0.0021743318,0.00001471548,0.0003450315,0.000034695342,0.0002076295,0.00016829393],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000731736,0.00016684656,0.00015327195,0.00041519964,0.00042344638,0.00006575316,0.0027645214,0.7419848,0.000056513076,0.007882629,0.004331757,0.24168213],"study_design_scores_gemma":[0.0014893005,0.00030123425,0.007094462,0.00041015155,0.00015426622,0.00000847762,0.0012603408,0.9156516,0.0008581986,0.0077384464,0.063999586,0.0010338873],"about_ca_topic_score_codex":0.000046865167,"about_ca_topic_score_gemma":0.00015126927,"teacher_disagreement_score":0.95051926,"about_ca_system_score_codex":0.00006684233,"about_ca_system_score_gemma":0.000018676046,"threshold_uncertainty_score":0.70479757},"labels":[],"label_agreement":null},{"id":"W4387571649","doi":"10.12927/hcpol.2023.27184","title":"Appendix 4: Program Integration Rating Tool","year":2023,"lang":"fr","type":"article","venue":"Healthcare policy","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science","score_opus":0.09879328276944219,"score_gpt":0.4364136695571837,"score_spread":0.33762038678774153,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4387571649","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.032969847,0.005572231,0.052889273,0.89951456,0.0019994846,0.002249209,0.00009137951,0.002866759,0.0018472868],"genre_scores_gemma":[0.9069785,0.006275287,0.034749735,0.0061612846,0.007207088,0.00085647433,0.00049509737,0.000113789574,0.037162777],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.995395,0.0008665695,0.0006767732,0.00078331615,0.0008717089,0.0014066169],"domain_scores_gemma":[0.9974126,0.00053566304,0.0001498032,0.0009390259,0.0005337414,0.0004291531],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0019625113,0.00026896453,0.00037783646,0.00071671564,0.0006039087,0.0007292084,0.00079426006,0.0002889525,0.00011529186],"category_scores_gemma":[0.0024105513,0.00027170614,0.00025294925,0.006717884,0.00021006947,0.00069821806,0.0005062396,0.0007553882,0.0069424785],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000042947845,0.00010103044,0.0013186618,0.00046728092,0.00001902827,0.000034300403,0.0022326068,0.00007309103,0.000021029466,0.16774958,0.008980242,0.8189989],"study_design_scores_gemma":[0.00041935497,0.0007422228,0.01007799,0.0007809212,0.000016889087,0.00006332066,0.00064243976,0.56676257,0.00035921764,0.040003423,0.3795132,0.0006184497],"about_ca_topic_score_codex":0.024944115,"about_ca_topic_score_gemma":0.0014278179,"teacher_disagreement_score":0.8933532,"about_ca_system_score_codex":0.0005111604,"about_ca_system_score_gemma":0.0020846897,"threshold_uncertainty_score":0.99997354},"labels":[],"label_agreement":null},{"id":"W4388092895","doi":"10.48550/arxiv.2310.18811","title":"Hierarchical Framework for Interpretable and Probabilistic Model-Based Safe Reinforcement Learning","year":2023,"lang":"en","type":"preprint","venue":"arXiv (Cornell University)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Science Foundation Ireland; European Commission; Canadian Institute of Steel Construction","keywords":"Reinforcement learning; Interpretability; Probabilistic logic; Computer science; Artificial intelligence; Machine learning; Black box; Domain (mathematical analysis); Focus (optics)","score_opus":0.09495309765304245,"score_gpt":0.24527691708822108,"score_spread":0.15032381943517864,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388092895","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.017843252,0.000026618654,0.9805732,0.00039336516,0.00012382727,0.0005617549,0.000006692431,0.00033339794,0.00013793162],"genre_scores_gemma":[0.9661595,0.00008805553,0.03172769,0.000066764784,0.00003505708,0.000014196629,0.000026926826,0.000021782902,0.0018599766],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99756044,0.00017751972,0.0002652371,0.0013185595,0.00017802994,0.00050023827],"domain_scores_gemma":[0.99695593,0.0014152989,0.00014589237,0.0010132903,0.00022838892,0.00024122075],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0007944762,0.00027730444,0.0004177307,0.00034768187,0.0002978747,0.00025277358,0.0013465637,0.00034871214,0.000012614331],"category_scores_gemma":[0.0011421425,0.00030106652,0.00028339907,0.00061045727,0.00020771047,0.00017924266,0.0022116073,0.0011477622,0.000025295725],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00004492104,0.000033907418,0.0005642044,0.00029400608,0.00006055298,0.00001594703,0.00012438842,0.8213322,0.0000022737045,0.17680918,0.000033013584,0.0006854013],"study_design_scores_gemma":[0.00017721817,0.00007237819,0.00004240077,0.00016692055,0.00003864501,2.468219e-7,0.000013362274,0.6808752,0.000008626348,0.31831324,0.00008307144,0.00020870028],"about_ca_topic_score_codex":0.00006179361,"about_ca_topic_score_gemma":0.000011739368,"teacher_disagreement_score":0.94884545,"about_ca_system_score_codex":0.0002428377,"about_ca_system_score_gemma":0.00042861406,"threshold_uncertainty_score":0.99994415},"labels":[],"label_agreement":null},{"id":"W4388100841","doi":"10.1002/qre.3466","title":"Special issue on systems reliability and safety","year":2023,"lang":"en","type":"article","venue":"Quality and Reliability Engineering International","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Reliability engineering; Reliability (semiconductor); Computer science; Engineering; Physics","score_opus":0.022380856244474855,"score_gpt":0.30645469886519855,"score_spread":0.2840738426207237,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388100841","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8050038,0.00019125611,0.12513831,0.04241669,0.011810338,0.0013573464,0.00020905463,0.0027569765,0.011116212],"genre_scores_gemma":[0.9819524,0.0010555737,0.0060201287,0.00031760917,0.007392415,0.00010513369,0.000081131555,0.000039102513,0.0030364513],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9973325,0.00020062365,0.0005771187,0.0007898795,0.0007594345,0.0003404058],"domain_scores_gemma":[0.99739736,0.0014584054,0.000070145274,0.00066508615,0.0002103431,0.00019863776],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0033991968,0.00020590218,0.00032305377,0.00024419307,0.00016791538,0.00029053597,0.0005796741,0.00015003182,0.000050450522],"category_scores_gemma":[0.0025541466,0.00018504774,0.00010482695,0.0005407111,0.00011855927,0.00035630044,0.0004368196,0.0003778246,0.00010281232],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00037920213,0.00081899704,0.052595187,0.0018608065,0.00034469806,0.0000570851,0.0029208225,0.39608055,0.00033049565,0.4249105,0.019837584,0.09986409],"study_design_scores_gemma":[0.00053621747,0.00010005366,0.16591243,0.000086524175,0.000007120225,0.000010142132,0.00007709843,0.5557914,0.00006477014,0.0025919294,0.27443138,0.00039093505],"about_ca_topic_score_codex":0.00014829599,"about_ca_topic_score_gemma":0.0000021147714,"teacher_disagreement_score":0.42231855,"about_ca_system_score_codex":0.00014990098,"about_ca_system_score_gemma":0.000043166052,"threshold_uncertainty_score":0.7546028},"labels":[],"label_agreement":null},{"id":"W4388469715","doi":"10.1109/tse.2023.3327575","title":"Identifying the Hazard Boundary of ML-Enabled Autonomous Systems Using Cooperative Coevolutionary Search","year":2023,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Alliance de recherche numérique du Canada; Natural Sciences and Engineering Research Council of Canada; Mitacs; Canada Research Chairs","keywords":"Computer science; Boundary (topology); Hazard; Context (archaeology); Component (thermodynamics); Genetic algorithm; Artificial intelligence; Metaheuristic; Machine learning; Algorithm; Mathematics","score_opus":0.03903614166885182,"score_gpt":0.2854306910932624,"score_spread":0.24639454942441058,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388469715","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.07473881,0.00029635467,0.92317986,0.000061901264,0.0007115393,0.0002751079,0.000024048328,0.0007071718,0.0000052093474],"genre_scores_gemma":[0.98767734,0.00008149367,0.011832711,0.000011157299,0.000048306792,0.00006368299,0.0000032616013,0.00002867629,0.00025339646],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9978389,0.00015580459,0.00040629803,0.00041993632,0.0007059787,0.00047308372],"domain_scores_gemma":[0.99796945,0.0009261783,0.00005000125,0.00066251954,0.00028320128,0.00010867886],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009325526,0.00019734002,0.00029576462,0.00054699666,0.0005909922,0.00025754553,0.0007470491,0.00009842812,0.000016188023],"category_scores_gemma":[0.000084088555,0.00016803964,0.0001976475,0.0025671138,0.00010601015,0.0004924188,0.000020983038,0.0005173591,0.00006343061],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000056921967,0.000042559306,0.00008249682,0.00012595377,0.00012649506,0.000015832378,0.0006026134,0.9953357,0.0016876586,0.00022077783,0.000045976278,0.0017082616],"study_design_scores_gemma":[0.0001775983,0.000047836525,0.00050495984,0.00014001117,0.000022215541,0.000025946232,0.00018204883,0.9933731,0.0051014074,0.000021106154,0.0002198311,0.00018395527],"about_ca_topic_score_codex":0.0001565075,"about_ca_topic_score_gemma":0.0000049602622,"teacher_disagreement_score":0.91293854,"about_ca_system_score_codex":0.00025400482,"about_ca_system_score_gemma":0.00031885153,"threshold_uncertainty_score":0.68524575},"labels":[],"label_agreement":null},{"id":"W4388561737","doi":"10.1109/dasc58513.2023.10311167","title":"Certification Considerations of Software-Defined Radio Using Model-Based Development and Automated Testing","year":2023,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Computer science; Certification; Avionics; Integration testing; Development testing; Software engineering; Process (computing); Context (archaeology); Unit testing; White-box testing; Conformance testing; Domain (mathematical analysis); Software; Embedded system; Verification and validation; System integration; Software development; Software quality; Software construction; Programming language; Operating system; Engineering","score_opus":0.13556169149693215,"score_gpt":0.3272776044751659,"score_spread":0.19171591297823376,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4388561737","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.24912882,0.000011837968,0.7495941,0.00028876722,0.000011060309,0.00007955671,7.5164644e-7,0.0008609203,0.000024224952],"genre_scores_gemma":[0.5367002,5.394471e-7,0.46324703,0.000020595895,0.0000017448928,0.0000060406387,0.000003087516,0.0000028706982,0.000017904847],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9989663,0.00005473376,0.00028941105,0.00026289528,0.00025684462,0.00016980927],"domain_scores_gemma":[0.998476,0.0008577716,0.00005996016,0.0002749585,0.0002696885,0.00006156461],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00053978374,0.00007252714,0.00012986273,0.00024648066,0.00022351123,0.000096796495,0.000141373,0.000038187525,0.00000467741],"category_scores_gemma":[0.0013314947,0.00006644159,0.000023783954,0.0011030876,0.00005651834,0.00016386909,0.00009736223,0.000050015395,0.000008938402],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000032313721,0.00010960516,0.01911764,0.0001417386,0.000047620404,0.0000061820483,0.00089701655,0.91955376,0.022763781,0.0058004274,0.00031908508,0.031239936],"study_design_scores_gemma":[0.00010522634,0.0000074548643,0.0043577435,0.00001789382,0.0000039704096,0.0000017255718,0.000009236081,0.98201877,0.011778426,0.001618495,0.0000062325366,0.000074832926],"about_ca_topic_score_codex":0.000048179943,"about_ca_topic_score_gemma":0.000007156761,"teacher_disagreement_score":0.28757137,"about_ca_system_score_codex":0.000042188905,"about_ca_system_score_gemma":0.000494841,"threshold_uncertainty_score":0.27094096},"labels":[],"label_agreement":null},{"id":"W4390122176","doi":"10.5267/j.ijiec.2023.11.001","title":"Software testing and release decision at different statistical confidence levels with consideration of debuggers’ learning and negligent factors","year":2023,"lang":"en","type":"article","venue":"International Journal of Industrial Engineering Computations","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Reliability (semiconductor); Debugging; Reliability engineering; Software quality; Computer science; Software; Confidence interval; Resource (disambiguation); Process (computing); Imperfect; Function (biology); Risk analysis (engineering); Software development; Engineering; Statistics; Mathematics","score_opus":0.061233479030421825,"score_gpt":0.3108060944516885,"score_spread":0.24957261542126669,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390122176","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4998905,0.000017589164,0.49979383,0.00011847682,0.00011562171,0.00003398895,0.0000055596583,0.000023751758,6.749329e-7],"genre_scores_gemma":[0.9618492,0.000018957471,0.038058706,0.0000044363164,0.00005190241,0.0000011633218,0.00000465972,0.000005758723,0.0000052268465],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99868697,0.000054460575,0.0004107542,0.0001476359,0.0005844695,0.000115682815],"domain_scores_gemma":[0.994353,0.004670197,0.00019698004,0.00005749942,0.0006027517,0.000119565564],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002858039,0.000096541786,0.00019011617,0.00036761598,0.00008987315,0.00014089962,0.0001802672,0.000048604208,0.0000062679364],"category_scores_gemma":[0.0057886234,0.00007717149,0.000031831838,0.00030655105,0.00005416379,0.00022160381,0.00015050275,0.00025908797,5.946727e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005697726,0.000048501228,0.13468349,0.000019268875,0.00021839111,0.00010136174,0.00065165054,0.7922201,0.001348049,0.0019233475,0.0000765438,0.06865232],"study_design_scores_gemma":[0.0010737989,0.00032473676,0.10242504,0.00043850392,0.0000273296,0.00016309776,0.000102498605,0.89272743,0.0013091513,0.0012125534,0.000044403707,0.00015144447],"about_ca_topic_score_codex":0.00002470102,"about_ca_topic_score_gemma":0.0000032085588,"teacher_disagreement_score":0.4619587,"about_ca_system_score_codex":0.000079572084,"about_ca_system_score_gemma":0.000100355945,"threshold_uncertainty_score":0.6929941},"labels":[],"label_agreement":null},{"id":"W4391720360","doi":"10.1007/978-3-031-17299-1_2464","title":"Reliability Generalization","year":2023,"lang":"en","type":"book-chapter","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Generalization; Reliability engineering; Reliability (semiconductor); Computer science; Mathematics; Engineering; Physics; Mathematical analysis; Thermodynamics","score_opus":0.038244011650895554,"score_gpt":0.28205656990692124,"score_spread":0.24381255825602569,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391720360","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0000038847506,0.000059382888,0.40683475,0.0008597325,0.0002546297,0.00017588644,0.0000051955994,0.0010250906,0.59078145],"genre_scores_gemma":[0.00023855572,0.0004093873,0.007042392,0.00015567866,0.00011861436,0.000009147966,0.00003771753,0.000028119493,0.9919604],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9979225,0.00003484115,0.00034022957,0.00074553915,0.00071873143,0.0002382069],"domain_scores_gemma":[0.99783814,0.00024500847,0.00007967153,0.0014075491,0.0003134167,0.00011623266],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006571301,0.0002048392,0.00029189678,0.0002409417,0.000101694764,0.000174175,0.0010674034,0.00029072643,0.00047986698],"category_scores_gemma":[0.00021166289,0.00017455146,0.00025469656,0.00018629522,0.000074011296,0.00019899095,0.00057497504,0.00028979828,0.003096585],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[7.68164e-7,0.000008701989,0.000019657375,0.00003660566,0.000028796945,0.000009793796,0.00001947295,0.00022623673,0.0000057225316,0.96783715,0.011903621,0.019903492],"study_design_scores_gemma":[0.00008215906,0.00003886293,0.00011800459,0.000042918182,0.000016529828,0.0000022586562,8.709568e-7,0.046337012,0.000044269498,0.6361046,0.31680635,0.0004061645],"about_ca_topic_score_codex":0.000044208766,"about_ca_topic_score_gemma":0.000023062667,"teacher_disagreement_score":0.40117893,"about_ca_system_score_codex":0.000099568744,"about_ca_system_score_gemma":0.00014625213,"threshold_uncertainty_score":0.9976796},"labels":[],"label_agreement":null},{"id":"W4392182143","doi":"10.1109/icra57147.2024.10611316","title":"Safety Optimized Reinforcement Learning via Multi-Objective Policy Optimization","year":2024,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Reinforcement learning; Computer science; Reinforcement; Artificial intelligence; Psychology; Social psychology","score_opus":0.02094009631972392,"score_gpt":0.31424824201442514,"score_spread":0.2933081456947012,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4392182143","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000022161701,0.00023338893,0.9883779,0.0022046608,0.00030730123,0.0005930716,0.0000022648571,0.000945052,0.007314191],"genre_scores_gemma":[0.11895886,0.0012146395,0.8612181,0.00026715983,0.00034046583,0.00020641915,0.00014948861,0.000057388173,0.01758749],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.996306,0.00037656576,0.0006734182,0.001267639,0.00083830894,0.00053807016],"domain_scores_gemma":[0.9977104,0.00030705577,0.0001796364,0.001135025,0.0004583362,0.00020956846],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0012445827,0.0003961167,0.00055439636,0.0008544285,0.0002566256,0.00080724136,0.0014558632,0.0003723662,0.0002415411],"category_scores_gemma":[0.00074057607,0.00034465626,0.00046883113,0.0012646404,0.00008396208,0.00020884452,0.0071965153,0.0017312159,0.0002696133],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001630544,0.000048698013,0.000016458938,0.00013200748,0.00021274843,0.000008766689,0.0006535934,0.9800077,0.000007944742,0.002766324,0.00006891481,0.016060533],"study_design_scores_gemma":[0.00037971145,0.000047058733,0.000018582215,0.00011372811,0.000032576983,0.0000033663948,0.000035857607,0.99629873,0.0001758657,0.0021763393,0.00035898242,0.00035922532],"about_ca_topic_score_codex":0.0016508776,"about_ca_topic_score_gemma":0.00001173982,"teacher_disagreement_score":0.12715982,"about_ca_system_score_codex":0.0008007261,"about_ca_system_score_gemma":0.00096157717,"threshold_uncertainty_score":0.9999005},"labels":[],"label_agreement":null},{"id":"W4394053924","doi":"10.5281/zenodo.883812","title":"Empirical Data From A Longitudinal Software Maintenance Experiment","year":2017,"lang":"en","type":"dataset","venue":"PolyPublie (École Polytechnique de Montréal)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Software; Longitudinal data; Empirical research; Computer science; Statistics; Data mining; Mathematics; Programming language","score_opus":0.05726297677289785,"score_gpt":0.34370171011894823,"score_spread":0.2864387333460504,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4394053924","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000079311496,0.002163202,0.47693425,0.0025148464,0.00024048571,0.00046253987,0.5171031,0.0004938464,0.000008429249],"genre_scores_gemma":[0.0005713095,0.001567464,0.15325707,0.0014125912,0.0005128856,0.00059416826,0.841543,0.000067963025,0.00047354447],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99212855,0.0004900837,0.0009960992,0.0029367532,0.0018164773,0.0016320077],"domain_scores_gemma":[0.97990996,0.0008251047,0.00074426347,0.017362803,0.00033371232,0.0008241897],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science"],"consensus_categories":["open_science"],"category_scores_codex":[0.0018452563,0.0009413505,0.0013182815,0.0006545027,0.00088517787,0.0022429721,0.018797446,0.001061485,0.00023017314],"category_scores_gemma":[0.003394789,0.0008715632,0.00047945033,0.00063272385,0.00040908394,0.0013973996,0.011013867,0.001981904,0.00054157985],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000040455434,0.00047174862,0.004909831,0.00006358171,0.00021986302,0.0005685777,0.000031513344,0.00005775727,0.000016251219,0.0000502852,0.9878287,0.0057414677],"study_design_scores_gemma":[0.0004965426,0.00014709288,0.0072758915,0.00031300937,0.00011904327,0.00008538244,0.000012982256,0.04490937,0.00016668445,0.0018174719,0.9434852,0.0011713078],"about_ca_topic_score_codex":0.05061874,"about_ca_topic_score_gemma":0.010750687,"teacher_disagreement_score":0.3244399,"about_ca_system_score_codex":0.0009516264,"about_ca_system_score_gemma":0.001541733,"threshold_uncertainty_score":0.9993735},"labels":[],"label_agreement":null},{"id":"W4398258577","doi":"10.7910/dvn/ii5jzg/uyd6xb","title":"MSP_F_20_NFL_4_10.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Artificial intelligence; History; Archaeology","score_opus":0.023225065483537803,"score_gpt":0.26859053903504077,"score_spread":0.24536547355150296,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398258577","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[4.943944e-7,0.0000023599425,0.01556171,0.000106925174,0.00040137224,0.00020106987,0.9834046,0.00021833486,0.00010316102],"genre_scores_gemma":[0.0000027991996,0.0005607874,0.0039209863,0.0010630265,0.00030749422,0.000026601425,0.9939174,0.000013995275,0.0001869223],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99644214,0.00025175296,0.0004412652,0.0011988918,0.0011081338,0.00055778865],"domain_scores_gemma":[0.9953764,0.00030261916,0.00017100894,0.0035739813,0.00014189181,0.00043409364],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00052630267,0.00038215067,0.00056270044,0.0002735208,0.0001817153,0.00060976297,0.0052498383,0.0003211283,0.007538689],"category_scores_gemma":[0.0010646881,0.00035967137,0.00030184462,0.00095473166,0.0001310157,0.0006402776,0.003085124,0.0009131287,0.51617527],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007133082,0.000074305644,0.000003556256,0.00012508604,0.000093174145,0.00026047713,0.000014952404,0.000007743698,0.0000040715877,0.00015789257,0.9972163,0.0020353505],"study_design_scores_gemma":[0.00020214173,0.00005681251,0.000009932736,0.000037710433,0.000050993058,0.000012120262,0.000006382907,0.001284163,0.000018743209,0.00017638311,0.9977644,0.00038018226],"about_ca_topic_score_codex":0.00032764257,"about_ca_topic_score_gemma":0.000044592613,"teacher_disagreement_score":0.5086366,"about_ca_system_score_codex":0.00010590186,"about_ca_system_score_gemma":0.0004156073,"threshold_uncertainty_score":0.9998855},"labels":[],"label_agreement":null},{"id":"W4398259737","doi":"10.7910/dvn/ii5jzg/xyl91v","title":"MSP_F_100_NFL_4_17.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Computer science; History; Archaeology","score_opus":0.0224901148611318,"score_gpt":0.2680181618805589,"score_spread":0.24552804701942707,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398259737","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[4.619359e-7,0.000002506518,0.017822433,0.000110570356,0.00042523415,0.00020544119,0.9811086,0.00022704422,0.00009772354],"genre_scores_gemma":[0.0000031697161,0.0006106112,0.0037104017,0.0010673004,0.00031422754,0.000027114062,0.99407166,0.000014551698,0.0001809869],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99637496,0.00026757433,0.0004652431,0.0012468055,0.0010648916,0.0005805212],"domain_scores_gemma":[0.9951714,0.0003321096,0.000180584,0.003709129,0.00015323662,0.00045349024],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005611695,0.0003983089,0.00059365534,0.00028661967,0.00018979081,0.0006325221,0.0054436354,0.00033379823,0.0072900895],"category_scores_gemma":[0.0011369688,0.0003760156,0.0003158959,0.0010075853,0.00014101311,0.0006719267,0.0032652428,0.00094514224,0.49283746],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007597425,0.00007699493,0.000004235613,0.00013114662,0.00009730312,0.00027111545,0.000015041168,0.000008519135,0.0000039852966,0.00015593992,0.9972785,0.0019496261],"study_design_scores_gemma":[0.00021990451,0.000057580713,0.000009318976,0.000039456983,0.000052376658,0.000012687063,0.000006726192,0.0012520695,0.00001925809,0.00016793783,0.99776566,0.00039701094],"about_ca_topic_score_codex":0.00036696056,"about_ca_topic_score_gemma":0.000051707022,"teacher_disagreement_score":0.48554736,"about_ca_system_score_codex":0.00011153633,"about_ca_system_score_gemma":0.00043734055,"threshold_uncertainty_score":0.9999374},"labels":[],"label_agreement":null},{"id":"W4398260233","doi":"10.7910/dvn/ii5jzg/etg9a6","title":"MSP_F_100_NFL_4_39.xlsx","year":2020,"lang":"tl","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.028991329473459743,"score_gpt":0.2755982653031389,"score_spread":0.24660693582967916,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398260233","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000014107541,0.000009591113,0.019668901,0.00031523846,0.0012673441,0.0006278948,0.9774983,0.00031160304,0.00028702273],"genre_scores_gemma":[0.00010339896,0.00270773,0.0048744236,0.0020579444,0.0011019626,0.000053091564,0.98815864,0.000058179456,0.00088463],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99034184,0.0010304621,0.0012967703,0.0030663915,0.0027029545,0.0015615807],"domain_scores_gemma":[0.9897333,0.001010828,0.0005352174,0.006864555,0.00048158408,0.0013745273],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0016642476,0.0010619034,0.0015107893,0.000575973,0.00064822036,0.0018800657,0.009398469,0.0008192752,0.042386305],"category_scores_gemma":[0.0034627263,0.0010675648,0.0008525897,0.0025368512,0.00062479853,0.0014708708,0.007850202,0.0023921418,0.85814667],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006683085,0.00036742626,0.00005415476,0.0005365183,0.00041233114,0.0010168502,0.00013640833,0.000049435574,0.000023166702,0.00040689664,0.99041617,0.0065138084],"study_design_scores_gemma":[0.00070132955,0.00025286622,0.00007220959,0.0002116076,0.00027873364,0.000041213407,0.00006539896,0.011205131,0.000051874613,0.00020041013,0.9858524,0.0010667961],"about_ca_topic_score_codex":0.00093809364,"about_ca_topic_score_gemma":0.00013042497,"teacher_disagreement_score":0.8157604,"about_ca_system_score_codex":0.00032891866,"about_ca_system_score_gemma":0.0013347694,"threshold_uncertainty_score":0.9999094},"labels":[],"label_agreement":null},{"id":"W4398271454","doi":"10.7910/dvn/ii5jzg/8yocmy","title":"MSP_F_100_NFL_4_10.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence","score_opus":0.022290049408088767,"score_gpt":0.26740691415760903,"score_spread":0.24511686474952027,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398271454","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.808152e-7,0.00000249615,0.014326435,0.00011119608,0.00042313352,0.00020812229,0.984596,0.00022998847,0.000102008264],"genre_scores_gemma":[0.000003186046,0.0005990318,0.0036602204,0.0010682761,0.00031689723,0.00002742164,0.9941211,0.000014751806,0.0001891411],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99624825,0.00027121886,0.000473313,0.001255823,0.0011661208,0.00058528397],"domain_scores_gemma":[0.9951395,0.00033237296,0.00018396482,0.0037321367,0.00015482712,0.00045723113],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005721374,0.00040208912,0.00059668906,0.00028934685,0.00019131239,0.0006459384,0.0054889256,0.0003359956,0.00762022],"category_scores_gemma":[0.0011402544,0.00037971645,0.00031880187,0.0010158108,0.00014230321,0.0006778075,0.0032818099,0.0009540043,0.50905246],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000074146615,0.00007718176,0.000004187513,0.00013315122,0.00009824684,0.0002779787,0.000015138294,0.000008436825,0.00000398368,0.00016013959,0.99720526,0.002008889],"study_design_scores_gemma":[0.00021225474,0.000058221103,0.000011420478,0.00003983308,0.000052831692,0.000012932602,0.0000066724383,0.001229443,0.000018970772,0.00017383961,0.99778277,0.00040080358],"about_ca_topic_score_codex":0.00037593048,"about_ca_topic_score_gemma":0.000051455485,"teacher_disagreement_score":0.50143224,"about_ca_system_score_codex":0.00011377404,"about_ca_system_score_gemma":0.0004443204,"threshold_uncertainty_score":0.9998919},"labels":[],"label_agreement":null},{"id":"W4398277276","doi":"10.7910/dvn/ii5jzg/z9n0pc","title":"MSP_F_100_NFL_4_32.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398277276","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.7341875e-7,0.0000025637078,0.014568853,0.000107447064,0.00042639795,0.00020574733,0.98436666,0.00022775501,0.00009392261],"genre_scores_gemma":[0.0000036632755,0.0005702637,0.0036207165,0.0010588315,0.00031344127,0.000027115379,0.9942098,0.000014572442,0.00018154064],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962889,0.0002687109,0.00046222092,0.0012452226,0.0011551703,0.00057977566],"domain_scores_gemma":[0.99518013,0.00033128008,0.00017950352,0.0037029986,0.00015290637,0.00045315677],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005594673,0.00039801988,0.0005909926,0.00028533553,0.00018902085,0.0006324584,0.0054457104,0.00033135526,0.0072793323],"category_scores_gemma":[0.0011222726,0.00037577059,0.0003156933,0.0010071425,0.00014058469,0.0006970303,0.0032502213,0.0009433133,0.50682527],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000075304547,0.00007635836,0.0000038921353,0.0001322457,0.000097238495,0.0002764218,0.000017173124,0.000008265583,0.000003947175,0.00015914455,0.9973605,0.0018572444],"study_design_scores_gemma":[0.00021094787,0.000058349866,0.000010695003,0.000039456987,0.000052244868,0.0000128837255,0.000007738012,0.0012473185,0.0000184389,0.00017461309,0.9977706,0.0003967026],"about_ca_topic_score_codex":0.00035805686,"about_ca_topic_score_gemma":0.000050281284,"teacher_disagreement_score":0.49954593,"about_ca_system_score_codex":0.00011165101,"about_ca_system_score_gemma":0.0004372918,"threshold_uncertainty_score":0.9999353},"labels":[],"label_agreement":null},{"id":"W4398286541","doi":"10.7910/dvn/ii5jzg/7fnwwf","title":"MSP_F_20_NFL_4_40.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.023513156280313474,"score_gpt":0.26932418256041757,"score_spread":0.2458110262801041,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398286541","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.1363327e-7,0.0000023105779,0.015792197,0.00010596117,0.00040807563,0.00020139218,0.9831668,0.0002183801,0.00010432808],"genre_scores_gemma":[0.0000028260063,0.00057114824,0.0039343988,0.0010744361,0.0003072567,0.000026620486,0.9938563,0.000014018119,0.0002129747],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9964409,0.00025178693,0.0004414684,0.0012000904,0.0011078563,0.0005579042],"domain_scores_gemma":[0.995374,0.0003023887,0.00017106217,0.0035766528,0.00014164699,0.0004342516],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005264081,0.0003822758,0.00056299893,0.00027368678,0.00018145383,0.0006099094,0.005249454,0.00032124663,0.0071684397],"category_scores_gemma":[0.0010631423,0.00035980492,0.00030193815,0.0009555563,0.00013080472,0.00064038334,0.0030858794,0.0009113692,0.58846784],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007151772,0.00007432325,0.0000035310316,0.00012522787,0.00009321796,0.000261246,0.0000151214945,0.000007804521,0.0000033387064,0.00016255028,0.9970628,0.0021836988],"study_design_scores_gemma":[0.00020226522,0.00005692903,0.000009559141,0.00003775288,0.000051021667,0.000012233936,0.0000064632864,0.0012761093,0.00001885685,0.00018284761,0.99776566,0.0003803081],"about_ca_topic_score_codex":0.00032771335,"about_ca_topic_score_gemma":0.000046136585,"teacher_disagreement_score":0.5812994,"about_ca_system_score_codex":0.00010593976,"about_ca_system_score_gemma":0.0004183964,"threshold_uncertainty_score":0.9998854},"labels":[],"label_agreement":null},{"id":"W4398300558","doi":"10.7910/dvn/ii5jzg/0wu5qu","title":"MSP_F_100_NFL_4_2.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Computer science; History; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398300558","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.89023e-7,0.000002372948,0.014198916,0.00010644621,0.00041995008,0.000203197,0.98474324,0.00022434555,0.00010096353],"genre_scores_gemma":[0.0000031418574,0.0006010699,0.0036613392,0.0010552432,0.00030600393,0.000026600179,0.99415594,0.000014324282,0.00017636159],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963344,0.00026547525,0.00045684987,0.0012329918,0.0011375829,0.00057274813],"domain_scores_gemma":[0.9952217,0.00032384775,0.00017717722,0.0036790008,0.00015078668,0.00044748816],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005544711,0.00039283326,0.0005838311,0.00028321208,0.00018643413,0.0006151311,0.005417486,0.00032846545,0.007295273],"category_scores_gemma":[0.0011102838,0.00037078725,0.00031173133,0.0009975708,0.00013929508,0.00066480326,0.003241375,0.0009318525,0.49630955],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000073388496,0.00007594783,0.0000041354324,0.00013045152,0.000096339776,0.00027400124,0.000015262776,0.000007845236,0.000003852091,0.00016542625,0.99726224,0.0019571576],"study_design_scores_gemma":[0.00020673011,0.000057189714,0.000011509901,0.00003890382,0.000051754054,0.000012595679,0.00000658097,0.0011887209,0.000018686322,0.00017491981,0.99784094,0.00039143843],"about_ca_topic_score_codex":0.0003815818,"about_ca_topic_score_gemma":0.000052529143,"teacher_disagreement_score":0.48901427,"about_ca_system_score_codex":0.0001090275,"about_ca_system_score_gemma":0.0004313343,"threshold_uncertainty_score":0.9999637},"labels":[],"label_agreement":null},{"id":"W4398322551","doi":"10.7910/dvn/ii5jzg/vsnztq","title":"MSP_F_20_NFL_4_16.xlsx","year":2020,"lang":"pt","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Mathematics; Artificial intelligence; Geography; Archaeology","score_opus":0.0322135567374304,"score_gpt":0.2809130324253887,"score_spread":0.24869947568795828,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398322551","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000009801818,0.00000869946,0.030301291,0.00027864674,0.0012012935,0.0006039544,0.96712154,0.0002839685,0.00019081323],"genre_scores_gemma":[0.000106635925,0.0018950093,0.0030997389,0.0020547104,0.0008418653,0.00005229223,0.9905592,0.000055963403,0.001334571],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9910177,0.0008953677,0.0012021982,0.0028933997,0.0025285983,0.001462755],"domain_scores_gemma":[0.99058473,0.00083686423,0.0004911817,0.0063577085,0.0004324396,0.00129707],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0014362427,0.0010042826,0.0014113081,0.00055849174,0.0006033223,0.0017053338,0.0085977875,0.0007771252,0.039684348],"category_scores_gemma":[0.002881966,0.0010097218,0.0007989118,0.002406333,0.0005054544,0.0013203347,0.0070186825,0.002265483,0.85633725],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000061669976,0.00035831297,0.00004825284,0.00051662436,0.0003882707,0.0009993489,0.00016877928,0.000050833063,0.000020194264,0.00024649588,0.9904524,0.006688825],"study_design_scores_gemma":[0.0006496431,0.0002539875,0.000057374396,0.00019441805,0.00026548712,0.000041585863,0.000078973004,0.014420281,0.00004665208,0.00011360205,0.9828751,0.0010028945],"about_ca_topic_score_codex":0.00067307526,"about_ca_topic_score_gemma":0.00009908716,"teacher_disagreement_score":0.8166529,"about_ca_system_score_codex":0.00029529745,"about_ca_system_score_gemma":0.0011844351,"threshold_uncertainty_score":0.999331},"labels":[],"label_agreement":null},{"id":"W4398329608","doi":"10.7910/dvn/ii5jzg/rnnvk8","title":"MSP_F_50_NFL_4_12.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Geography; Artificial intelligence; Archaeology","score_opus":0.02345769106095168,"score_gpt":0.2691457205047474,"score_spread":0.2456880294437957,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398329608","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.660102e-7,0.0000024986066,0.014031388,0.000111776164,0.00041811002,0.00020578009,0.9849057,0.00022378047,0.0001003723],"genre_scores_gemma":[0.0000037838472,0.00060177833,0.003765024,0.001075238,0.00031290448,0.00002734102,0.9940469,0.000014235,0.00015274252],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99633235,0.00025845243,0.00045773666,0.0012353628,0.0011404026,0.00057568203],"domain_scores_gemma":[0.9952328,0.00031396924,0.00017651398,0.0036802837,0.00014666408,0.0004497746],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00054486946,0.00039382765,0.00058085535,0.0002818548,0.00018950283,0.00063243555,0.0054284865,0.00033172002,0.0072891274],"category_scores_gemma":[0.0011222215,0.00037149058,0.00031216006,0.0009696711,0.0001379748,0.0006664281,0.0032438932,0.0009460225,0.49888948],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010611048,0.000077683355,0.000003959263,0.0001300118,0.00009636935,0.0002722947,0.000015186906,0.000008179758,0.0000033183455,0.00017033365,0.9972366,0.0019754176],"study_design_scores_gemma":[0.00020824185,0.00008121272,0.000010559798,0.00004025062,0.000052581006,0.000012795677,0.0000066247794,0.0011823718,0.000016086644,0.00017837045,0.99781877,0.00039215208],"about_ca_topic_score_codex":0.00034080123,"about_ca_topic_score_gemma":0.00004813293,"teacher_disagreement_score":0.49160036,"about_ca_system_score_codex":0.000109037224,"about_ca_system_score_gemma":0.0004185519,"threshold_uncertainty_score":0.9999526},"labels":[],"label_agreement":null},{"id":"W4398331983","doi":"10.7910/dvn/ii5jzg/0wo44w","title":"MSP_F_100_NFL_4_1.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398331983","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.720255e-7,0.0000024373999,0.014438762,0.00011002771,0.00042047832,0.00020358687,0.98450285,0.00022495941,0.000096312906],"genre_scores_gemma":[0.000003049193,0.0006123042,0.0036442308,0.0010870142,0.00030885823,0.000026604246,0.9941266,0.000014316753,0.00017700421],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99632305,0.00026671504,0.0004568616,0.0012337215,0.0011462931,0.0005733485],"domain_scores_gemma":[0.99521065,0.00033323164,0.00017708118,0.003679898,0.00015124025,0.0004478806],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00055780733,0.0003929698,0.00058459543,0.00028298103,0.00018735988,0.00062552415,0.0054127085,0.00032857628,0.007089812],"category_scores_gemma":[0.0011223075,0.00037080378,0.00031192423,0.0009986719,0.00013746344,0.0006647031,0.0032362947,0.0009356042,0.49571192],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007273148,0.00007679552,0.000004268074,0.00013184901,0.00009596378,0.0002732401,0.00001471413,0.0000082100905,0.000003841435,0.00016263813,0.99723434,0.0019868503],"study_design_scores_gemma":[0.00020606522,0.000057521294,0.00001124708,0.000039314997,0.000051552415,0.0000121366165,0.0000066851107,0.0012075918,0.00001835978,0.00017495555,0.9978231,0.00039144684],"about_ca_topic_score_codex":0.00036544047,"about_ca_topic_score_gemma":0.000052465737,"teacher_disagreement_score":0.48862213,"about_ca_system_score_codex":0.000110152636,"about_ca_system_score_gemma":0.00043302987,"threshold_uncertainty_score":0.99996847},"labels":[],"label_agreement":null},{"id":"W4398347322","doi":"10.7910/dvn/ii5jzg/b8n0xq","title":"MSP_F_100_NFL_4_40.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398347322","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.042837e-7,0.000002415641,0.014514022,0.000110285815,0.00042677872,0.00020654527,0.98441166,0.0002278777,0.00009983159],"genre_scores_gemma":[0.0000031949514,0.00060597353,0.0036596293,0.0010734326,0.00031442454,0.000027110142,0.9941218,0.000014598154,0.00017984839],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99627876,0.00026929812,0.00046331112,0.0012489236,0.0011586911,0.0005810204],"domain_scores_gemma":[0.99516463,0.00033231542,0.00017991589,0.003715962,0.00015327767,0.00045391815],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005623766,0.000398692,0.0005924363,0.00028704837,0.00018991734,0.00063315814,0.005461233,0.00033327274,0.007189798],"category_scores_gemma":[0.0011320277,0.0003763748,0.00031622555,0.0010096862,0.00014117926,0.0006732004,0.003267855,0.0009453175,0.4971392],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000074027243,0.00007648625,0.0000041124995,0.00013232927,0.000097401215,0.00027681372,0.000015212088,0.000008474078,0.000004053067,0.00016411567,0.9972459,0.0019677042],"study_design_scores_gemma":[0.00021074804,0.00005789175,0.0000111888185,0.00003953681,0.000052357635,0.000012892674,0.0000066927737,0.001221406,0.000018896659,0.0001791677,0.9977919,0.0003973247],"about_ca_topic_score_codex":0.0003758433,"about_ca_topic_score_gemma":0.000052968444,"teacher_disagreement_score":0.48994938,"about_ca_system_score_codex":0.000112810834,"about_ca_system_score_gemma":0.00044406042,"threshold_uncertainty_score":0.9999197},"labels":[],"label_agreement":null},{"id":"W4398392323","doi":"10.7910/dvn/ii5jzg/doctby","title":"MSP_F_100_NFL_4_46.xlsx","year":2020,"lang":"pt","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Operations research; Mathematics; Geography; Artificial intelligence; Archaeology","score_opus":0.03104582651259454,"score_gpt":0.2799452977929295,"score_spread":0.24889947128033493,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398392323","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000010878877,0.000008712508,0.02792195,0.00029983427,0.0012618054,0.0006162185,0.96939415,0.00029517515,0.00019130448],"genre_scores_gemma":[0.00010886633,0.0020767155,0.0028737613,0.0020305945,0.0008710942,0.000053182186,0.99059975,0.000058002355,0.0013280242],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9905537,0.0009857897,0.0012663968,0.003015059,0.00265305,0.0015260333],"domain_scores_gemma":[0.990107,0.0009253286,0.00051878096,0.006620538,0.00046880054,0.0013595434],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0015658394,0.001048977,0.0014815322,0.00059042877,0.0006352761,0.0017737618,0.008951247,0.0008076506,0.039802417],"category_scores_gemma":[0.0031204866,0.0010575708,0.00084313547,0.0025725288,0.00054791674,0.0014087626,0.0074439324,0.002354941,0.85474265],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000638912,0.00036303233,0.00005520609,0.0005466368,0.0004082903,0.0009739931,0.00016456073,0.000055097276,0.000020371192,0.0002717908,0.9905597,0.0065174624],"study_design_scores_gemma":[0.0006834421,0.00025212247,0.00006709752,0.00020720657,0.0002718404,0.000040217503,0.000081967046,0.014392726,0.00004841541,0.00011606579,0.9827902,0.0010487093],"about_ca_topic_score_codex":0.00077113614,"about_ca_topic_score_gemma":0.000118238546,"teacher_disagreement_score":0.8149402,"about_ca_system_score_codex":0.00031415725,"about_ca_system_score_gemma":0.0012682561,"threshold_uncertainty_score":0.99994665},"labels":[],"label_agreement":null},{"id":"W4398405777","doi":"10.7910/dvn/ii5jzg/ji7owm","title":"MSP_F_100_NFL_4_23.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Artificial intelligence; History; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398405777","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.4601187e-7,0.0000024578321,0.014027295,0.00011090267,0.00042515126,0.00020585449,0.98491424,0.00022767666,0.000085792904],"genre_scores_gemma":[0.0000034893792,0.00061919785,0.0035212662,0.0010613761,0.0003146467,0.000027079626,0.9942584,0.000014594254,0.00017994804],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99628747,0.00026877783,0.00046177313,0.0012449737,0.0011576714,0.00057935837],"domain_scores_gemma":[0.995175,0.00033438543,0.00017940319,0.0037062361,0.00015240286,0.00045256264],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.000560714,0.0003980288,0.00059098913,0.0002867436,0.00018780265,0.0006328587,0.0054455576,0.0003331365,0.0069584316],"category_scores_gemma":[0.0011302648,0.00037588106,0.0003158142,0.0010083702,0.00013877817,0.0006705051,0.0032583713,0.0009401998,0.49728703],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007233433,0.000075147254,0.0000050627586,0.00013164367,0.000097211036,0.00027903137,0.000014868522,0.000008152806,0.000004256996,0.00015776306,0.99727494,0.0019447028],"study_design_scores_gemma":[0.0002082116,0.000057418332,0.00001285964,0.00004045989,0.00005226544,0.000013105587,0.0000054111524,0.001232418,0.000019944853,0.00016833769,0.9977928,0.0003967746],"about_ca_topic_score_codex":0.0003707896,"about_ca_topic_score_gemma":0.000052056905,"teacher_disagreement_score":0.4903286,"about_ca_system_score_codex":0.00011122093,"about_ca_system_score_gemma":0.00042971657,"threshold_uncertainty_score":0.99993545},"labels":[],"label_agreement":null},{"id":"W4398542541","doi":"10.7910/dvn/ii5jzg/zlipsl","title":"MSP_F_100_NFL_4_3.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398542541","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.219932e-7,0.0000023314649,0.014389597,0.00011680181,0.00039713577,0.00020564294,0.98457414,0.00022058884,0.00009311507],"genre_scores_gemma":[0.0000041465278,0.00061746547,0.0036240977,0.0010417218,0.00030532715,0.000027190626,0.99419314,0.0000144016785,0.00017249669],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962859,0.00027425468,0.0004551541,0.0012360555,0.001166615,0.0005819769],"domain_scores_gemma":[0.99519706,0.00034127384,0.00017354215,0.0036820252,0.00015077753,0.00045535289],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00056347233,0.00039178616,0.00058173103,0.00029019223,0.00018852916,0.00060062815,0.0054834965,0.00032006166,0.0070910356],"category_scores_gemma":[0.0011342696,0.0003696681,0.0003115185,0.0009988435,0.0001442532,0.0006534823,0.0033631886,0.00094471715,0.4852552],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007403276,0.000075961885,0.000004330684,0.00013457304,0.00009539731,0.00027917954,0.000015520958,0.0000104901765,0.0000044454973,0.00013904697,0.99717766,0.0020559814],"study_design_scores_gemma":[0.00020538908,0.000058071277,0.000011811135,0.00004041528,0.00004983674,0.00001275586,0.0000057693665,0.0012762639,0.000019660598,0.00018552126,0.9977453,0.00038924583],"about_ca_topic_score_codex":0.0003841519,"about_ca_topic_score_gemma":0.000053871052,"teacher_disagreement_score":0.47816417,"about_ca_system_score_codex":0.00010970573,"about_ca_system_score_gemma":0.0004371928,"threshold_uncertainty_score":0.9998973},"labels":[],"label_agreement":null},{"id":"W4398548649","doi":"10.7910/dvn/ii5jzg/vas3zc","title":"MSP_F_100_NFL_4_48.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Mathematics; Algorithm; Artificial intelligence; Geography; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398548649","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[6.024735e-7,0.000002417405,0.014218512,0.0001108944,0.00042731466,0.00020834635,0.98470306,0.00022783398,0.000101007456],"genre_scores_gemma":[0.0000032948878,0.00060707243,0.0036613778,0.001074391,0.00031495263,0.000027294327,0.99411875,0.000014590963,0.00017829356],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962848,0.00026888863,0.00046295585,0.00124627,0.0011564944,0.00058063085],"domain_scores_gemma":[0.995172,0.0003311718,0.00017974123,0.0037100941,0.00015328085,0.00045371454],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00056107325,0.00039847536,0.0005917298,0.00028697445,0.00019058993,0.0006329825,0.0054581654,0.00033345574,0.007228779],"category_scores_gemma":[0.0011291,0.0003761463,0.0003161534,0.0010091465,0.00014162814,0.0006730123,0.0032644656,0.0009447653,0.5018487],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000074032896,0.000076437114,0.0000041506437,0.000133288,0.00009728308,0.00027663074,0.000015434654,0.00000836172,0.000004019509,0.00015950688,0.99726886,0.0019486209],"study_design_scores_gemma":[0.00021052253,0.00005696043,0.000011341252,0.00003960181,0.000052365922,0.000012882621,0.000006712564,0.0012342666,0.00001908738,0.00017383872,0.9977853,0.00039712476],"about_ca_topic_score_codex":0.00036639848,"about_ca_topic_score_gemma":0.00005162539,"teacher_disagreement_score":0.49461994,"about_ca_system_score_codex":0.000111538655,"about_ca_system_score_gemma":0.0004388646,"threshold_uncertainty_score":0.99992275},"labels":[],"label_agreement":null},{"id":"W4398616256","doi":"10.7910/dvn/ii5jzg/emv7pe","title":"MSP_F_100_NFL_4_5.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.022543176851258364,"score_gpt":0.2681597623408774,"score_spread":0.24561658548961904,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398616256","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.7523926e-7,0.0000023670884,0.014500997,0.00011141508,0.00042223162,0.00020356724,0.98443186,0.00022455987,0.00010243967],"genre_scores_gemma":[0.0000031342083,0.0005940495,0.0036412082,0.0010770131,0.0003117757,0.000025213538,0.9941533,0.00001433017,0.00017994088],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9963167,0.00026679135,0.0004616077,0.0012345775,0.0011466037,0.00057376974],"domain_scores_gemma":[0.99521077,0.00032637364,0.0001790008,0.0036845505,0.00015122644,0.0004480834],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00056257664,0.0003931781,0.0005916703,0.00028305923,0.00018724542,0.0006258557,0.0054138335,0.00032833708,0.006961303],"category_scores_gemma":[0.0011203507,0.00037103446,0.00031199597,0.0009981131,0.00013930435,0.0006640235,0.0032432457,0.0009344266,0.4961584],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000007270032,0.00007538382,0.0000039138963,0.00013158543,0.000095878924,0.00027360363,0.000015080086,0.000008372318,0.000003863346,0.00016191225,0.9971982,0.0020249397],"study_design_scores_gemma":[0.00020779831,0.0000569507,0.0000110992905,0.00003903729,0.000051333347,0.000012529248,0.0000066841694,0.001276143,0.000018079436,0.00017199478,0.99775666,0.0003916935],"about_ca_topic_score_codex":0.00037626748,"about_ca_topic_score_gemma":0.000052758045,"teacher_disagreement_score":0.4891971,"about_ca_system_score_codex":0.00011045669,"about_ca_system_score_gemma":0.00043340647,"threshold_uncertainty_score":0.99996734},"labels":[],"label_agreement":null},{"id":"W4398616439","doi":"10.7910/dvn/ii5jzg/ifjy0h","title":"MSP_F_100_NFL_4_11.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Operations research; History; Mathematics; Artificial intelligence; Archaeology","score_opus":0.022497111215832387,"score_gpt":0.26800169829012155,"score_spread":0.24550458707428915,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398616439","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.106387e-7,0.0000024449737,0.014553935,0.00011149878,0.0004211393,0.00020615161,0.98436135,0.00022782947,0.00011512192],"genre_scores_gemma":[0.0000032747798,0.0006015155,0.0036295545,0.0010706884,0.00031903363,0.000027098493,0.99415916,0.0000145867125,0.00017507917],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962852,0.00026905694,0.0004626349,0.0012462017,0.0011568288,0.00058010605],"domain_scores_gemma":[0.9951734,0.0003314691,0.0001797335,0.0037113929,0.0001527712,0.0004512629],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005600814,0.0003984213,0.00059180084,0.0002856772,0.0001845532,0.000631704,0.0054497044,0.00033307477,0.0072394675],"category_scores_gemma":[0.0011165335,0.00037612094,0.00031605922,0.0009930608,0.00014039587,0.0006720048,0.0032602064,0.00094393606,0.49998927],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000073973015,0.00007642831,0.0000052478736,0.00013227362,0.00009737922,0.00027670967,0.000014593411,0.000006385249,0.000003996908,0.00014230357,0.99762434,0.0016129203],"study_design_scores_gemma":[0.00021036908,0.000057729205,0.00001086648,0.000039310242,0.000052337764,0.000012891731,0.000008795195,0.0010034206,0.000018757115,0.00017168204,0.9980167,0.00039714],"about_ca_topic_score_codex":0.00029388306,"about_ca_topic_score_gemma":0.000046761426,"teacher_disagreement_score":0.4927498,"about_ca_system_score_codex":0.000111785776,"about_ca_system_score_gemma":0.00043989808,"threshold_uncertainty_score":0.9999313},"labels":[],"label_agreement":null},{"id":"W4398624170","doi":"10.7910/dvn/ii5jzg/fkwyil","title":"MSP_F_100_NFL_4_16.xlsx","year":2020,"lang":"pt","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; History; Artificial intelligence; Archaeology","score_opus":0.03091838749115236,"score_gpt":0.27974777761232117,"score_spread":0.2488293901211688,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398624170","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000011388163,0.000009063383,0.027807688,0.00028701418,0.0012531689,0.0006178353,0.96953183,0.00029553563,0.00018645183],"genre_scores_gemma":[0.00011877352,0.0019964792,0.00286259,0.0020377643,0.00085528527,0.000052896543,0.9906857,0.000057886296,0.0013326417],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9906099,0.00095697865,0.0012607123,0.0030083482,0.0026423952,0.0015216783],"domain_scores_gemma":[0.9901414,0.0009188536,0.0005163809,0.0066011795,0.00046769503,0.0013544895],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0015336697,0.0010468758,0.0014842378,0.00058551756,0.0006310264,0.0017693564,0.008937061,0.0008058958,0.039960027],"category_scores_gemma":[0.0030670338,0.0010557744,0.00083633134,0.0025405898,0.00054523745,0.0013875659,0.007427483,0.0023487164,0.852653],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006377268,0.00036851515,0.00005611389,0.0005456479,0.00040549558,0.0010581894,0.00016973741,0.000054898766,0.000019507383,0.0002471796,0.9904214,0.006589537],"study_design_scores_gemma":[0.00067696336,0.00025799935,0.000065278306,0.0002036429,0.00027247277,0.000043824828,0.000081796075,0.013782799,0.000046756224,0.00011126885,0.9834094,0.0010477763],"about_ca_topic_score_codex":0.00076935557,"about_ca_topic_score_gemma":0.00011365739,"teacher_disagreement_score":0.812693,"about_ca_system_score_codex":0.0003142666,"about_ca_system_score_gemma":0.0012566617,"threshold_uncertainty_score":0.9999529},"labels":[],"label_agreement":null},{"id":"W4398627424","doi":"10.7910/dvn/ii5jzg/1qlk2b","title":"MSP_F_100_NFL_4_6.xlsx","year":2020,"lang":"pt","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Computer science; History; Archaeology","score_opus":0.03104582651259454,"score_gpt":0.2799452977929295,"score_spread":0.24889947128033493,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398627424","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000010760025,0.000008599922,0.027954657,0.00029535245,0.001246294,0.0006073516,0.96939576,0.00029059305,0.00019061047],"genre_scores_gemma":[0.00010941528,0.0020304804,0.0028771712,0.002002895,0.0008559818,0.000052304888,0.9907046,0.000057039688,0.0013101059],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9906478,0.00097645505,0.0012514063,0.0029867492,0.0026284687,0.0015091496],"domain_scores_gemma":[0.9901785,0.0009288244,0.00051175675,0.006573477,0.00046307186,0.0013443329],"candidate_categories":["metaepi_narrow","scholarly_communication","open_science","research_integrity","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0015535891,0.0010357014,0.0014650237,0.0005805177,0.0006286426,0.0017580548,0.008886039,0.0007970493,0.03986689],"category_scores_gemma":[0.0030993242,0.0010437694,0.0008330512,0.002536249,0.00054171134,0.0013968702,0.007397357,0.0023283968,0.85136425],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000628413,0.0003584216,0.000054819422,0.000540429,0.000403055,0.0009643718,0.00016450536,0.000054727618,0.000020088653,0.00027203956,0.9907876,0.006317092],"study_design_scores_gemma":[0.0006753132,0.0002484039,0.00006454585,0.00020450514,0.00026819092,0.00003951967,0.00008162508,0.014404693,0.00004777188,0.000112857924,0.9828175,0.0010350753],"about_ca_topic_score_codex":0.0007681706,"about_ca_topic_score_gemma":0.00011630256,"teacher_disagreement_score":0.81149733,"about_ca_system_score_codex":0.00031184097,"about_ca_system_score_gemma":0.0012563121,"threshold_uncertainty_score":0.9999733},"labels":[],"label_agreement":null},{"id":"W4398686768","doi":"10.7910/dvn/ii5jzg/qicure","title":"MSP_F_100_NFL_4_12.xlsx","year":2020,"lang":"en","type":"dataset","venue":"Harvard Dataverse","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Context (archaeology); Resolution (logic); Computer science; Operations research; Mathematics; Artificial intelligence; Geography; Archaeology","score_opus":0.02250162810344538,"score_gpt":0.2680108294734913,"score_spread":0.24550920137004592,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4398686768","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[5.8292613e-7,0.0000022714464,0.0141554475,0.00011424797,0.00042856933,0.00022650564,0.9847488,0.00022792358,0.00009567826],"genre_scores_gemma":[0.0000032487826,0.00053567335,0.0036599066,0.0010703003,0.00031321615,0.000030878942,0.99418706,0.00001458963,0.00018514255],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9962783,0.0002728133,0.00046271196,0.0012471016,0.0011585597,0.0005805164],"domain_scores_gemma":[0.9951675,0.00033253405,0.000179284,0.0037143677,0.00015273939,0.00045358055],"candidate_categories":["metaepi_narrow","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0005643175,0.00039845976,0.0005916469,0.00028666356,0.00018927535,0.0006328788,0.0054590805,0.00033536306,0.0070246235],"category_scores_gemma":[0.0011442462,0.00037619035,0.0003160379,0.0009914757,0.0001409571,0.00067226816,0.0032654293,0.0009511288,0.49556285],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000074593336,0.0000771264,0.000004273417,0.00013669365,0.00009732056,0.00027643327,0.000015149455,0.00000858424,0.0000038616577,0.00016141219,0.99725,0.0019616417],"study_design_scores_gemma":[0.000210903,0.000057646314,0.000011601599,0.000037992228,0.000052293388,0.000012884252,0.0000065195372,0.001214372,0.000018947881,0.00017266326,0.997807,0.00039713204],"about_ca_topic_score_codex":0.00037248418,"about_ca_topic_score_gemma":0.000052324685,"teacher_disagreement_score":0.48853824,"about_ca_system_score_codex":0.000111941634,"about_ca_system_score_gemma":0.00043327556,"threshold_uncertainty_score":0.99992186},"labels":[],"label_agreement":null},{"id":"W4400432787","doi":"","title":"Perspectives on AI-ML Safety Assurance","year":2024,"lang":"en","type":"article","venue":"HAL (Le Centre pour la Communication Scientifique Directe)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Safran Electronics (Canada)","funders":"","keywords":"Safety assurance; Computer science; Engineering; Reliability engineering","score_opus":0.010536414860755622,"score_gpt":0.25686873259485027,"score_spread":0.24633231773409464,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4400432787","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008699717,0.0026192816,0.8926362,0.056723434,0.0001284282,0.00015985064,0.000010579501,0.0006478957,0.038374607],"genre_scores_gemma":[0.9609086,0.0010097696,0.029529218,0.00015733596,0.000021525439,0.00002252338,0.00001528035,0.00001720399,0.008318559],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99527526,0.0027527,0.00027230158,0.0007951352,0.0005786452,0.0003259311],"domain_scores_gemma":[0.99356,0.0034213758,0.00005492732,0.0017194658,0.0010915153,0.00015272321],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005024058,0.00016530155,0.00018866551,0.00022453224,0.00034902236,0.00088387856,0.0014718361,0.00008329197,0.00013485482],"category_scores_gemma":[0.002156328,0.00015162595,0.0001942531,0.0013107172,0.00018737621,0.0004429498,0.00039914227,0.00041933378,0.00032880958],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000060169314,0.00031491736,0.00059766084,0.000059384158,0.000059945665,0.000013502803,0.009883164,0.000071221315,0.000952487,0.8551682,0.0022621441,0.13061136],"study_design_scores_gemma":[0.00077541795,0.0000036049116,0.012505455,0.002776084,0.00003744137,0.000038341845,0.0007386432,0.679434,0.04665672,0.030844627,0.22512738,0.0010622748],"about_ca_topic_score_codex":0.00020325852,"about_ca_topic_score_gemma":0.00015688321,"teacher_disagreement_score":0.9522089,"about_ca_system_score_codex":0.00014170125,"about_ca_system_score_gemma":0.00024804787,"threshold_uncertainty_score":0.8523261},"labels":[],"label_agreement":null},{"id":"W4401282669","doi":"10.1007/978-3-031-66997-2_17","title":"A Framework for Formal Probabilistic Risk Assessment Using HOL Theorem Proving","year":2024,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"HOL; Computer science; Probabilistic logic; Automated theorem proving; Formal verification; Theoretical computer science; Programming language; Artificial intelligence","score_opus":0.03149742351460235,"score_gpt":0.3305850709995267,"score_spread":0.2990876474849244,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401282669","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00018976416,0.00042642473,0.9957966,0.0005092437,0.0010966077,0.0012254124,0.000020235795,0.00024663025,0.00048908696],"genre_scores_gemma":[0.10146471,0.00003215339,0.8975049,0.00021646838,0.0005494741,0.000055564564,0.0000041133394,0.000045579818,0.00012698362],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9945938,0.00006853378,0.00067438505,0.00210487,0.001458196,0.0011002241],"domain_scores_gemma":[0.9945039,0.0028175707,0.0002980281,0.0016721094,0.0004926866,0.00021569249],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0034926995,0.00056070636,0.0006641817,0.0009558504,0.00060675305,0.0018827156,0.0034551306,0.00043898734,0.000014135796],"category_scores_gemma":[0.0009950971,0.00045702886,0.00038753208,0.0010895985,0.0008232711,0.0008769321,0.0023008792,0.0017975129,0.00001731622],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000075315265,0.000045632663,0.000099485194,0.00036632264,0.00004614079,0.000035993125,0.0005329377,0.07916861,0.000025851823,0.44167423,0.000002686652,0.4779946],"study_design_scores_gemma":[0.00006426011,0.000108145665,0.000017156292,0.00048878376,0.00002213403,0.000014105341,1.9965775e-7,0.5071757,0.000047076624,0.49162474,0.00015650288,0.00028118677],"about_ca_topic_score_codex":0.00003088806,"about_ca_topic_score_gemma":0.000027909911,"teacher_disagreement_score":0.4777134,"about_ca_system_score_codex":0.0008506212,"about_ca_system_score_gemma":0.0015656853,"threshold_uncertainty_score":0.99978817},"labels":[],"label_agreement":null},{"id":"W4401543655","doi":"10.1145/3643991.3645067","title":"Predicting the Impact of Crashes Across Release Channels","year":2024,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Computer science; Environmental science","score_opus":0.026123118609519772,"score_gpt":0.3617541329787454,"score_spread":0.33563101436922566,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401543655","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79827815,0.00058034575,0.19877586,0.0012104551,0.00011218684,0.000087859364,0.0000045367387,0.0002414508,0.0007091816],"genre_scores_gemma":[0.99892676,0.00003875036,0.00037888947,0.000015977157,0.00006642249,0.0000060949633,4.8754373e-7,0.000003620234,0.0005630081],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9989991,0.00007086161,0.00016114199,0.00023222905,0.0002940847,0.00024257443],"domain_scores_gemma":[0.9987207,0.0006539042,0.000017815419,0.00046655827,0.000080145364,0.000060847316],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008572018,0.000068232504,0.00010275489,0.00004522054,0.00011016475,0.0003507066,0.00071891973,0.000030871885,0.00006203407],"category_scores_gemma":[0.0003005147,0.000033419597,0.00022768416,0.0008430043,0.000082325285,0.00029582682,0.00030036137,0.00015053875,0.000047850022],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000026525244,0.00042657033,0.17513613,0.0005161027,0.0011220765,0.000108808665,0.025950141,0.044923272,0.004014162,0.024291715,0.013294669,0.7101898],"study_design_scores_gemma":[0.000039780545,0.00007735586,0.00813418,0.000038920378,0.00000378014,0.0000074942222,0.00012933243,0.98555934,0.0020966926,0.003649471,0.00020274102,0.000060917344],"about_ca_topic_score_codex":0.0011902661,"about_ca_topic_score_gemma":0.000011286957,"teacher_disagreement_score":0.94063604,"about_ca_system_score_codex":0.00003124645,"about_ca_system_score_gemma":0.000110987035,"threshold_uncertainty_score":0.3381872},"labels":[],"label_agreement":null},{"id":"W4402391956","doi":"10.1145/3691642","title":"An Operational Quality Model of Embedded Software Aligned with ISO 25000","year":2024,"lang":"en","type":"article","venue":"ACM Transactions on Embedded Computing Systems","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École de Technologie Supérieure","funders":"","keywords":"Computer science; Quality (philosophy); Software; Software quality; Reliability engineering; Software engineering; Engineering; Software development; Programming language; Physics","score_opus":0.03932399085442554,"score_gpt":0.3357699320259824,"score_spread":0.2964459411715568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402391956","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.073085666,0.00015758042,0.9249498,0.00027180312,0.00029428667,0.0003402455,0.000041186933,0.00078553235,0.000073896044],"genre_scores_gemma":[0.88733715,0.000006646878,0.112276435,0.00006158019,0.000065280576,0.000034376582,0.000015377476,0.00002775651,0.00017538198],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9960741,0.0005343674,0.00080587086,0.0009783357,0.001164674,0.00044268832],"domain_scores_gemma":[0.996247,0.0010379474,0.00012008104,0.0019183556,0.00046871943,0.00020792635],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001518519,0.00030226517,0.00051830243,0.0004251231,0.00040714376,0.0006461771,0.0015966446,0.00015817498,0.000018418037],"category_scores_gemma":[0.00009005201,0.0002510882,0.00023295396,0.0013247133,0.00012701284,0.0007249969,0.000032675503,0.00033395615,0.00003471106],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000033017408,0.00031677698,0.000070815135,0.000312484,0.00022150044,0.000007959684,0.0030404653,0.97151417,0.001836323,0.005098046,0.00004370496,0.017504761],"study_design_scores_gemma":[0.00032404286,0.0002662195,0.000049384904,0.0002906953,0.00003392942,0.000025112799,0.00043140427,0.995221,0.0020918748,0.00093179615,0.000029717625,0.00030482246],"about_ca_topic_score_codex":0.00026364467,"about_ca_topic_score_gemma":0.000019863926,"teacher_disagreement_score":0.8142515,"about_ca_system_score_codex":0.00015008631,"about_ca_system_score_gemma":0.00052948826,"threshold_uncertainty_score":0.99999416},"labels":[],"label_agreement":null},{"id":"W4402416800","doi":"10.24840/978-989-20-8313-1","title":"Proceedings of the 6th International Conference on Integrity-Reliability-Failure (IRF2018)","year":2018,"lang":"en","type":"book","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Reliability engineering; Reliability (semiconductor); Computer science; Forensic engineering; Engineering; Physics","score_opus":0.031114628789009405,"score_gpt":0.28707819316420524,"score_spread":0.25596356437519585,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402416800","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00037258654,0.00001578138,0.010229482,0.010424863,0.00070825557,0.0005558226,0.00003262896,0.00018951729,0.97747105],"genre_scores_gemma":[0.012991529,0.00007995381,0.010912602,0.00046307474,0.00047147373,0.00003533813,0.000013739901,0.00002730037,0.975005],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99647737,0.000056108238,0.00056092476,0.00092227967,0.0016651247,0.00031817146],"domain_scores_gemma":[0.995891,0.00033394902,0.00033700623,0.0010263633,0.002294104,0.00011758464],"candidate_categories":["open_science","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012472263,0.00033189476,0.00047480868,0.00025999075,0.00013134594,0.00031501628,0.00568187,0.0004323552,0.00095614087],"category_scores_gemma":[0.000884506,0.0002022199,0.00044231457,0.00038736477,0.000660436,0.00031874466,0.0015866156,0.0013006044,0.00025659113],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000177917,0.00019095108,0.00060028484,0.0001728028,0.00013959043,8.031569e-7,0.00028075167,0.0000037999896,0.00006243748,0.6529746,0.34123367,0.004322516],"study_design_scores_gemma":[0.00042361155,0.00046343548,0.00035523606,0.0012878418,0.000053166546,0.000012110175,0.000088574314,0.015476196,0.0020665748,0.4196239,0.55939746,0.0007518577],"about_ca_topic_score_codex":0.000039653012,"about_ca_topic_score_gemma":0.000035306617,"teacher_disagreement_score":0.23335068,"about_ca_system_score_codex":0.00032128766,"about_ca_system_score_gemma":0.0009810333,"threshold_uncertainty_score":0.99995714},"labels":[],"label_agreement":null},{"id":"W4402416801","doi":"10.24840/978-989-54756-1-2","title":"Proceedings of the 7th International Conference on Integrity-Reliability-Failure (IRF2020)","year":2020,"lang":"en","type":"book","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Reliability engineering; Reliability (semiconductor); Computer science; Forensic engineering; Engineering; Physics","score_opus":0.033671283190196896,"score_gpt":0.2778249172421859,"score_spread":0.24415363405198898,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402416801","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00014915242,0.00002443538,0.009107283,0.051830094,0.0005060425,0.0006535234,0.000047298865,0.00025932546,0.9374229],"genre_scores_gemma":[0.038856305,0.00021277429,0.015298472,0.0018386866,0.00057991594,0.00006217413,0.00003457609,0.000049771614,0.9430673],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.996485,0.000050156228,0.0005619915,0.00095056137,0.0016832652,0.0002690429],"domain_scores_gemma":[0.9972423,0.0003164073,0.00029709633,0.00071752776,0.0012761415,0.00015053013],"candidate_categories":["open_science"],"consensus_categories":[],"category_scores_codex":[0.00063912675,0.0003274751,0.0005208853,0.00017306437,0.000099188816,0.00030602113,0.005426671,0.0003518776,0.0005429514],"category_scores_gemma":[0.0009420564,0.00020726722,0.0004931246,0.00049114286,0.00033406774,0.00027185841,0.0017209338,0.001833627,0.00017077693],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017265755,0.00012433332,0.00039391388,0.00023690109,0.00014445609,0.0000023139485,0.0003547889,0.000011583055,0.00008238317,0.8171767,0.17703013,0.0044252267],"study_design_scores_gemma":[0.0005656146,0.00046215957,0.00041816375,0.0011594782,0.000082925086,0.000012979685,0.00019014986,0.05404511,0.0019819688,0.33835495,0.60176855,0.0009579277],"about_ca_topic_score_codex":0.000026930935,"about_ca_topic_score_gemma":0.000014775777,"teacher_disagreement_score":0.47882175,"about_ca_system_score_codex":0.00027379463,"about_ca_system_score_gemma":0.0011012356,"threshold_uncertainty_score":0.99995446},"labels":[],"label_agreement":null},{"id":"W4402757116","doi":"10.1016/j.jss.2024.112219","title":"A model-driven formal methods approach to software architectural security vulnerabilities specification and verification","year":2024,"lang":"en","type":"article","venue":"Journal of Systems and Software","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":5,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University; Université du Québec en Outaouais","funders":"","keywords":"Computer science; Formal methods; Software engineering; Software security assurance; Formal specification; Programming language; Model checking; Formal verification; Computer security; Information security; Security service","score_opus":0.04063376757938385,"score_gpt":0.32362344089337886,"score_spread":0.282989673313995,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402757116","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.11258483,0.004037294,0.88259125,0.0002912803,0.00017815497,0.00019916041,0.000009168921,0.000085472326,0.000023371776],"genre_scores_gemma":[0.6629626,0.00012863969,0.3366721,0.00001681858,0.00013265287,0.0000117263735,0.0000021534706,0.000008924373,0.00006436301],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9980115,0.00031052087,0.00057503453,0.0003677819,0.00047654434,0.0002585785],"domain_scores_gemma":[0.9984214,0.00051958516,0.00013322278,0.00033444303,0.00034988803,0.00024149353],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0021726037,0.00016006263,0.00036921375,0.0003603404,0.00019467191,0.000756672,0.00039698236,0.000095747644,0.0000011697034],"category_scores_gemma":[0.0004968489,0.000118223994,0.0001249223,0.00044058074,0.00007728942,0.0009413314,0.00016680773,0.0003802188,0.0000018425474],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012511291,0.00028630072,0.006312944,0.004777264,0.00045843713,0.00003195229,0.056914754,0.12006481,0.0012661743,0.051766664,0.0012056753,0.7567899],"study_design_scores_gemma":[0.00020363902,0.0001987957,0.0019033153,0.0002980107,0.000035693538,0.0005478971,0.00091632205,0.9855504,0.00008434463,0.007809291,0.0022076142,0.00024466324],"about_ca_topic_score_codex":0.000036501402,"about_ca_topic_score_gemma":0.0000017581599,"teacher_disagreement_score":0.8654856,"about_ca_system_score_codex":0.000093490766,"about_ca_system_score_gemma":0.00013232125,"threshold_uncertainty_score":0.7296606},"labels":[],"label_agreement":null},{"id":"W4402927940","doi":"10.3390/math12193014","title":"A Network Reliability Analysis Method for Complex Real-Time Systems: Case Studies in Railway and Maritime Systems","year":2024,"lang":"en","type":"article","venue":"Mathematics","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"National Key Research and Development Program of China","keywords":"Reliability (semiconductor); Reliability engineering; Computer science; Engineering","score_opus":0.07972522513810179,"score_gpt":0.3872078090780016,"score_spread":0.3074825839398998,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4402927940","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.008464798,0.0032936053,0.9864639,0.00029458065,0.00013959483,0.00087167433,0.000031342937,0.00027587934,0.00016460074],"genre_scores_gemma":[0.2143163,0.00034580487,0.7831903,0.00002580784,0.00018556509,0.00054769113,0.000016310707,0.000035998975,0.0013362315],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99716634,0.00051697576,0.0008026916,0.0006808549,0.00038938038,0.000443786],"domain_scores_gemma":[0.99147993,0.0071751033,0.00009900001,0.0008725429,0.0002617825,0.0001116128],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0066974834,0.00022058706,0.000994046,0.00039199754,0.00017293237,0.00061868335,0.0003984928,0.000105739135,0.00000681584],"category_scores_gemma":[0.0008354086,0.00017443746,0.00022390005,0.0024597617,0.00011786606,0.00021910576,0.00035377365,0.00016222772,0.00001830812],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000056240584,0.0012925223,0.0039428454,0.07105413,0.01206642,0.0053245015,0.033064496,0.4458859,0.00029073443,0.38818336,0.026281731,0.012557127],"study_design_scores_gemma":[0.0001166319,0.000044620323,0.00003250761,0.0002623233,0.00032239626,0.00029327258,0.000606259,0.98361886,0.000001500935,0.014092131,0.00042898473,0.00018049455],"about_ca_topic_score_codex":0.0004113503,"about_ca_topic_score_gemma":0.000040937684,"teacher_disagreement_score":0.53773296,"about_ca_system_score_codex":0.0001775886,"about_ca_system_score_gemma":0.000061900944,"threshold_uncertainty_score":0.71133536},"labels":[],"label_agreement":null},{"id":"W4403181934","doi":"10.23919/icn.2024.0016","title":"Automated and Controlled Patch Generation for Enhanced Fixing of Communication Software Vulnerabilities","year":2024,"lang":"en","type":"article","venue":"Intelligent and Converged Networks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Brock University","funders":"National Natural Science Foundation of China","keywords":"Computer science; Software; Software engineering; Operating system","score_opus":0.022527795594571902,"score_gpt":0.2925877064000147,"score_spread":0.2700599108054428,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403181934","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.119841196,0.0106523475,0.8683188,0.0003864681,0.00016229108,0.00036143564,0.0000028460943,0.00025982343,0.000014797069],"genre_scores_gemma":[0.9911683,0.0034420593,0.0049866457,0.00003916734,0.000055928715,0.000109718545,0.000022566373,0.000007338252,0.00016825461],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99884844,0.00016203949,0.000358254,0.00030498736,0.00015126354,0.00017500787],"domain_scores_gemma":[0.99792147,0.0014531922,0.000057446257,0.0002976677,0.00021438164,0.000055829805],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008988371,0.00011293157,0.0002787325,0.00010083996,0.00015479574,0.00022646057,0.00021314851,0.000091877104,0.000014016097],"category_scores_gemma":[0.00025230917,0.000094378745,0.000093963135,0.00025096125,0.000104127015,0.00021618111,0.000118503165,0.00013552197,0.0000012586931],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00028288845,0.00015280517,0.0027035035,0.00063340226,0.0007200118,0.0000027545123,0.0057099396,0.01847972,0.00642982,0.028986318,0.002645621,0.9332532],"study_design_scores_gemma":[0.00032372636,0.00007447508,0.00006001547,0.00008260752,0.000023886867,9.980441e-7,0.00006822525,0.98965126,0.006484123,0.002625391,0.0005056782,0.00009960448],"about_ca_topic_score_codex":0.000044826404,"about_ca_topic_score_gemma":0.000028827697,"teacher_disagreement_score":0.97117156,"about_ca_system_score_codex":0.000023700812,"about_ca_system_score_gemma":0.000042602784,"threshold_uncertainty_score":0.38486534},"labels":[],"label_agreement":null},{"id":"W4403318507","doi":"10.1145/3699711","title":"A Systematic Literature Review on Automated Software Vulnerability Detection Using Machine Learning","year":2024,"lang":"en","type":"review","venue":"ACM Computing Surveys","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":62,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University","funders":"","keywords":"Computer science; Systematic review; Vulnerability (computing); Artificial intelligence; Software; Machine learning; Software engineering; Computer security; Programming language; MEDLINE","score_opus":0.06293239003634729,"score_gpt":0.3830288452040654,"score_spread":0.3200964551677181,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403318507","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.000008062935,0.8890093,0.10553831,0.00002932303,0.00062836113,0.0015189091,0.00001936752,0.0032432966,0.0000050622257],"genre_scores_gemma":[0.00008447512,0.99222666,0.007105054,0.00005371997,0.00016137007,0.000057538706,0.00013645264,0.00009414147,0.00008057955],"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","domain_scores_codex":[0.9755561,0.018500455,0.0020353636,0.0019170417,0.0012109219,0.00078014284],"domain_scores_gemma":[0.9878966,0.0072129685,0.0009028762,0.0032356654,0.0005238933,0.00022796815],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.020993382,0.001003139,0.004088278,0.00085505727,0.00061331835,0.0012730134,0.0032060137,0.00054675265,0.000006952374],"category_scores_gemma":[0.022752278,0.0007113905,0.0016274819,0.0064635132,0.00007089928,0.00026788015,0.0021069096,0.0032285876,0.0003121576],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[2.9960069e-7,0.000041766383,0.000008949351,0.63569564,0.0002860331,0.00007121665,0.000043864286,0.00016787423,1.1195904e-7,0.000007955368,0.000028085014,0.36364824],"study_design_scores_gemma":[0.00004990916,0.00007459685,0.00000405088,0.725173,0.00083743606,0.00024567195,9.4558845e-7,0.2661949,5.4537185e-7,0.00006107976,0.0067651886,0.0005926571],"about_ca_topic_score_codex":0.00008775578,"about_ca_topic_score_gemma":0.000014309441,"teacher_disagreement_score":0.36305556,"about_ca_system_score_codex":0.00070283125,"about_ca_system_score_gemma":0.00038769876,"threshold_uncertainty_score":0.9997638},"labels":[],"label_agreement":null},{"id":"W4403647320","doi":"10.1145/3691621.3694960","title":"6th International Workshop on Automated and verifiable Software sYstem DEvelopment (ASYDE 2024)","year":2024,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Verifiable secret sharing; Computer science; Software engineering; Software development; Software; Operating system; Programming language; Set (abstract data type)","score_opus":0.017595912305175987,"score_gpt":0.28778797092549224,"score_spread":0.27019205862031626,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403647320","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009668739,0.0005739959,0.97567254,0.0015298598,0.0012081316,0.00016547376,0.0000031885343,0.0038712353,0.007306814],"genre_scores_gemma":[0.86317736,0.00008477538,0.11376598,0.00022109695,0.00011605185,0.000066479384,0.00001746462,0.000020105545,0.022530688],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99852586,0.000043575426,0.00020689431,0.0004853746,0.00050859404,0.00022968689],"domain_scores_gemma":[0.999185,0.0003210314,0.000015410731,0.0002900705,0.000081200335,0.0001072714],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00045046466,0.000115849034,0.00012752414,0.00023511967,0.00011595194,0.000754266,0.0005275842,0.00006009164,0.00012718712],"category_scores_gemma":[0.000112680194,0.000089412475,0.00004415327,0.0005765816,0.000029702098,0.00029276568,0.00032508324,0.00016021814,0.00042048586],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000029606827,0.0002499125,0.00372968,0.0009353664,0.00075562665,0.0004191017,0.0017850867,0.004709386,0.00018990408,0.07746869,0.07295874,0.8367689],"study_design_scores_gemma":[0.0001170753,0.000021140804,0.0009046059,0.000397968,0.0000048980437,0.000021571992,0.000110266206,0.92518836,0.0006098561,0.00017273161,0.07226162,0.00018988914],"about_ca_topic_score_codex":0.000024447832,"about_ca_topic_score_gemma":0.0000053899857,"teacher_disagreement_score":0.920479,"about_ca_system_score_codex":0.00019931549,"about_ca_system_score_gemma":0.00017350852,"threshold_uncertainty_score":0.72734046},"labels":[],"label_agreement":null},{"id":"W4403972617","doi":"10.2139/ssrn.5007553","title":"Integrating Failure Modes and Effect Analysis for Technical Language Processing","year":2024,"lang":"en","type":"preprint","venue":"SSRN Electronic Journal","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Trois-Rivières","funders":"","keywords":"Computer science; Natural language processing","score_opus":0.007768599163640487,"score_gpt":0.30839508718690817,"score_spread":0.3006264880232677,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4403972617","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.05641121,0.023396024,0.91799843,0.0015680989,0.00003960142,0.00031046502,0.00000574511,0.0002002728,0.00007014097],"genre_scores_gemma":[0.9813872,0.0007244365,0.017361186,0.00001821552,0.00022244111,0.00005974119,0.000012129955,0.000024242361,0.00019042466],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9961706,0.00024118097,0.0004678813,0.00083048217,0.0005319161,0.0017579693],"domain_scores_gemma":[0.99852616,0.0004580333,0.00019614611,0.00048788078,0.00020020551,0.00013159763],"candidate_categories":["scholarly_communication","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.005880044,0.00033110648,0.0006744397,0.0008184332,0.00030301424,0.0014738638,0.0012276312,0.00031533762,0.000002325136],"category_scores_gemma":[0.00058963645,0.00024327732,0.00073042663,0.0011630792,0.000080048696,0.00018218857,0.0013043884,0.0062288684,0.0000029964112],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000055374774,0.00010546598,0.0038935894,0.0014597279,0.0050238697,0.000035163477,0.0011430823,0.0044201924,0.0016245408,0.102990925,0.000073595256,0.8791745],"study_design_scores_gemma":[0.00025566376,0.0003363199,0.000093650226,0.00034681967,0.0012276089,0.00021340868,0.00081679976,0.59606415,0.00024995345,0.3998877,0.000059617694,0.0004483335],"about_ca_topic_score_codex":0.000078069934,"about_ca_topic_score_gemma":0.0011385783,"teacher_disagreement_score":0.924976,"about_ca_system_score_codex":0.00084141176,"about_ca_system_score_gemma":0.002522503,"threshold_uncertainty_score":0.9995627},"labels":[],"label_agreement":null},{"id":"W4404838900","doi":"10.1142/s0218539324500578","title":"Machine Learning-Based Reliability Evaluation for Software Defect Prediction and Model Validation Assessment","year":2024,"lang":"en","type":"article","venue":"International Journal of Reliability Quality and Safety Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"","keywords":"Reliability engineering; Reliability (semiconductor); Computer science; Software quality; Verification and validation; Model validation; Software; Machine learning; Predictive modelling; Artificial intelligence; Engineering; Software development; Programming language","score_opus":0.03188100837465722,"score_gpt":0.3559982628762541,"score_spread":0.32411725450159684,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404838900","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08103604,0.00070650026,0.91426903,0.0030937707,0.00044947772,0.0002753955,0.000049552404,0.00011102744,0.000009196136],"genre_scores_gemma":[0.9115717,0.00034276964,0.08785224,0.00003361882,0.000107384316,0.00002515471,0.000043027427,0.000011120887,0.000012990186],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9969824,0.00029684033,0.00089095347,0.00044531538,0.0012050639,0.0001794058],"domain_scores_gemma":[0.99533445,0.0026565595,0.00017957955,0.00024760634,0.0014513636,0.00013042832],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.012267638,0.00016809937,0.00029059238,0.00029663762,0.00011605532,0.00034892335,0.0003671598,0.00012529649,0.000007904345],"category_scores_gemma":[0.0057691187,0.00014958507,0.0002732782,0.00022784634,0.000059479113,0.00089405634,0.00011858108,0.0005232318,4.648113e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00008724141,0.00008206582,0.00833646,0.00038097316,0.00012112992,0.0000016743576,0.00019214669,0.9513726,0.00020819451,0.0037658499,0.000011422301,0.035440195],"study_design_scores_gemma":[0.0006015178,0.00014390486,0.0102778645,0.0001637631,0.000057067664,0.000014574656,0.000011060789,0.974974,0.0001818834,0.012550482,0.0008945289,0.00012935637],"about_ca_topic_score_codex":0.000027526105,"about_ca_topic_score_gemma":0.0000026312664,"teacher_disagreement_score":0.83053565,"about_ca_system_score_codex":0.00048500267,"about_ca_system_score_gemma":0.00039383283,"threshold_uncertainty_score":0.69065905},"labels":[],"label_agreement":null},{"id":"W4405846117","doi":"10.1109/seaa64295.2024.00014","title":"Model-Based Reliability, Availability, and Maintainability Analysis for Satellite Systems with Collaborative Maneuvers via Stochastic Games","year":2024,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"","keywords":"Maintainability; Computer science; Reliability (semiconductor); Reliability engineering; Satellite; Distributed computing; Software engineering; Engineering; Aerospace engineering","score_opus":0.015379014910279794,"score_gpt":0.27788855987946554,"score_spread":0.2625095449691857,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4405846117","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.040425844,0.00086522935,0.9547561,0.0018350518,0.000065727676,0.0013972482,0.000045885667,0.00047819904,0.00013076038],"genre_scores_gemma":[0.9609555,0.000022072189,0.038085297,0.000068185116,0.00002172685,0.00036693766,0.000019570747,0.000019188967,0.00044151556],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99591655,0.00036450505,0.0005798049,0.0017375689,0.00078202406,0.0006195253],"domain_scores_gemma":[0.9945153,0.0026107484,0.00007630357,0.0014408623,0.0010358834,0.00032089176],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.002844295,0.00036369817,0.00070904405,0.0006377041,0.0002549542,0.0010906806,0.0007140441,0.00014987425,0.000018977033],"category_scores_gemma":[0.00044473316,0.00026327727,0.0003041741,0.004371933,0.0005579527,0.0006324498,0.0002538942,0.0002613917,0.000011339376],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017570607,0.00026997973,0.0190464,0.0014875171,0.0009844757,0.000015279904,0.00078836345,0.94869876,0.000041686075,0.02002431,0.000104397164,0.00836311],"study_design_scores_gemma":[0.00027918155,0.0003081931,0.0013875954,0.00004059026,0.00035590932,0.0000031259087,0.00020312874,0.9869531,0.000059097423,0.009814057,0.00026154116,0.00033450377],"about_ca_topic_score_codex":0.0002646037,"about_ca_topic_score_gemma":0.0002568056,"teacher_disagreement_score":0.92052966,"about_ca_system_score_codex":0.00042728757,"about_ca_system_score_gemma":0.00072850054,"threshold_uncertainty_score":0.99998194},"labels":[],"label_agreement":null},{"id":"W4406385688","doi":"10.51501/jotnafe.v39i1.144","title":"Resolving Schedule Delay Claims with Forensic Analysis","year":2022,"lang":"en","type":"article","venue":"Journal of the National Academy of Forensic Engineers","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Response Biomedical (Canada)","funders":"","keywords":"Schedule; Forensic science; Psychology; Computer science; Criminology; History; Archaeology","score_opus":0.01863120476250374,"score_gpt":0.273908310249016,"score_spread":0.2552771054865123,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406385688","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.66388184,0.0011718853,0.30233985,0.03152189,0.00029920155,0.0002886541,0.000032064025,0.000057507008,0.00040713724],"genre_scores_gemma":[0.96440524,0.000012464179,0.035074055,0.00033082496,0.00007411199,0.000003632476,7.9595833e-7,0.0000065724357,0.00009228673],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99649477,0.00015032779,0.00051441876,0.00017495493,0.0024539824,0.00021151642],"domain_scores_gemma":[0.99837464,0.00050260633,0.00047521625,0.00011634824,0.00044280398,0.00008836686],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002065465,0.00011323224,0.00032325636,0.00077325205,0.00022553359,0.000033985627,0.0014092455,0.000050882794,0.00003387895],"category_scores_gemma":[0.00034765055,0.00007379865,0.0004974022,0.003393578,0.0001191245,0.00032943572,0.00029387826,0.0008321083,7.6305594e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003984878,0.00006928989,0.003065108,0.000013985582,0.0013686705,0.000004443253,0.00024235537,0.97445935,0.00041491087,0.015538555,0.002590875,0.0021925836],"study_design_scores_gemma":[0.0017918954,0.0005074252,0.049629834,0.00009148462,0.0009463289,0.00064277754,0.00034411493,0.89488417,0.011639632,0.03382786,0.0051999735,0.0004945271],"about_ca_topic_score_codex":0.0000038502653,"about_ca_topic_score_gemma":8.601343e-7,"teacher_disagreement_score":0.30052346,"about_ca_system_score_codex":0.00026092908,"about_ca_system_score_gemma":0.00018820456,"threshold_uncertainty_score":0.36151427},"labels":[],"label_agreement":null},{"id":"W4406657102","doi":"10.1016/b978-0-434-98607-1.50051-0","title":"10.1016/b978-0-434-98607-1.50051-0","year":2000,"lang":"en","type":"book-chapter","venue":"Time to knit","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Computer science","score_opus":0.012723263470418278,"score_gpt":0.20591872035540593,"score_spread":0.19319545688498765,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406657102","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[2.1790902e-7,0.0003857122,0.00044673152,0.0005996997,0.0000041673047,0.00031303603,0.000029259054,0.00042576916,0.9977954],"genre_scores_gemma":[0.0000010082222,0.000003904187,0.0010746127,0.000071166294,0.00026705378,0.000022250739,0.000037246686,0.000057013804,0.9984657],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9966114,0.00006454916,0.0005025423,0.0011081347,0.0010952508,0.0006181116],"domain_scores_gemma":[0.99698627,0.00032420296,0.00011263324,0.0018963357,0.00023326994,0.0004472809],"candidate_categories":["metaepi_narrow","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.00052078575,0.0004809897,0.0006690431,0.00053319085,0.00018220834,0.00036969804,0.0024815048,0.00040830424,0.99028647],"category_scores_gemma":[0.000100074496,0.00046884734,0.00045511796,0.00026431674,0.00012602916,0.00025713776,0.0007496192,0.00067439204,0.99676996],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000160682,0.000031756717,1.5718554e-8,0.000020391035,0.00009044081,0.000039028964,0.0000126169225,0.000058882935,0.0000015336557,0.00044508066,0.10853025,0.8907539],"study_design_scores_gemma":[0.00011520984,0.00015342752,9.836533e-7,0.000098599536,0.00003374036,0.0000127086505,8.4161286e-8,0.0012490106,0.000010342007,0.0026695502,0.9951338,0.000522502],"about_ca_topic_score_codex":0.00003322786,"about_ca_topic_score_gemma":5.9933336e-7,"teacher_disagreement_score":0.89023143,"about_ca_system_score_codex":0.00016370397,"about_ca_system_score_gemma":0.0002477841,"threshold_uncertainty_score":0.9997763},"labels":[],"label_agreement":null},{"id":"W4406857082","doi":"10.1109/tcns.2025.3534459","title":"Worst-Case Integrity Attacks and Resilient State Estimation With Partially Secured Measurements","year":2025,"lang":"en","type":"article","venue":"IEEE Transactions on Control of Network Systems","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"Natural Sciences and Engineering Research Council of Canada; National Natural Science Foundation of China","keywords":"Computer science; State (computer science); Resilience (materials science); Reliability engineering; Estimation; Computer security; Engineering; Algorithm; Systems engineering; Materials science","score_opus":0.02421469377263241,"score_gpt":0.28098299423808265,"score_spread":0.25676830046545024,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406857082","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012847643,0.0003538212,0.9853863,0.0003342032,0.00034686335,0.00058494735,0.00000893204,0.000084262436,0.000053031246],"genre_scores_gemma":[0.996801,0.000020723348,0.0029251159,0.00004153502,0.000016817548,0.000091728165,5.855582e-7,0.0000060640737,0.00009639583],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99795634,0.0003992921,0.00046983044,0.00039684633,0.00049143273,0.0002862811],"domain_scores_gemma":[0.9984046,0.00049989764,0.00013425971,0.0005083338,0.00034694784,0.00010594366],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012409135,0.00016254085,0.00038289482,0.00017697633,0.0002403195,0.00016025569,0.00025573574,0.00007731482,0.0000020994755],"category_scores_gemma":[0.000022050614,0.00012936025,0.000085512336,0.00075493194,0.00008919354,0.00022699096,0.0000031188915,0.00028891966,0.0000031477848],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027972044,0.00017745326,0.0009274261,0.00013211384,0.00038373773,0.00003323204,0.0001686988,0.9647503,0.00006924511,0.0002819789,0.000107096406,0.032689005],"study_design_scores_gemma":[0.0013627299,0.00024114133,0.00028223792,0.00037319775,0.00008518016,0.000033861736,0.000030010722,0.99681467,0.0003008465,0.00026723646,0.000082142906,0.00012674244],"about_ca_topic_score_codex":0.00028975998,"about_ca_topic_score_gemma":0.00062141917,"teacher_disagreement_score":0.9839534,"about_ca_system_score_codex":0.000090272435,"about_ca_system_score_gemma":0.00015685038,"threshold_uncertainty_score":0.5275158},"labels":[],"label_agreement":null},{"id":"W4406916423","doi":"10.1080/0305215x.2025.2450686","title":"Effective prompting with ChatGPT for problem formulation in engineering optimization","year":2025,"lang":"en","type":"article","venue":"Engineering Optimization","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Mathematical optimization; Engineering optimization; Multidisciplinary design optimization; Computer science; Optimization problem; Mathematics; Multidisciplinary approach","score_opus":0.0037620645025599393,"score_gpt":0.22476219727686314,"score_spread":0.2210001327743032,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4406916423","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0012415245,0.000059980557,0.997041,0.00021425074,0.00006292223,0.0010436762,8.00871e-7,0.00029210356,0.000043745604],"genre_scores_gemma":[0.28437847,0.000008977145,0.71509945,0.000011520748,0.000022616385,0.00041191338,0.00002729718,0.000015069897,0.000024693369],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99893475,0.000021025873,0.00024667932,0.000350572,0.00017372998,0.00027322958],"domain_scores_gemma":[0.9992434,0.0002784158,0.00003799111,0.00022653848,0.00017682466,0.000036835772],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00043994232,0.00014992837,0.00017959176,0.00054119685,0.00007092262,0.00015122474,0.00022078978,0.00008116601,0.0000016970964],"category_scores_gemma":[0.00033474012,0.00014622565,0.00004312236,0.0014984909,0.000005396519,0.0006623453,0.000059551774,0.000126608,4.790894e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010330847,0.000027615124,0.00041506047,0.00021487297,0.000021743228,5.1326646e-7,0.00014936311,0.9925402,0.00009126959,0.00406821,0.0000036779004,0.0024571533],"study_design_scores_gemma":[0.0005367276,0.00005539811,0.0004501807,0.00023428579,0.000008433972,8.385507e-7,0.000004482345,0.9977028,0.0007703022,0.0000405168,0.000047699934,0.00014837961],"about_ca_topic_score_codex":0.000009470401,"about_ca_topic_score_gemma":0.0000016712446,"teacher_disagreement_score":0.28313696,"about_ca_system_score_codex":0.000199712,"about_ca_system_score_gemma":0.000050907776,"threshold_uncertainty_score":0.5962909},"labels":[],"label_agreement":null},{"id":"W4407427851","doi":"10.2139/ssrn.5135862","title":"A Protocol for Structured Robustness Reproductions and Replicability Assessments","year":2025,"lang":"en","type":"article","venue":"SSRN Electronic Journal","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Deutsche Forschungsgemeinschaft","keywords":"Robustness (evolution); Protocol (science); Computer science; Medicine; Biology; Pathology","score_opus":0.02425805445412338,"score_gpt":0.3878035321090887,"score_spread":0.3635454776549653,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407427851","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0056378744,0.000054596247,0.96660864,0.003466907,0.00006332933,0.024024231,7.9920477e-7,0.00006427068,0.00007932253],"genre_scores_gemma":[0.57438236,0.00026544256,0.122625485,0.0003605957,0.0007448906,0.29142588,0.0000063197895,0.000049394916,0.010139638],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99758077,0.00016052868,0.00030781125,0.00057697215,0.00024492675,0.0011290142],"domain_scores_gemma":[0.9986409,0.00015467886,0.00009884867,0.00071881845,0.00031949516,0.00006725996],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0036281368,0.00011393517,0.00018758111,0.00016259376,0.0004828908,0.00025548265,0.0006072155,0.000064610314,0.000004242115],"category_scores_gemma":[0.0006498746,0.0000917518,0.00011449845,0.0006250221,0.00007218348,0.00034721952,0.00016584582,0.0009310001,4.6373185e-7],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001767036,0.00034394822,0.012299821,0.00017371075,0.00041627517,0.0000011486104,0.00006420748,0.00094374164,0.00042282196,0.4812356,0.00051207317,0.5034099],"study_design_scores_gemma":[0.0013267861,0.0003098099,0.0032610528,0.000030687122,0.000023720457,0.000109043016,0.00013055353,0.025537845,0.00039850853,0.9614484,0.007259018,0.00016460396],"about_ca_topic_score_codex":0.000019096007,"about_ca_topic_score_gemma":0.00009798691,"teacher_disagreement_score":0.8439832,"about_ca_system_score_codex":0.0005427133,"about_ca_system_score_gemma":0.0026080913,"threshold_uncertainty_score":0.4626642},"labels":[],"label_agreement":null},{"id":"W4407830460","doi":"10.1109/jbhi.2025.3544560","title":"GPT-Based Automated Induction: Vulnerability Detection in Medical Software","year":2025,"lang":"en","type":"article","venue":"IEEE Journal of Biomedical and Health Informatics","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Brandon University","funders":"Ningbo Municipal Bureau of Science and Technology","keywords":"Computer science; Software; Vulnerability (computing); Computer security; Operating system","score_opus":0.029098518179370866,"score_gpt":0.36749565567955955,"score_spread":0.3383971375001887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407830460","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.09960699,0.00021824172,0.8784659,0.020860206,0.0006375188,0.000104253144,0.0000013212998,0.00009036367,0.000015208812],"genre_scores_gemma":[0.95725286,0.00040755424,0.037989397,0.0041905087,0.00014005248,0.0000054521615,0.00000200758,0.000003364966,0.000008826141],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.996826,0.00019291061,0.0015742324,0.00009654015,0.0010239613,0.0002863388],"domain_scores_gemma":[0.9983685,0.0003923361,0.00036155476,0.00019793253,0.00024660595,0.00043309672],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005782734,0.000098409524,0.0003749211,0.0006971232,0.00015707697,0.000075750104,0.0004495983,0.00019965778,0.000014558485],"category_scores_gemma":[0.0011267464,0.00007198591,0.000077698125,0.0014191323,0.00020912371,0.00038426777,0.000074215444,0.00076987804,0.0000029017438],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000028328843,0.00030306538,0.0032085488,0.00086854165,0.000028114559,0.000012015227,0.0005841966,0.0001702161,0.000007663214,0.000121287965,0.0017309164,0.9929371],"study_design_scores_gemma":[0.0020822822,0.0007238817,0.02141947,0.0007764799,0.000008648795,0.00013806247,0.00018401364,0.9612634,0.00034095976,0.0025480245,0.010363952,0.00015082542],"about_ca_topic_score_codex":0.00006391544,"about_ca_topic_score_gemma":0.00003528881,"teacher_disagreement_score":0.9927863,"about_ca_system_score_codex":0.00022577272,"about_ca_system_score_gemma":0.0019263054,"threshold_uncertainty_score":0.3417183},"labels":[],"label_agreement":null},{"id":"W4408758447","doi":"10.1093/qopen/qoaf004","title":"A protocol for structured robustness reproductions and replicability assessments","year":2025,"lang":"en","type":"article","venue":"Q Open","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"Deutsche Forschungsgemeinschaft","keywords":"Robustness (evolution); Protocol (science); Computer science; Medicine; Biology; Pathology","score_opus":0.0868313040377392,"score_gpt":0.4803263243522517,"score_spread":0.3934950203145125,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4408758447","genre_codex":"methods","genre_gemma":"protocol","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0014494006,0.0000026750965,0.84416497,0.0052984953,0.000046833018,0.14756584,0.000004848796,0.0000809066,0.0013860209],"genre_scores_gemma":[0.012782702,6.332976e-7,0.28617063,0.00020779516,0.000042971766,0.6944075,0.0000053237536,0.0000068113377,0.006375576],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99869734,0.00009894475,0.0001891301,0.0007250671,0.0001306053,0.00015889753],"domain_scores_gemma":[0.99829,0.00015329635,0.000041128395,0.0012708589,0.00019658958,0.000048115668],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0010823937,0.00007232617,0.00015982827,0.00006153361,0.00025762946,0.0005513307,0.0010960478,0.00004113362,0.000017357568],"category_scores_gemma":[0.0005729407,0.000058968973,0.000041164272,0.0005437628,0.000058419355,0.00042662205,0.0010394097,0.00008864488,0.0000011248995],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00041105758,0.0011485374,0.087431915,0.0012761025,0.00033831212,0.000003685493,0.0002722439,0.002243789,0.0008331177,0.16483068,0.041971657,0.6992389],"study_design_scores_gemma":[0.005037187,0.00033745138,0.09622686,0.00017758622,0.00004084718,0.0000101645555,0.000103779836,0.2872507,0.006611946,0.19851713,0.4050746,0.00061176723],"about_ca_topic_score_codex":0.00006248264,"about_ca_topic_score_gemma":0.000018262685,"teacher_disagreement_score":0.6986271,"about_ca_system_score_codex":0.000039790426,"about_ca_system_score_gemma":0.00020276965,"threshold_uncertainty_score":0.5316494},"labels":[],"label_agreement":null},{"id":"W4409218899","doi":"10.32920/28745432.v1","title":"Comparison of Various RIP Software in Regards to Their Resolution Capabilities","year":2025,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Resolution (logic); Software; Computer science; Artificial intelligence; Operating system","score_opus":0.04410717357010306,"score_gpt":0.35643005407175216,"score_spread":0.3123228805016491,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409218899","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04723925,0.0006775397,0.9465774,0.0015819713,0.00032768687,0.0006443176,0.000044522232,0.00028430903,0.0026229874],"genre_scores_gemma":[0.8820307,0.0000641335,0.11517086,0.00013415495,0.000038742026,0.00014077786,0.000023898347,0.00000902585,0.0023876594],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99659306,0.0004410391,0.00086923083,0.0009803831,0.00068383117,0.00043244782],"domain_scores_gemma":[0.9962197,0.0009913363,0.00014571508,0.002029733,0.0004937314,0.00011981187],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00148891,0.00027817365,0.0009030847,0.0008473312,0.00007220361,0.00015244144,0.0019345979,0.00036093587,0.00003667285],"category_scores_gemma":[0.0015476246,0.0002416924,0.0002887354,0.0011696313,0.00011772561,0.0001191425,0.0039472873,0.0008340773,0.00001301468],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015059095,0.0023959302,0.31842276,0.0048012384,0.0004175301,0.000018811452,0.037986115,0.37893826,0.00010680935,0.046052646,0.013961135,0.19674817],"study_design_scores_gemma":[0.0007840054,0.0004883552,0.05750315,0.0023794824,0.00005379808,0.0000033594529,0.0019311296,0.753716,0.008469582,0.16167721,0.01140978,0.0015841342],"about_ca_topic_score_codex":0.0054636253,"about_ca_topic_score_gemma":0.0017569805,"teacher_disagreement_score":0.8347915,"about_ca_system_score_codex":0.00039238355,"about_ca_system_score_gemma":0.0008173499,"threshold_uncertainty_score":0.985593},"labels":[],"label_agreement":null},{"id":"W4409218984","doi":"10.32920/28745432","title":"Comparison of Various RIP Software in Regards to Their Resolution Capabilities","year":2025,"lang":"en","type":"preprint","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Resolution (logic); Software; Computer science; Artificial intelligence; Operating system","score_opus":0.04410717357010306,"score_gpt":0.35643005407175216,"score_spread":0.3123228805016491,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409218984","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.04723925,0.0006775397,0.9465774,0.0015819713,0.00032768687,0.0006443176,0.000044522232,0.00028430903,0.0026229874],"genre_scores_gemma":[0.8820307,0.0000641335,0.11517086,0.00013415495,0.000038742026,0.00014077786,0.000023898347,0.00000902585,0.0023876594],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99659306,0.0004410391,0.00086923083,0.0009803831,0.00068383117,0.00043244782],"domain_scores_gemma":[0.9962197,0.0009913363,0.00014571508,0.002029733,0.0004937314,0.00011981187],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00148891,0.00027817365,0.0009030847,0.0008473312,0.00007220361,0.00015244144,0.0019345979,0.00036093587,0.00003667285],"category_scores_gemma":[0.0015476246,0.0002416924,0.0002887354,0.0011696313,0.00011772561,0.0001191425,0.0039472873,0.0008340773,0.00001301468],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00015059095,0.0023959302,0.31842276,0.0048012384,0.0004175301,0.000018811452,0.037986115,0.37893826,0.00010680935,0.046052646,0.013961135,0.19674817],"study_design_scores_gemma":[0.0007840054,0.0004883552,0.05750315,0.0023794824,0.00005379808,0.0000033594529,0.0019311296,0.753716,0.008469582,0.16167721,0.01140978,0.0015841342],"about_ca_topic_score_codex":0.0054636253,"about_ca_topic_score_gemma":0.0017569805,"teacher_disagreement_score":0.8347915,"about_ca_system_score_codex":0.00039238355,"about_ca_system_score_gemma":0.0008173499,"threshold_uncertainty_score":0.985593},"labels":[],"label_agreement":null},{"id":"W4409495297","doi":"10.1109/icstw64639.2025.10962516","title":"Adaptive Test Healing using LLM/GPT and Reinforcement Learning","year":2025,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Nutrasource","funders":"","keywords":"Reinforcement learning; Computer science; Test (biology); Reinforcement; Artificial intelligence; Engineering; Structural engineering; Biology","score_opus":0.028884838810211186,"score_gpt":0.3116515862055184,"score_spread":0.2827667473953072,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4409495297","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.009240459,0.00021966684,0.9854503,0.0006059673,0.000028708917,0.00007558609,6.6904185e-8,0.00010238914,0.0042768535],"genre_scores_gemma":[0.9635686,0.00006129728,0.033985678,0.00020000078,0.000012954861,0.0000032408552,4.2036655e-7,0.0000021895146,0.0021655732],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99908143,0.000054235687,0.0001636645,0.0002803495,0.00020504578,0.00021526402],"domain_scores_gemma":[0.9991173,0.00048146505,0.00002680664,0.00021981832,0.000094959854,0.0000596505],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004798704,0.00007340087,0.00012222817,0.00015274464,0.00025994386,0.00017570418,0.0002461552,0.000033876862,0.000018264187],"category_scores_gemma":[0.00034863708,0.00006319552,0.00003931136,0.0005960011,0.000045365832,0.0002156928,0.0003988246,0.00016352623,0.000007169746],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002633864,0.00028330908,0.10208141,0.00028445324,0.0002546842,0.000030415558,0.002015199,0.38933867,0.0033289345,0.15042508,0.0005244286,0.35140708],"study_design_scores_gemma":[0.00011591465,0.000063761225,0.00047489346,0.00004054511,0.000005038688,0.0000014045702,0.00012346717,0.99608874,0.0008223967,0.0011379789,0.0010542978,0.00007153874],"about_ca_topic_score_codex":0.00034726446,"about_ca_topic_score_gemma":0.000009113406,"teacher_disagreement_score":0.9543282,"about_ca_system_score_codex":0.00006418937,"about_ca_system_score_gemma":0.00009954072,"threshold_uncertainty_score":0.25770387},"labels":[],"label_agreement":null},{"id":"W4411017475","doi":"10.1145/3721890.3721897","title":"Summary of the 6th International Workshop on Automated and verifiable Software sYstem DEvelopment (ASYDE) co-located with the 39th IEEE/ACM ASE 2024","year":2025,"lang":"en","type":"article","venue":"ACM SIGSOFT Software Engineering Notes","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Verifiable secret sharing; Software engineering; Computer science; Software; Systems engineering; Engineering; Operating system; Programming language","score_opus":0.012769823327909989,"score_gpt":0.2569284274883022,"score_spread":0.2441586041603922,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411017475","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39295557,0.0014483049,0.59569615,0.0035390658,0.0015661708,0.0011225366,0.000059348396,0.003576537,0.000036279373],"genre_scores_gemma":[0.96569943,0.000025996853,0.033465967,0.00015195413,0.000039505485,0.00009418583,0.000019785835,0.00002898084,0.00047417605],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99773276,0.00009696824,0.00040653875,0.000577113,0.0007654322,0.0004212003],"domain_scores_gemma":[0.9734866,0.024219168,0.00012854686,0.0017546572,0.00030857153,0.00010246516],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0006552313,0.00032347994,0.00035136656,0.0002855412,0.00031506657,0.00024555883,0.0028392724,0.00013630457,0.000010013369],"category_scores_gemma":[0.02782862,0.0001997479,0.00010452704,0.0014129578,0.00013282082,0.0002220095,0.00091803237,0.00046371997,0.000012112883],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00024103894,0.0005484247,0.63829523,0.0028295182,0.00198019,0.00009561711,0.0026778348,0.25110137,0.0014019974,0.0015203822,0.01873123,0.080577165],"study_design_scores_gemma":[0.004715932,0.00050729007,0.5720242,0.016196873,0.00041090624,0.00012492009,0.000772005,0.2608672,0.09209921,0.00015335715,0.04878264,0.0033454224],"about_ca_topic_score_codex":0.000064278946,"about_ca_topic_score_gemma":0.000010759335,"teacher_disagreement_score":0.5727439,"about_ca_system_score_codex":0.00022496769,"about_ca_system_score_gemma":0.0003661582,"threshold_uncertainty_score":0.9803604},"labels":[],"label_agreement":null},{"id":"W4411134448","doi":"10.24840/978-972-752-323-8","title":"Proceedings of the 8th International Conference on Integrity-Reliability-Failure (IRF2025)","year":2025,"lang":"en","type":"book","venue":"FEUP eBooks","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Toronto","funders":"","keywords":"Reliability engineering; Reliability (semiconductor); Forensic engineering; Psychology; Computer science; Engineering; Physics","score_opus":0.02489195301957386,"score_gpt":0.2794001358274377,"score_spread":0.25450818280786386,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411134448","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00028418173,0.000024926328,0.001920817,0.0039900397,0.00066957163,0.0005978204,0.000056019795,0.00014686884,0.99230975],"genre_scores_gemma":[0.012062013,0.00002169694,0.0022712601,0.00034353475,0.00017558134,0.000060935912,0.000011423749,0.000016695738,0.98503685],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99676645,0.000068051115,0.0005853725,0.00090259657,0.0013497627,0.0003277434],"domain_scores_gemma":[0.99682075,0.0003919123,0.00032861324,0.0010064982,0.0013541543,0.00009806664],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.00086410553,0.0003574007,0.00054111745,0.00034562024,0.00015889836,0.00029742406,0.0049090385,0.0004499811,0.00011241372],"category_scores_gemma":[0.0006489185,0.00025013648,0.0005087191,0.00021503148,0.0003983851,0.000119055156,0.0016616597,0.0018246509,0.00004701912],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000025573747,0.00010165577,0.00026715526,0.00041132397,0.00018334875,0.0000025647385,0.0005980403,0.0000055666906,0.000091755115,0.9041928,0.07775888,0.016361354],"study_design_scores_gemma":[0.00057843886,0.00022928239,0.00026857285,0.0035402945,0.00009783464,0.000007262663,0.00013159841,0.0060724583,0.0024174605,0.4583869,0.5275164,0.00075346744],"about_ca_topic_score_codex":0.000041707473,"about_ca_topic_score_gemma":0.000029340976,"teacher_disagreement_score":0.44975755,"about_ca_system_score_codex":0.0003890398,"about_ca_system_score_gemma":0.0018330888,"threshold_uncertainty_score":0.9999951},"labels":[],"label_agreement":null},{"id":"W4411337211","doi":"10.1109/sp61157.2025.00191","title":"SV-TrustEval-C: Evaluating Structure and Semantic Reasoning in Large Language Models for Source Code Vulnerability Analysis","year":2025,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Open Text (Canada); University of Ottawa","funders":"","keywords":"Computer science; Programming language; Natural language processing; Code (set theory); Source code; Artificial intelligence; Set (abstract data type)","score_opus":0.022681027767030418,"score_gpt":0.3626329116289783,"score_spread":0.3399518838619479,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411337211","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.414579,0.000231919,0.5844781,0.00031768036,0.0000135468335,0.00020281548,0.000008817528,0.00006519802,0.00010291392],"genre_scores_gemma":[0.95551145,0.0000086368045,0.04374807,0.00014976921,0.000011722332,0.000028001235,0.000011053329,0.000005133122,0.0005261509],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9976922,0.0003467426,0.0003790305,0.0007567746,0.00038490552,0.00044029413],"domain_scores_gemma":[0.99789065,0.0010350123,0.000061714076,0.0007536998,0.00018034971,0.000078590805],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0028781379,0.00015490895,0.00043073847,0.000525344,0.00022761455,0.00024974378,0.00053350796,0.000105075625,0.000034139703],"category_scores_gemma":[0.0012610513,0.00013045088,0.00017761302,0.0024810492,0.00004720278,0.00034394485,0.00041578073,0.00024325245,7.1884716e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000682448,0.00028160884,0.39518905,0.00065576687,0.0010451594,0.0000074147133,0.00688503,0.42511317,0.001155524,0.03361805,0.000092066075,0.13588893],"study_design_scores_gemma":[0.0004643658,0.000021669643,0.009458353,0.000028548851,0.00011721218,6.1217963e-7,0.00034433583,0.9782015,0.0003688374,0.010849478,0.000015605647,0.00012947856],"about_ca_topic_score_codex":0.00045651576,"about_ca_topic_score_gemma":0.0016277665,"teacher_disagreement_score":0.5530883,"about_ca_system_score_codex":0.0000913707,"about_ca_system_score_gemma":0.00012628494,"threshold_uncertainty_score":0.5319632},"labels":[],"label_agreement":null},{"id":"W4411436444","doi":"10.1007/978-3-031-87496-3_13","title":"ExploitabilityBirthMark: An Early Predictor of the Likelihood of Exploitation","year":2025,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec en Outaouais; Sheridan College","funders":"","keywords":"Exploit; Computer science; Vulnerability (computing); Workload; Computer security; ENCODE; Data science; Risk analysis (engineering); Machine learning","score_opus":0.016284124536406652,"score_gpt":0.2614974433642664,"score_spread":0.24521331882785974,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411436444","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.012242797,0.0002131515,0.9854775,0.0003708306,0.0005779398,0.00043101388,0.000012969441,0.000057676905,0.0006160899],"genre_scores_gemma":[0.88074666,0.000041159357,0.11868976,0.00016258615,0.00011266361,0.00001657601,0.0000029531323,0.000016371563,0.00021127549],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9958911,0.00015170251,0.0007759883,0.0011303363,0.0016135991,0.00043728706],"domain_scores_gemma":[0.9950594,0.0011480725,0.00037847445,0.0024768026,0.00081624713,0.00012100681],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001692428,0.000312808,0.00057770463,0.0007295803,0.00014919958,0.00014947813,0.0048640845,0.0002513384,0.000016658836],"category_scores_gemma":[0.00067797536,0.00023167703,0.00027916374,0.0014734229,0.001154677,0.0006995751,0.0016554267,0.0006123444,0.000002441213],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000023425077,0.00025791186,0.0077179736,0.00042874945,0.000054707085,0.0000052914943,0.0036140732,0.010825486,0.0010757584,0.025989389,0.0000128109405,0.94999444],"study_design_scores_gemma":[0.00050954596,0.0007415746,0.01556018,0.0014747429,0.000038633876,0.0000048680645,0.0000026734892,0.38879767,0.017651375,0.5743957,0.0001466508,0.0006763777],"about_ca_topic_score_codex":0.000114665534,"about_ca_topic_score_gemma":0.00010129943,"teacher_disagreement_score":0.94931805,"about_ca_system_score_codex":0.0001727634,"about_ca_system_score_gemma":0.0011806786,"threshold_uncertainty_score":0.9447515},"labels":[],"label_agreement":null},{"id":"W4411656424","doi":"10.51847/m2llulhwsi","title":"10.51847/M2LluLHWSI","year":2000,"lang":"en","type":"article","venue":"Time to knit","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Generalization; Computer science; Software maintenance; Software; Software engineering; Mathematics; Software development; Programming language","score_opus":0.008915289742088406,"score_gpt":0.2140413400058043,"score_spread":0.2051260502637159,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4411656424","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00089265057,0.000057368703,0.00083303667,0.0014751712,0.0000018868321,0.000113764356,0.00000193117,0.0003069364,0.99631727],"genre_scores_gemma":[0.000630378,6.6848077e-7,0.0021679138,0.00008878731,0.000047374124,0.000011589066,0.0000020506802,0.0000059855415,0.9970453],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9987884,0.00007147889,0.00014387966,0.0003355506,0.0003574009,0.00030325417],"domain_scores_gemma":[0.9989511,0.00011970179,0.000012052257,0.0006634412,0.00006685603,0.00018690097],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0003273454,0.0000903116,0.00013741104,0.0000986477,0.00009629473,0.00015418878,0.0009326455,0.00004211901,0.9673238],"category_scores_gemma":[0.000086859916,0.000081922524,0.000084334664,0.000741811,0.000030566134,0.00020981171,0.00014750195,0.00009776886,0.9848089],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000058592036,0.000041175466,7.428583e-7,0.0000024728547,0.000009203417,0.0000056689523,0.0000150780215,0.00018976328,0.000026836171,0.000014307765,0.008181265,0.99150765],"study_design_scores_gemma":[0.00007434788,0.00006779776,0.00007830647,0.000005349222,0.0000026859507,0.0000035824996,1.7554703e-7,0.012470678,0.000045481866,0.00003398351,0.9871101,0.0001075382],"about_ca_topic_score_codex":0.000024484125,"about_ca_topic_score_gemma":9.63614e-8,"teacher_disagreement_score":0.99140006,"about_ca_system_score_codex":0.000032251697,"about_ca_system_score_gemma":0.000045826186,"threshold_uncertainty_score":0.33407035},"labels":[],"label_agreement":null},{"id":"W4412659517","doi":"10.37116/revistaenergia.v22.n1.2025.700","title":"Análisis de Estabilidad Transitoria Utilizando el Concepto de Inercia y Minería de Datos","year":2025,"lang":"es","type":"article","venue":"Revista Técnica Energía","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Petro-Canada","funders":"","keywords":"Humanities; Philosophy","score_opus":0.010585436233582083,"score_gpt":0.32635659496119745,"score_spread":0.31577115872761535,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412659517","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5076733,0.023547592,0.45404828,0.008066899,0.0007385016,0.00084364833,0.00014262162,0.0005991769,0.0043399916],"genre_scores_gemma":[0.98086834,0.006304273,0.009464585,0.00082242215,0.000564811,0.00009528818,0.000034839555,0.000047149497,0.0017982727],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.99324125,0.0016586873,0.0011152698,0.0014785184,0.00085275975,0.0016535171],"domain_scores_gemma":[0.9942199,0.0019196636,0.00021347226,0.0024434163,0.00049716834,0.0007063764],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0031760414,0.0006101894,0.0011650933,0.00044789634,0.0004632002,0.0016083632,0.0029546407,0.0005598473,0.00021481459],"category_scores_gemma":[0.0025457703,0.0006006144,0.0007951869,0.0030255953,0.0004616013,0.00045269684,0.0006759877,0.0008573364,0.000059330963],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00048120247,0.0023753105,0.2233467,0.0047556125,0.0015096641,0.00038794903,0.0037845785,0.0059243063,0.04721632,0.3519259,0.035322234,0.32297024],"study_design_scores_gemma":[0.0025219943,0.00030398695,0.018736068,0.0020449506,0.00078846945,0.00005056685,0.0003538249,0.16680564,0.012431729,0.0036882202,0.7907666,0.0015079385],"about_ca_topic_score_codex":0.0010042338,"about_ca_topic_score_gemma":0.000058323967,"teacher_disagreement_score":0.75544435,"about_ca_system_score_codex":0.0009424729,"about_ca_system_score_gemma":0.002944442,"threshold_uncertainty_score":0.9996445},"labels":[],"label_agreement":null},{"id":"W4412736375","doi":"10.1016/j.cie.2025.111412","title":"Simultaneously anomaly detection and forecasting for predictive maintenance using a zero-cost differentiable architecture search-based network","year":2025,"lang":"en","type":"article","venue":"Computers & Industrial Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Regina","funders":"Natural Sciences and Engineering Research Council of Canada; Conselho Nacional de Desenvolvimento Científico e Tecnológico; Canadian Network for Research and Innovation in Machining Technology, Natural Sciences and Engineering Research Council of Canada","keywords":"Anomaly detection; Zero (linguistics); Anomaly (physics); Architecture; Differentiable function; Predictive maintenance; Computer science; Data mining; Engineering; Reliability engineering; Mathematics; Geography; Physics","score_opus":0.032442288847473556,"score_gpt":0.24436534126975712,"score_spread":0.21192305242228357,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412736375","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.08204035,0.00009358615,0.9163915,0.00011651921,0.00049078994,0.00064675143,0.000005933029,0.00020881183,0.0000057901175],"genre_scores_gemma":[0.93050176,0.0000022222978,0.06906704,0.0000436257,0.00029970185,0.000041439347,0.0000044748886,0.000016344182,0.000023399656],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99815637,0.00008226378,0.00033201286,0.00056275306,0.00023739981,0.0006292105],"domain_scores_gemma":[0.9977629,0.0015463274,0.00006739603,0.0003292499,0.00015925914,0.00013483688],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005786103,0.00022904432,0.00034413594,0.0003059591,0.0003021714,0.0003258648,0.00047572676,0.00017766628,6.5574636e-7],"category_scores_gemma":[0.0005906531,0.00022738447,0.00013659726,0.0010290107,0.000057446046,0.00016833385,0.00029417092,0.000495804,2.452018e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000058223908,0.000016193642,0.00047099532,0.00006785335,0.000071127746,0.0000052106616,0.000038830527,0.9113084,0.0006966915,0.00016860028,0.000028014416,0.08706983],"study_design_scores_gemma":[0.0010802877,0.00010588442,0.00020802589,0.00035377158,0.000024140307,0.0000069819102,0.0000049658765,0.99607444,0.0010305247,0.00025544368,0.0006643101,0.00019123605],"about_ca_topic_score_codex":0.000065541564,"about_ca_topic_score_gemma":0.0000059938757,"teacher_disagreement_score":0.8484614,"about_ca_system_score_codex":0.0002015191,"about_ca_system_score_gemma":0.0001517344,"threshold_uncertainty_score":0.9272469},"labels":[],"label_agreement":null},{"id":"W4412945237","doi":"10.18653/v1/2025.acl-long.657","title":"Code-Switching Red-Teaming: LLM Evaluation for Safety and Multilingual Understanding","year":2025,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Kootenay Association for Science & Technology","funders":"","keywords":"Computer science; Code-switching; Code (set theory); Programming language; Software engineering; Linguistics","score_opus":0.08500175203487303,"score_gpt":0.383077898694948,"score_spread":0.298076146660075,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4412945237","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013693741,0.00012245592,0.9818333,0.0019279649,0.000089918394,0.00032064962,9.323562e-7,0.000100386664,0.0019106605],"genre_scores_gemma":[0.97045004,0.000031520663,0.028937839,0.00009758592,0.000018565424,0.000014874193,0.000003076722,0.000003171515,0.0004433264],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9987397,0.00010321681,0.00022672667,0.00039514154,0.00032675016,0.00020847296],"domain_scores_gemma":[0.9984322,0.0010152934,0.000035259698,0.00029640805,0.00017007765,0.000050746505],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0025888938,0.00008118453,0.00014402857,0.00018365217,0.00035105986,0.0002062077,0.00028620637,0.00005516904,0.000009784171],"category_scores_gemma":[0.0011238956,0.000070889095,0.00006249681,0.00043148847,0.00002950401,0.00024734755,0.00018822201,0.00009822808,0.0000012484883],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006397431,0.00011162197,0.0040904083,0.00013464493,0.00012270879,0.000001100253,0.00111459,0.0020249523,0.0013285442,0.17799646,0.0003203087,0.8126907],"study_design_scores_gemma":[0.0005600927,0.00003908704,0.0006409047,0.000032808886,0.00001965385,6.328522e-7,0.0003698484,0.95995665,0.00086422043,0.03656495,0.00086852896,0.00008262239],"about_ca_topic_score_codex":0.00003463421,"about_ca_topic_score_gemma":0.00007270071,"teacher_disagreement_score":0.9579317,"about_ca_system_score_codex":0.0002262755,"about_ca_system_score_gemma":0.00016952146,"threshold_uncertainty_score":0.28907734},"labels":[],"label_agreement":null},{"id":"W4413149443","doi":"10.1016/j.eng.2025.07.037","title":"Can large language models solve complex engineering issues? Practical applications in reliability systems engineering","year":2025,"lang":"en","type":"article","venue":"Engineering","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":1,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Alberta","funders":"National Natural Science Foundation of China","keywords":"Computer science; Reliability (semiconductor); Complex system; Reliability engineering; Systems engineering; Software engineering; Engineering; Artificial intelligence","score_opus":0.01648185438212156,"score_gpt":0.30314236168240044,"score_spread":0.28666050730027887,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413149443","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.016831353,0.00070947316,0.9803028,0.0008385373,0.00016407328,0.0004367237,0.000010665154,0.0006204229,0.00008593109],"genre_scores_gemma":[0.94339925,0.000031305975,0.056081373,0.00002039586,0.00008865051,0.00025204176,0.000017368402,0.000022937034,0.00008670641],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9978669,0.000035671623,0.000468632,0.00058804994,0.00037563496,0.0006650535],"domain_scores_gemma":[0.99823475,0.00051843765,0.000038047252,0.00095868646,0.00009645443,0.00015359867],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009264686,0.000256858,0.0004004057,0.00051358616,0.000072601244,0.0002307489,0.00069374504,0.00013527992,0.0000055769287],"category_scores_gemma":[0.00062251656,0.00027332705,0.00011732708,0.0016437754,0.000013246603,0.00044908543,0.00037167006,0.0005751441,0.000009526753],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000011371437,0.0000828245,0.00039069256,0.00030876897,0.00003191845,0.000014095417,0.0004099275,0.9164093,0.0011516922,0.080957055,0.000063747175,0.00017884275],"study_design_scores_gemma":[0.00019343116,0.0000080946265,0.0016737941,0.00010115221,0.000008104448,0.000003887717,0.0000844407,0.9933871,0.00031953395,0.00010847623,0.0038764728,0.00023548455],"about_ca_topic_score_codex":0.00033294328,"about_ca_topic_score_gemma":0.000014621438,"teacher_disagreement_score":0.92656785,"about_ca_system_score_codex":0.0003314455,"about_ca_system_score_gemma":0.00009763865,"threshold_uncertainty_score":0.99997187},"labels":[],"label_agreement":null},{"id":"W4413180511","doi":"10.23919/cycon65856.2025.11103589","title":"Vulnerability Patch Verification for Military Software Systems Through AI-Driven Code-Level Rule Generation","year":2025,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Defence Research and Development Canada; McGill University; Queen's University","funders":"","keywords":"Computer science; Code (set theory); Code review; Vulnerability (computing); Static program analysis; Software; Software engineering; Programming language; Software development; Computer security","score_opus":0.07424323189957673,"score_gpt":0.34586518636938024,"score_spread":0.27162195446980353,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413180511","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.02115499,0.00056792266,0.9736111,0.0030044818,0.00042733926,0.00072900514,0.000052257925,0.0003088805,0.0001440164],"genre_scores_gemma":[0.74765706,0.000107649386,0.24841815,0.00065270957,0.00015315431,0.00058628764,0.0001819655,0.000011724296,0.0022312843],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9977192,0.00027213414,0.0004814997,0.0007970194,0.0003997578,0.0003303792],"domain_scores_gemma":[0.99747485,0.00052420556,0.000049656166,0.0011627661,0.0007201973,0.00006830775],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009907801,0.0001652212,0.00028364622,0.00010901301,0.00041766354,0.00019316825,0.000832333,0.00014948552,0.00001664082],"category_scores_gemma":[0.00075866195,0.00014534444,0.00017515376,0.0007375749,0.0000776343,0.000727778,0.00018419002,0.00017017178,0.000030056934],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001504238,0.002225149,0.11331404,0.002783242,0.0007539388,0.0000058095393,0.0028192094,0.080937415,0.010203706,0.3382158,0.1426116,0.30597967],"study_design_scores_gemma":[0.00039132035,0.0000572875,0.004633709,0.000029068877,0.000019063402,8.5745563e-7,0.000061738494,0.97109437,0.0034418325,0.009855324,0.010211464,0.0002039712],"about_ca_topic_score_codex":0.0012230008,"about_ca_topic_score_gemma":0.0002425345,"teacher_disagreement_score":0.8901569,"about_ca_system_score_codex":0.00020978683,"about_ca_system_score_gemma":0.00029253788,"threshold_uncertainty_score":0.59269744},"labels":[],"label_agreement":null},{"id":"W4413238534","doi":"10.1007/978-3-032-00624-0","title":"Availability, Reliability and Security","year":2025,"lang":"en","type":"book","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Royal Holloway, University of London; Universitat Politècnica de València; SBA Research; Università degli Studi del Molise; Universität Stuttgart; Universitatea Politehnica Timisoara; Trakya Üniversitesi; Ruhr-Universität Bochum; North-West University; Università degli Studi di Cagliari; Technische Universität Braunschweig; Tartu Ülikool; Vysoké Učení Technické v Brně; Universiteit Gent; Cochin University of Science and Technology; Sapienza Università di Roma; Rheinische Friedrich-Wilhelms-Universität Bonn; Newcastle University; Technische Universität Darmstadt; Università degli Studi di Trento; Università della Calabria; Université de Toulouse; Università degli Studi di Padova; Università degli Studi di Verona; Universidad Carlos III de Madrid; Oak Ridge National Laboratory; KU Leuven; Tel Aviv University; Università degli Studi Mediterranea di Reggio Calabria; Chalmers Tekniska Högskola; Aalborg Universitet; Universidade de Lisboa; Universität Hamburg; Norges Teknisk-Naturvitenskapelige Universitet; Sungkyunkwan University; Deakin University; University of Macedonia; Universität Regensburg; Institut Polytechnique de Paris; Université Grenoble Alpes; Universitetet i Stavanger; Politechnika Warszawska; Högskolan i Halmstad; Louisiana State University; Kuwait University; Szegedi Tudományegyetem; Hunan University; Universidade de Coimbra; Universität Passau; Universidade da Beira Interior; Università degli Studi di Salerno; Edinburgh Napier University; Michigan Technological University; Universitat Rovira i Virgili; Polytechnique Montréal; Università degli Studi di Milano; DePaul University","keywords":"Computer science; Reliability (semiconductor); Computer security; Reliability engineering","score_opus":0.011896334982947832,"score_gpt":0.27546631425388046,"score_spread":0.2635699792709326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413238534","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00066404534,0.00085285836,0.99219716,0.002093626,0.00067683065,0.00050115096,0.000008655579,0.00022937816,0.0027763247],"genre_scores_gemma":[0.31956333,0.00083683716,0.66715854,0.0044807163,0.0010335553,0.00010573624,0.00003139554,0.00006511025,0.0067247543],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9938943,0.00026622065,0.00069003005,0.0028293203,0.0014487609,0.0008713686],"domain_scores_gemma":[0.99330264,0.0027502787,0.0001639384,0.0029424487,0.00056214223,0.0002785695],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0042325146,0.00050939666,0.00079999206,0.0009880505,0.00037040675,0.00092262356,0.00445597,0.00047753507,0.000022315144],"category_scores_gemma":[0.0018935973,0.0004549193,0.00019682242,0.002594682,0.0019730553,0.0006996609,0.004867577,0.0015317758,0.000025859033],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011547246,0.00021390554,0.0039021668,0.0007300007,0.000026320395,0.00005696649,0.0011983946,0.006978881,0.00001627765,0.005049618,0.0006841282,0.9811318],"study_design_scores_gemma":[0.00019586117,0.00010003827,0.00065153406,0.00028138535,0.000009807114,0.000013956282,1.3826872e-7,0.46298596,0.00025388622,0.5315296,0.0034653684,0.0005124505],"about_ca_topic_score_codex":0.00009975188,"about_ca_topic_score_gemma":0.00012467652,"teacher_disagreement_score":0.9806194,"about_ca_system_score_codex":0.00083312334,"about_ca_system_score_gemma":0.0031743585,"threshold_uncertainty_score":0.99979025},"labels":[],"label_agreement":null},{"id":"W4413240493","doi":"10.1007/978-3-032-00627-1","title":"Availability, Reliability and Security","year":2025,"lang":"en","type":"book","venue":"Lecture notes in computer science","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Royal Holloway, University of London; Universitat Politècnica de València; Universität Rostock; Università degli Studi del Molise; Universität Stuttgart; Universitat Rovira i Virgili; Universitatea Politehnica Timisoara; Trakya Üniversitesi; Università della Calabria; Università degli Studi di Salerno; Università degli Studi di Verona; Ruhr-Universität Bochum; SBA Research; United Arab Emirates University; North-West University; Deakin University; Umeå Universitet; Tartu Ülikool; Vysoké Učení Technické v Brně; Universiteit Gent; Cochin University of Science and Technology; Sapienza Università di Roma; Rheinische Friedrich-Wilhelms-Universität Bonn; Newcastle University; Technische Universität Darmstadt; Università degli Studi di Trento; Universidade de Coimbra; Linköpings Universitet; Kuwait University; Szegedi Tudományegyetem; Hunan University; Université de Toulouse; Università degli Studi di Padova; Aalborg Universitet; Universidade de Lisboa; Universität Hamburg; Norges Teknisk-Naturvitenskapelige Universitet; Sungkyunkwan University; University of Macedonia; Universität Regensburg; Chalmers Tekniska Högskola; Università degli Studi Mediterranea di Reggio Calabria; Institut Polytechnique de Paris; Universidad Carlos III de Madrid; Institut Mines-Télécom; Université Grenoble Alpes; Universitetet i Stavanger; Politechnika Warszawska; Technische Universität Braunschweig; Università degli Studi di Cagliari; Politecnico di Torino; Università degli Studi di Milano; University of South Alabama; Polytechnique Montréal; Universidad de Málaga; Edinburgh Napier University; Michigan Technological University; Högskolan i Halmstad; Louisiana State University; Oak Ridge National Laboratory; KU Leuven; DePaul University; Universität Passau; Universidade da Beira Interior","keywords":"Computer science; Reliability (semiconductor); Computer security; Reliability engineering","score_opus":0.011896334982947832,"score_gpt":0.27546631425388046,"score_spread":0.2635699792709326,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413240493","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00066404534,0.00085285836,0.99219716,0.002093626,0.00067683065,0.00050115096,0.000008655579,0.00022937816,0.0027763247],"genre_scores_gemma":[0.31956333,0.00083683716,0.66715854,0.0044807163,0.0010335553,0.00010573624,0.00003139554,0.00006511025,0.0067247543],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9938943,0.00026622065,0.00069003005,0.0028293203,0.0014487609,0.0008713686],"domain_scores_gemma":[0.99330264,0.0027502787,0.0001639384,0.0029424487,0.00056214223,0.0002785695],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0042325146,0.00050939666,0.00079999206,0.0009880505,0.00037040675,0.00092262356,0.00445597,0.00047753507,0.000022315144],"category_scores_gemma":[0.0018935973,0.0004549193,0.00019682242,0.002594682,0.0019730553,0.0006996609,0.004867577,0.0015317758,0.000025859033],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000011547246,0.00021390554,0.0039021668,0.0007300007,0.000026320395,0.00005696649,0.0011983946,0.006978881,0.00001627765,0.005049618,0.0006841282,0.9811318],"study_design_scores_gemma":[0.00019586117,0.00010003827,0.00065153406,0.00028138535,0.000009807114,0.000013956282,1.3826872e-7,0.46298596,0.00025388622,0.5315296,0.0034653684,0.0005124505],"about_ca_topic_score_codex":0.00009975188,"about_ca_topic_score_gemma":0.00012467652,"teacher_disagreement_score":0.9806194,"about_ca_system_score_codex":0.00083312334,"about_ca_system_score_gemma":0.0031743585,"threshold_uncertainty_score":0.99979025},"labels":[],"label_agreement":null},{"id":"W4413838025","doi":"10.1115/1.4069585","title":"Adaptive Kalman Filter by Reinforcement Learning for Monitoring Aircraft Engines' Performance Against Abrupt Events","year":2025,"lang":"en","type":"article","venue":"Journal of Engineering for Gas Turbines and Power","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Safran Electronics (Canada)","funders":"","keywords":"Kalman filter; Reinforcement learning; Reinforcement; Computer science; Extended Kalman filter; Aeronautics; Artificial intelligence; Control theory (sociology); Engineering; Structural engineering","score_opus":0.00945800043395988,"score_gpt":0.2515829847906222,"score_spread":0.2421249843566623,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4413838025","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.20370445,0.0010187094,0.7941916,0.0003377601,0.0004987651,0.00017578567,0.0000013256193,0.000028963768,0.00004262663],"genre_scores_gemma":[0.98382735,0.0003656014,0.014845477,0.000030352854,0.00012674826,0.000021385948,0.0000019605397,0.0000106596435,0.000770487],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99896795,0.00001099265,0.00036768935,0.00016315852,0.00020794774,0.0002822453],"domain_scores_gemma":[0.9990738,0.00028679037,0.000110660556,0.00013709726,0.00030062447,0.00009103004],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005160472,0.00014745095,0.00025507674,0.00022290704,0.000119255026,0.00007467721,0.00031259257,0.000059116912,0.0000016213086],"category_scores_gemma":[0.00022542765,0.00012189795,0.00017210057,0.00022227167,0.0000109372295,0.00036957007,0.00007724677,0.0002206769,3.6065381e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00017213845,0.00008879442,0.010423035,0.0005688505,0.00046593882,0.0000041166077,0.00052548765,0.9432316,0.003468876,0.00077341817,0.0023385026,0.03793924],"study_design_scores_gemma":[0.0011683112,0.0007137236,0.0027946273,0.0005503921,0.000036678477,0.000008806492,0.00005664382,0.9352013,0.004618721,0.00009980749,0.05450475,0.0002462386],"about_ca_topic_score_codex":0.0000017437953,"about_ca_topic_score_gemma":6.524484e-8,"teacher_disagreement_score":0.7801229,"about_ca_system_score_codex":0.00005920947,"about_ca_system_score_gemma":0.00004049739,"threshold_uncertainty_score":0.49708542},"labels":[],"label_agreement":null},{"id":"W4414015671","doi":"10.1016/j.jss.2025.112610","title":"Model-based dependability and performance analysis for satellite systems with collaborative maintenance maneuvers via stochastic games","year":2025,"lang":"en","type":"article","venue":"Journal of Systems and Software","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Concordia University","funders":"Université Toulouse III - Paul Sabatier","keywords":"Dependability; Computer science; Satellite; Systems engineering; Reliability engineering; Engineering; Operations research; Software engineering; Aerospace engineering","score_opus":0.011838451434125567,"score_gpt":0.25311009328637246,"score_spread":0.2412716418522469,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4414015671","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15835707,0.0037872996,0.8372186,0.00014638426,0.00009096084,0.0003550046,0.00001598689,0.00002296757,0.00000573138],"genre_scores_gemma":[0.9850466,0.00016902012,0.014555381,0.00002800115,0.000021783988,0.000033459863,0.0000016342431,0.000006263318,0.00013785144],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9981437,0.00014861448,0.00060211454,0.00037490885,0.00044835723,0.0002823084],"domain_scores_gemma":[0.99693584,0.0007420086,0.00038891635,0.00036791,0.001402048,0.00016325543],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014295335,0.00019120138,0.0007295974,0.00048895343,0.00023820209,0.00038230995,0.00036283734,0.0000905329,3.8320468e-7],"category_scores_gemma":[0.0002397102,0.0001291958,0.00013469957,0.0012681888,0.00015353462,0.00044146588,0.00006195638,0.00019503948,2.0656253e-7],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000388083,0.000081135026,0.16723254,0.001793958,0.0012170825,0.000014788475,0.00056534674,0.8131848,0.000018386514,0.0007436296,0.000058985413,0.014701269],"study_design_scores_gemma":[0.00079078105,0.00037339405,0.013084608,0.00046021154,0.0002944237,0.000025742012,0.00036039972,0.9840978,0.000013255195,0.00018305624,0.00015420931,0.00016216832],"about_ca_topic_score_codex":0.000052591302,"about_ca_topic_score_gemma":0.000027676157,"teacher_disagreement_score":0.82668954,"about_ca_system_score_codex":0.00012878099,"about_ca_system_score_gemma":0.00039351257,"threshold_uncertainty_score":0.5268452},"labels":[],"label_agreement":null},{"id":"W58149836","doi":"","title":"A Proposal for a Lightweight Rigorous UML-Based Development Method for Reliable Systems","year":2001,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University","funders":"","keywords":"Computer science; Unified Modeling Language; Consistency (knowledge bases); UML tool; Software engineering; Documentation; Extreme programming; Programming language; Applications of UML; Software; Software development; Class diagram; Coding (social sciences); Process (computing); Software development process; Artificial intelligence","score_opus":0.03009346698855788,"score_gpt":0.3254070185507338,"score_spread":0.2953135515621759,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W58149836","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0003709584,0.00015357645,0.99452096,0.0023881113,0.00017640283,0.0014879745,0.0000026677935,0.00030491262,0.0005944054],"genre_scores_gemma":[0.009458665,0.0000056337117,0.9826627,0.00018904611,0.00008712209,0.0013317808,0.000011872606,0.000015532947,0.00623765],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9977428,0.0000925143,0.00043114895,0.00066262344,0.00046982794,0.0006010966],"domain_scores_gemma":[0.9978125,0.00080414535,0.00007953052,0.0006110632,0.00051745214,0.00017530571],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002084814,0.00017774795,0.00033225247,0.00019960226,0.00034486063,0.0003450098,0.0008814576,0.00010556305,0.000018535322],"category_scores_gemma":[0.00025164118,0.00013074455,0.00018242365,0.00064167,0.0000249199,0.00023576712,0.00010892052,0.00008668805,0.000040850322],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005826173,0.002530268,0.0061531942,0.0023882203,0.0007164667,0.00005357217,0.0013871503,0.016769959,0.0019404377,0.46588603,0.05041477,0.45117733],"study_design_scores_gemma":[0.0006918978,0.00016860917,0.000028716217,0.000030946576,0.000011207511,0.000007689217,0.000013181995,0.76089156,0.005360328,0.0030952129,0.2294902,0.0002104605],"about_ca_topic_score_codex":0.00017854813,"about_ca_topic_score_gemma":0.000064930326,"teacher_disagreement_score":0.7441216,"about_ca_system_score_codex":0.00017982407,"about_ca_system_score_gemma":0.0010032414,"threshold_uncertainty_score":0.5331608},"labels":[],"label_agreement":null},{"id":"W6925962417","doi":"10.21227/107d-hk82","title":"CLEAR: A Dataset for Compositional Language and Elementary Acoustic Reasoning","year":2019,"lang":"en","type":"dataset","venue":"IEEE DataPort","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke","funders":"","keywords":"Task (project management); Acoustic model; Code (set theory); Order (exchange); Scheme (mathematics)","score_opus":0.025908271525664432,"score_gpt":0.3388993667689945,"score_spread":0.3129910952433301,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6925962417","genre_codex":"dataset","genre_gemma":"dataset","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"dataset","genre_consensus":"dataset","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00011969725,0.000162377,0.036930006,0.00012870406,0.00025596196,0.00042277764,0.9619477,0.000030696367,0.0000021219907],"genre_scores_gemma":[0.000061587445,0.00014597122,0.0071840086,0.0005745209,0.00022956122,0.0000576714,0.9916957,0.000011953416,0.00003901266],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9975773,0.00008055634,0.00036736368,0.0009235884,0.0006293982,0.0004217826],"domain_scores_gemma":[0.9972495,0.00048380072,0.00016175046,0.0018688212,0.00007631424,0.0001597997],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0008517641,0.0002639038,0.00041849923,0.0001990084,0.00018214113,0.00032213551,0.0016752006,0.00016049454,0.00012377715],"category_scores_gemma":[0.00013481047,0.00025267346,0.00009169747,0.00021907504,0.00008963853,0.00047419866,0.00073507486,0.00040151493,0.00019908565],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002209739,0.00005863049,0.000025332398,0.00029228514,0.00010917596,0.00005721027,0.00001231153,0.00002863656,0.00006444514,0.00000874826,0.99844444,0.0008767134],"study_design_scores_gemma":[0.00032930094,0.00012437203,0.000053841915,0.00009367941,0.00010701023,0.000051698094,0.000021720676,0.0053311805,0.000039186234,0.000033679462,0.9934984,0.0003159377],"about_ca_topic_score_codex":0.00043891175,"about_ca_topic_score_gemma":0.00009530596,"teacher_disagreement_score":0.029748056,"about_ca_system_score_codex":0.00006196558,"about_ca_system_score_gemma":0.00021115376,"threshold_uncertainty_score":0.99999255},"labels":[],"label_agreement":null},{"id":"W6931256440","doi":"10.5281/zenodo.5822995","title":"blue-marble/gridpath: GridPath v0.12.0","year":2022,"lang":"en","type":"other","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Environment and Climate Change Canada","funders":"","keywords":"Limiting; Downtime; Troubleshooting; Process (computing); Table (database)","score_opus":0.02765081010975657,"score_gpt":0.24152723980910173,"score_spread":0.21387642969934517,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6931256440","genre_codex":"other","genre_gemma":"other","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"other","genre_consensus":"other","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00006540283,0.0006772754,0.044699863,0.0012357705,0.00037217993,0.00081663067,0.0005006832,0.0049092546,0.9467229],"genre_scores_gemma":[0.0051966864,0.0023900024,0.005853292,0.000923863,0.0016221827,0.0000012207483,0.008162871,0.021854335,0.9539955],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99478704,0.0011563747,0.00041403362,0.0013038989,0.0015390997,0.0007995759],"domain_scores_gemma":[0.9967943,0.00009243847,0.00025494114,0.0021108727,0.00037157148,0.00037587088],"candidate_categories":["metaepi_narrow","sts","scholarly_communication","open_science","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.001666005,0.00036966187,0.0004363987,0.0010306261,0.0022986776,0.0017644144,0.0061144116,0.00023291285,0.14517358],"category_scores_gemma":[0.0011771213,0.0003877398,0.00023189727,0.0016755663,0.00024234086,0.00028616367,0.007407194,0.0011282098,0.018237308],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000010342229,0.00019320643,0.0000029988616,0.00011042072,0.000110068984,0.00006637857,0.00041217334,0.000045298948,0.000056795696,0.004236723,0.8989886,0.09576701],"study_design_scores_gemma":[0.00032777048,0.00021708819,0.000034454464,0.000045322977,0.000020073165,0.000078056364,0.000094490555,0.0012607275,0.000025033803,0.00022764916,0.99726117,0.00040817342],"about_ca_topic_score_codex":0.00011634691,"about_ca_topic_score_gemma":0.0000023666955,"teacher_disagreement_score":0.12693627,"about_ca_system_score_codex":0.00040058515,"about_ca_system_score_gemma":0.000026648133,"threshold_uncertainty_score":0.9998574},"labels":[],"label_agreement":null},{"id":"W6958165376","doi":"10.60692/wya39-9dj52","title":"COSMOGLOBE DR1 results","year":2023,"lang":"en","type":"article","venue":"Greater South Information System","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Canadian Institute for Theoretical Astrophysics; University of Waterloo; University of Toronto; University of British Columbia","funders":"","keywords":"CMB cold spot; Planck; Cosmic microwave background; Polarization (electrochemistry); Cosmic background radiation; Synchrotron; Galactic plane; Birefringence","score_opus":0.048745617426740113,"score_gpt":0.2501955800094639,"score_spread":0.2014499625827238,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6958165376","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23919678,0.00000282756,0.73673385,0.001316951,0.0008086593,0.0005300676,0.00012639484,0.0041047116,0.017179735],"genre_scores_gemma":[0.99852264,2.2413191e-7,0.0007962159,0.00008996243,0.000041181007,0.000033725966,0.000021156502,0.0000032669452,0.0004916311],"study_design_codex":"qualitative","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99822223,0.000091846334,0.00054493354,0.00020199969,0.0006158293,0.00032315604],"domain_scores_gemma":[0.99869806,0.000033889693,0.00013863186,0.0007800515,0.00023362304,0.000115762625],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0012041143,0.00011040055,0.00016644175,0.00046114816,0.00018533318,0.0005004239,0.0006964159,0.000075840726,0.000004385274],"category_scores_gemma":[0.00013791249,0.00008860598,0.00009592724,0.0016951592,0.000024182305,0.001198544,0.0002713164,0.0000960169,0.010183966],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0002766428,0.000020195956,0.27368775,0.0024878348,0.00047107026,0.00014677565,0.538314,0.012937138,0.000009616545,0.038960807,0.052585535,0.08010263],"study_design_scores_gemma":[0.002881373,0.0001319707,0.16252016,0.00026610243,0.000019874147,0.00006259686,0.011292043,0.7991952,0.0009936765,0.00009291971,0.021773307,0.0007707951],"about_ca_topic_score_codex":0.000009351716,"about_ca_topic_score_gemma":8.29095e-8,"teacher_disagreement_score":0.78625804,"about_ca_system_score_codex":0.000088218156,"about_ca_system_score_gemma":0.0000536168,"threshold_uncertainty_score":0.9905867},"labels":[],"label_agreement":null},{"id":"W6989357711","doi":"","title":"The Application of Subjective Estimates of Effectiveness to Controlling Software Inspections","year":2000,"lang":"en","type":"article","venue":"NPARC","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":true,"route_about_ca":false,"ca_institutions":"","funders":"","keywords":"Software; Key (lock); Automation; Control (management); Measure (data warehouse)","score_opus":0.008884768602547173,"score_gpt":0.2821429677639557,"score_spread":0.27325819916140853,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W6989357711","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.15435731,0.000047244463,0.84458447,0.00016009648,0.000015961516,0.00025454714,0.0000028561558,0.000050239203,0.0005272766],"genre_scores_gemma":[0.98140436,0.0000219717,0.018456459,0.000008261251,0.000010653394,0.000078837424,7.7157193e-7,0.0000032339378,0.000015427266],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9991381,0.00010844921,0.00017510561,0.00019696643,0.00024509753,0.00013623612],"domain_scores_gemma":[0.99747896,0.0017626028,0.000045242326,0.00042159998,0.0002488593,0.00004274027],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005975595,0.000058453414,0.00015985634,0.000064678694,0.00016456551,0.000029954852,0.00047219972,0.000027702254,0.000013218786],"category_scores_gemma":[0.00037537902,0.000041996525,0.00007734349,0.00074237795,0.0000951535,0.0000935078,0.00005452403,0.0000730489,0.00001757194],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000098588884,0.00017403993,0.022351777,0.000066170076,0.00009132151,4.4589345e-7,0.0005440402,0.010178745,0.025292417,0.022797702,0.000013089436,0.91839164],"study_design_scores_gemma":[0.0009880656,0.0005153228,0.18218589,0.00014802376,0.000049053473,0.0000054254892,0.00009196965,0.2753223,0.26021177,0.2784767,0.001694211,0.0003113267],"about_ca_topic_score_codex":0.000108133194,"about_ca_topic_score_gemma":0.000018814468,"teacher_disagreement_score":0.91808033,"about_ca_system_score_codex":0.00003664648,"about_ca_system_score_gemma":0.000050370512,"threshold_uncertainty_score":0.17125686},"labels":[],"label_agreement":null},{"id":"W7035980055","doi":"","title":"American and Canadian Investors Meet Ethiopian Government to Explore Phase 2 of Power Sector Transformation","year":2014,"lang":"en","type":"other","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Government (linguistics); Phase (matter); Power (physics); Transformation (genetics); Public sector; Private sector","score_opus":0.03175943770839744,"score_gpt":0.299434835007008,"score_spread":0.26767539729861056,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7035980055","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.010825271,0.00013077195,0.41182184,0.007227967,0.00026173878,0.0009871772,0.00013333096,0.00027187794,0.56834],"genre_scores_gemma":[0.8351505,0.00026109867,0.040267564,0.0049571553,0.0001968709,0.00012839003,0.00003928721,0.0003018103,0.11869735],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.9985186,0.00009302383,0.00020927185,0.00032892489,0.00059017853,0.00026002253],"domain_scores_gemma":[0.99885243,0.0000546479,0.000083406594,0.00048160672,0.000035916466,0.00049199146],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00029916735,0.00016037759,0.00032139273,0.00032595516,0.000035688485,0.00005848389,0.00048076443,0.00009807778,0.00050728663],"category_scores_gemma":[0.000067698595,0.00013572183,0.00006804769,0.00045465655,0.00009969612,0.000077354605,0.00005325781,0.000111091635,0.00003067017],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000026599268,0.00038754768,0.0015654864,0.00057944946,0.0004923826,0.00001334718,0.020005655,0.00004379241,0.00032126656,0.035326686,0.6591337,0.28210413],"study_design_scores_gemma":[0.00037185493,0.00045375686,0.00026795204,0.00009933156,0.000014814856,0.000001382672,0.00041577427,0.0018745711,0.00042892722,0.00012143218,0.9956,0.00035020968],"about_ca_topic_score_codex":0.105757564,"about_ca_topic_score_gemma":0.20240547,"teacher_disagreement_score":0.8243252,"about_ca_system_score_codex":0.00015603603,"about_ca_system_score_gemma":0.00017645226,"threshold_uncertainty_score":0.90019727},"labels":[],"label_agreement":null},{"id":"W7095822863","doi":"","title":"Application of tabular methods to the specification and verification of a nuclear reactor shutdown system","year":2001,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Shutdown; Suite; Software; Nuclear power; Selection (genetic algorithm); Verification and validation","score_opus":0.023399760528346015,"score_gpt":0.3190532773774816,"score_spread":0.2956535168491356,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7095822863","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.03632827,0.000062148894,0.96125615,0.0014204336,0.000014673483,0.0002506532,7.5284754e-7,0.00004391824,0.0006230042],"genre_scores_gemma":[0.8615076,0.00004217222,0.13837266,0.000020365182,0.00001492311,0.0000142513745,0.0000012257939,0.000002995438,0.000023798775],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99901575,0.00017428347,0.00025577203,0.00023442597,0.00023581876,0.000083950894],"domain_scores_gemma":[0.9986773,0.00019264831,0.00010187412,0.00077977363,0.00020283958,0.000045549048],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012774006,0.00005027106,0.00012578814,0.00008989244,0.00004581553,0.00003224307,0.0004834327,0.00003474127,0.0000068254735],"category_scores_gemma":[0.00013029584,0.00003434917,0.000037209436,0.0007791315,0.00004210339,0.00011879083,0.000090600945,0.000046689038,0.000016074218],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000263393,0.00012240821,0.0031756342,0.00011638596,0.00003579289,2.0943733e-7,0.0015918303,0.0001705712,0.15855262,0.2592858,0.00020962206,0.5767128],"study_design_scores_gemma":[0.00026093877,0.00017145614,0.109771825,0.000047517962,0.0000357446,0.000010611276,0.0016493846,0.7565014,0.04889259,0.0013050615,0.08112678,0.00022669601],"about_ca_topic_score_codex":0.00018139006,"about_ca_topic_score_gemma":0.0000047246067,"teacher_disagreement_score":0.82517934,"about_ca_system_score_codex":0.000030502459,"about_ca_system_score_gemma":0.000015971009,"threshold_uncertainty_score":0.14007184},"labels":[],"label_agreement":null},{"id":"W7096527887","doi":"","title":"International Workshop on Software Measurement (IWSM'01) 137 Montral, Qubec, Canada -- August 28-29, 2001","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Functional requirement; Non-functional requirement; Correctness; Requirements analysis; Software requirements specification; Software; CLARITY; Software architecture; Software requirements","score_opus":0.11556615707686638,"score_gpt":0.3365428672379701,"score_spread":0.22097671016110373,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7096527887","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.015702542,0.0003055466,0.9605549,0.009839868,0.0012830921,0.0003274776,0.000010400962,0.00041953468,0.011556629],"genre_scores_gemma":[0.83637506,0.00057177315,0.123065285,0.008480601,0.0013375115,0.00007108826,0.00006669737,0.00007533188,0.029956637],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99489456,0.000097072276,0.0004991813,0.0007346105,0.0029043357,0.0008702171],"domain_scores_gemma":[0.997441,0.0004834073,0.00009054809,0.0009643658,0.0005669647,0.0004537437],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0020459162,0.0002567374,0.00026731656,0.0002536356,0.00023838515,0.00030536216,0.0018896948,0.00011524111,0.0011314462],"category_scores_gemma":[0.0012251606,0.00021954959,0.00016607779,0.00073001784,0.00006356998,0.00043639253,0.00035082863,0.0004316199,0.00014193181],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0001047377,0.0007650292,0.010676861,0.000037295686,0.0003966763,0.00016072778,0.00022822927,0.0034699792,0.00029275657,0.006426622,0.6963829,0.28105822],"study_design_scores_gemma":[0.0011361676,0.00010282253,0.01917077,0.00016512671,0.000028405746,0.000020059088,0.00018920495,0.016809193,0.0025765635,0.0016712248,0.957097,0.0010334946],"about_ca_topic_score_codex":0.23969418,"about_ca_topic_score_gemma":0.66916347,"teacher_disagreement_score":0.8374896,"about_ca_system_score_codex":0.001296029,"about_ca_system_score_gemma":0.00089622004,"threshold_uncertainty_score":0.99978167},"labels":[],"label_agreement":null},{"id":"W7100661179","doi":"","title":"2.3 – High-Availability Solutions to Common Software Failures","year":2006,"lang":"en","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Software; Information system; Software system; Key (lock); Reliability (semiconductor)","score_opus":0.01734075737018135,"score_gpt":0.2648676183310695,"score_spread":0.24752686096088813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7100661179","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.059800703,0.000059491183,0.9292711,0.009043422,0.00009498565,0.00020233863,0.000006663839,0.0007140184,0.00080730044],"genre_scores_gemma":[0.83866173,0.0000018655471,0.15947604,0.00032800608,0.000068462614,0.000034822566,0.000008026817,0.0000061137025,0.0014149428],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.9979819,0.00014125723,0.00030531923,0.00054876594,0.000518644,0.00050410326],"domain_scores_gemma":[0.99806166,0.0004022769,0.000030178866,0.00113236,0.00020260096,0.00017094894],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0006607772,0.0001348387,0.0002200994,0.00016565199,0.0003337369,0.00033423855,0.0010943701,0.00007116149,0.00022859055],"category_scores_gemma":[0.00028389305,0.000113978866,0.000133998,0.0011077448,0.00008535021,0.00034705806,0.00064565026,0.0001576349,0.0008272439],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001632053,0.0012643905,0.106179364,0.000066555236,0.00006289319,0.00002530463,0.00018600082,0.014290236,0.0012118302,0.61752665,0.14765993,0.11151051],"study_design_scores_gemma":[0.0011077166,0.00055337773,0.36473417,0.00005364741,0.000043131018,0.000024021718,0.00009432132,0.038105723,0.010658641,0.35519496,0.22759171,0.0018385885],"about_ca_topic_score_codex":0.0072061494,"about_ca_topic_score_gemma":0.0018427265,"teacher_disagreement_score":0.778861,"about_ca_system_score_codex":0.00010419278,"about_ca_system_score_gemma":0.00008345651,"threshold_uncertainty_score":0.9999507},"labels":[],"label_agreement":null},{"id":"W7126196663","doi":"10.5281/zenodo.18435371","title":"Applying Agile Methodologies to Embedded Software Development: A Case Study","year":2016,"lang":"en","type":"article","venue":"Open MIND","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact","funders":"","keywords":"Agile software development; Scrum; Extreme programming practices; Unavailability; Agile Unified Process; Firmware; Extreme programming; Software","score_opus":0.2087852583063768,"score_gpt":0.43293630838208214,"score_spread":0.22415105007570535,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7126196663","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.4007834,0.000013835218,0.5978097,0.00019484805,0.000044002398,0.00092673866,0.0000018600475,0.000016719143,0.00020894781],"genre_scores_gemma":[0.27984232,9.5052997e-7,0.7184429,0.000041796895,0.000013648495,0.0003724504,4.159021e-7,0.000006075113,0.0012794439],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978509,0.00039410754,0.0002958191,0.0007153433,0.00037091586,0.00037290793],"domain_scores_gemma":[0.9975039,0.0012561209,0.000054400818,0.000853209,0.0001517476,0.00018062246],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.002494519,0.000141123,0.00027096717,0.00016389738,0.00029561642,0.00053593447,0.0019920452,0.000045931083,0.0004309977],"category_scores_gemma":[0.001651679,0.00009107019,0.00004902771,0.0006917324,0.000034106703,0.0005634871,0.0026297977,0.000086874454,0.0010627736],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000072794146,0.00008505883,0.002233024,0.0000025505833,0.00004271546,0.0007030904,0.0055126003,0.000015123173,0.0003232411,0.0000049799123,0.00009454672,0.9909758],"study_design_scores_gemma":[0.010679832,0.0030435065,0.011475001,0.0007997654,0.00021478333,0.0038037584,0.08120489,0.0019336111,0.22785574,0.007004733,0.6459612,0.0060231816],"about_ca_topic_score_codex":0.00010986035,"about_ca_topic_score_gemma":0.0002721771,"teacher_disagreement_score":0.9849526,"about_ca_system_score_codex":0.00008170639,"about_ca_system_score_gemma":0.00023080708,"threshold_uncertainty_score":0.99971503},"labels":[],"label_agreement":null},{"id":"W7126273757","doi":"10.5281/zenodo.18435372","title":"Applying Agile Methodologies to Embedded Software Development: A Case Study","year":2016,"lang":"en","type":"article","venue":"Zenodo (CERN European Organization for Nuclear Research)","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact","funders":"","keywords":"Agile software development; Scrum; Extreme programming practices; Unavailability; Agile Unified Process; Firmware; Extreme programming; Software","score_opus":0.10454425921272248,"score_gpt":0.3259491331514609,"score_spread":0.22140487393873842,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7126273757","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.079455085,0.000025812342,0.91576463,0.0005002554,0.000057410864,0.0010049236,0.000015966627,0.0017641125,0.0014117975],"genre_scores_gemma":[0.843279,0.000015364427,0.15471719,0.00020168502,0.00008373997,0.000002275688,0.000037896047,0.0006716047,0.0009912718],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99666625,0.0010366358,0.0003423085,0.0007781205,0.00064032956,0.00053633546],"domain_scores_gemma":[0.9975544,0.00033622832,0.00007516758,0.0009824013,0.0007519835,0.00029978825],"candidate_categories":["sts","scholarly_communication","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0026729077,0.00016993703,0.00022065316,0.00045025384,0.0024596506,0.001061788,0.0022823547,0.000053542804,0.001567357],"category_scores_gemma":[0.004857382,0.0001317543,0.00006278156,0.00138447,0.000091435955,0.00050137995,0.0041317227,0.0001832154,0.0058350735],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017742728,0.00015456094,0.000035878893,0.000023087527,0.00006453477,0.00038356736,0.005663589,0.00003832213,0.0016490456,0.0003719567,0.0065943496,0.98500335],"study_design_scores_gemma":[0.0009953508,0.00059678685,0.00055589393,0.000063607,0.000016762464,0.0012529104,0.0044856546,0.0002991207,0.003400598,0.0008551176,0.98693395,0.000544267],"about_ca_topic_score_codex":0.000026874266,"about_ca_topic_score_gemma":0.0000019836743,"teacher_disagreement_score":0.9844591,"about_ca_system_score_codex":0.00021126363,"about_ca_system_score_gemma":0.000014515678,"threshold_uncertainty_score":0.9999752},"labels":[],"label_agreement":null},{"id":"W7130321097","doi":"10.69894/20190106","title":"Software, hardware short takes","year":2019,"lang":"en","type":"article","venue":"Oil IT Journal.","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Flow (mathematics); Software; Set (abstract data type)","score_opus":0.02069827051356022,"score_gpt":0.2799031839506562,"score_spread":0.259204913437096,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7130321097","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.103872985,0.0031460244,0.8150666,0.015060301,0.0025385672,0.00016069229,0.0000072736734,0.00065047183,0.059497133],"genre_scores_gemma":[0.74478847,0.002748438,0.15705606,0.0037119172,0.0010745073,0.00002527271,0.00000936576,0.000077919765,0.090508044],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9978265,0.00013899944,0.00032817357,0.00037176284,0.00087292003,0.0004616146],"domain_scores_gemma":[0.99827844,0.00027967527,0.0000693844,0.0007543448,0.00035383686,0.00026429625],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0011261025,0.00014489863,0.00025956752,0.0002161785,0.00021857617,0.0006526838,0.0015475992,0.000082339444,0.00083508686],"category_scores_gemma":[0.00037697438,0.000116234936,0.00025626484,0.0006085929,0.000040273684,0.00079587597,0.00018273154,0.00054649106,0.0014144892],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000092626115,0.00014149917,0.040001918,0.00006510768,0.00013034555,0.00013878998,0.00049077143,0.00063859863,0.00039297636,0.00085253,0.030404117,0.9267341],"study_design_scores_gemma":[0.0010013318,0.00038156082,0.012348792,0.00037815663,0.000047577112,0.0009765738,0.00040232227,0.021641284,0.002098586,0.0049035684,0.9547553,0.0010649188],"about_ca_topic_score_codex":0.000010188502,"about_ca_topic_score_gemma":0.0000044686913,"teacher_disagreement_score":0.9256692,"about_ca_system_score_codex":0.0000856178,"about_ca_system_score_gemma":0.00019571198,"threshold_uncertainty_score":0.999363},"labels":[],"label_agreement":null},{"id":"W7137121728","doi":"10.1109/phm-xian66756.2025.11427646","title":"How Much Confidence Do We Have for GenAI-Based Reasoningƒ A Statistical Inference Perspective for Reliability Estimation","year":2025,"lang":"","type":"article","venue":"","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Future Earth","funders":"","keywords":"Reliability (semiconductor); Perspective (graphical); Fiducial inference; Statistical inference; Estimation; Inference; Interval estimation; Point estimation","score_opus":0.03371601620258263,"score_gpt":0.37813023416168384,"score_spread":0.3444142179591012,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7137121728","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0005017958,0.0009231784,0.9435816,0.050365504,0.00030896938,0.00351149,0.00028536204,0.00018387403,0.00033823957],"genre_scores_gemma":[0.6362447,0.00011555938,0.36089304,0.00019417398,0.000052090392,0.0008000507,0.000023148983,0.000015041451,0.0016621875],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99350965,0.0005998923,0.00094305136,0.0026028347,0.0010997774,0.0012447734],"domain_scores_gemma":[0.9696109,0.021487517,0.00030976176,0.001986675,0.0061830026,0.0004221206],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0035604315,0.0005882162,0.0009812426,0.0004931531,0.0010888453,0.0026848998,0.0017237938,0.00042417098,0.00012542927],"category_scores_gemma":[0.05176446,0.0005489454,0.0005864285,0.0014782256,0.0011358768,0.0009579904,0.0004410957,0.00057928014,0.0000147595365],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00043992998,0.0006392992,0.0011443496,0.0015196792,0.00015579638,0.000002992269,0.000787976,0.0023828188,0.00005529845,0.86159194,0.0017855368,0.12949438],"study_design_scores_gemma":[0.0011130769,0.0005609014,0.0004141733,0.00040549843,0.00012547891,7.309835e-7,0.0010728326,0.6666111,0.0019342214,0.32629883,0.0010890627,0.00037407945],"about_ca_topic_score_codex":0.00055513333,"about_ca_topic_score_gemma":0.00018458384,"teacher_disagreement_score":0.66422826,"about_ca_system_score_codex":0.0011943714,"about_ca_system_score_gemma":0.004365326,"threshold_uncertainty_score":0.9996962},"labels":[],"label_agreement":null},{"id":"W7153004372","doi":"10.63282/3117-5481/aijcst-v6i6p108","title":"Integrating Observability, Defect Prediction, and Decision Intelligence for Reliable AI-Driven Software Systems","year":2024,"lang":"","type":"article","venue":"American International Journal of Computer Science and Technology","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Software quality; Software system; Software quality control; Software; Software quality analyst; Software development; Software architecture; Verification and validation; Debugging","score_opus":0.015694975163232976,"score_gpt":0.321151208990072,"score_spread":0.305456233826839,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7153004372","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0937671,0.004051044,0.8918946,0.006520357,0.0033626175,0.00027053565,0.000019664856,0.00010971244,0.0000043741948],"genre_scores_gemma":[0.7859908,0.0019273965,0.21149425,0.00017531426,0.00037542262,0.000015000252,9.682632e-7,0.000012918714,0.000007973617],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9952732,0.00007858745,0.0012837081,0.001080958,0.0017259077,0.00055766833],"domain_scores_gemma":[0.99064237,0.0017076641,0.0005104365,0.0005058806,0.006353516,0.00028011293],"candidate_categories":["metaepi_narrow","sts","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0039836126,0.0003039401,0.0006425636,0.002617521,0.00043902174,0.0029697358,0.0030801797,0.00015072683,0.0000040556383],"category_scores_gemma":[0.002500495,0.00024916767,0.00021515289,0.004157047,0.004005933,0.0022260332,0.0016833508,0.0008283802,0.0000040125024],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00003605315,0.000084100946,0.014451034,0.00008564692,0.0002193596,0.000055828197,0.00040879467,0.002159851,0.0001458109,0.023349844,0.00038543958,0.9586182],"study_design_scores_gemma":[0.0001960691,0.001917258,0.0007812919,0.0011991271,0.00005785395,0.0012434676,0.0002725325,0.9515503,0.0002764468,0.037345793,0.0049303207,0.00022952585],"about_ca_topic_score_codex":0.00012665722,"about_ca_topic_score_gemma":0.000010147371,"teacher_disagreement_score":0.9583887,"about_ca_system_score_codex":0.00043182026,"about_ca_system_score_gemma":0.0015874807,"threshold_uncertainty_score":0.99999607},"labels":[],"label_agreement":null},{"id":"W7153116829","doi":"10.63282/3050-9262.ijaidsml-v5i3p121","title":"AI-Augmented Software Engineering: A Holistic Approach to Reliability, Security, and Lifecycle Optimization","year":2024,"lang":"","type":"article","venue":"International Journal of Artificial Intelligence Data Science and Machine Learning","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Software quality; Software development; Quality (philosophy); Vulnerability (computing); Implementation; Model-driven architecture; Software system; Software; System lifecycle","score_opus":0.0680045401483019,"score_gpt":0.3636618901234947,"score_spread":0.2956573499751928,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7153116829","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.013213865,0.0018816212,0.97664243,0.006445885,0.0014325646,0.00021167712,0.00008549091,0.000064477346,0.000022012364],"genre_scores_gemma":[0.94248796,0.0018301731,0.054789625,0.00030549688,0.0005065959,0.000003748362,0.000038941867,0.000020471278,0.00001699513],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99417937,0.000192291,0.0011849217,0.0012391077,0.0026658683,0.00053845276],"domain_scores_gemma":[0.99535763,0.00081490417,0.00025607113,0.0006166713,0.0023547364,0.0005999888],"candidate_categories":["metaresearch","metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.008200362,0.00031386176,0.00041392772,0.0016293499,0.00046670472,0.0050575496,0.0038948613,0.00010873505,0.000043175005],"category_scores_gemma":[0.017421098,0.00028212933,0.000097103024,0.0025924705,0.00076000264,0.005047884,0.0037807662,0.0013206592,0.000018478835],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009914259,0.00036462184,0.0012895559,0.00020963035,0.0001665802,0.00011124884,0.0039522923,0.6624794,0.00044025155,0.008440008,0.0001332659,0.322314],"study_design_scores_gemma":[0.00004582907,0.0002788161,0.00012389087,0.00042332913,0.000056125064,0.00030490343,0.00031309674,0.9933483,0.00029911892,0.0018359941,0.0026969644,0.0002736329],"about_ca_topic_score_codex":0.0002598566,"about_ca_topic_score_gemma":0.0000095610285,"teacher_disagreement_score":0.9292741,"about_ca_system_score_codex":0.0002587611,"about_ca_system_score_gemma":0.0009429038,"threshold_uncertainty_score":0.9999631},"labels":[],"label_agreement":null},{"id":"W7153281641","doi":"10.63282/3050-9416.ijaibdcms-v5i4p123","title":"An Initiative-Scale Structure for Reliable AI: Governance-Centrical Architecture for Reliability, Difficult, and Active Policy","year":2024,"lang":"","type":"article","venue":"International Journal of AI BigData Computational and Management Studies","topic":"Software Reliability and Analysis Research","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Artificial Intelligence in Medicine (Canada)","funders":"","keywords":"Interoperability; Enterprise architecture; Enterprise system; Architecture; Software architecture; Enterprise software; Reference architecture; Corporate governance; Software","score_opus":0.027245628665976115,"score_gpt":0.3730370554082643,"score_spread":0.3457914267422882,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W7153281641","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"methods","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.011770516,0.0068727895,0.8671168,0.109554164,0.0015450286,0.0010316919,0.0020366383,0.000040015013,0.000032335225],"genre_scores_gemma":[0.89825296,0.015203336,0.080630004,0.0031976283,0.0021324325,0.00007969816,0.00021919131,0.000038239235,0.0002465305],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9963706,0.00013271246,0.00093949394,0.0008586682,0.0012961199,0.0004024187],"domain_scores_gemma":[0.9944884,0.0016053305,0.00035922063,0.00022517634,0.0031057792,0.00021605565],"candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.00080855883,0.00037694647,0.00060236914,0.0006438067,0.00040818608,0.0011207194,0.0010270062,0.00010456215,0.0000103381135],"category_scores_gemma":[0.0006406078,0.00029662633,0.00025712288,0.00065689284,0.00045147346,0.0016272664,0.0009899914,0.00048902794,0.0000010215024],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.001872538,0.0010291467,0.0013354467,0.0029749644,0.009224035,0.00013379245,0.006097605,0.09348651,0.00003709111,0.26773036,0.033464678,0.5826138],"study_design_scores_gemma":[0.0031063133,0.0014396841,0.012948097,0.001182954,0.00050173927,0.0001591789,0.00084321573,0.42531314,0.00007079934,0.45959765,0.094282,0.0005552345],"about_ca_topic_score_codex":0.000022846736,"about_ca_topic_score_gemma":0.000019407185,"teacher_disagreement_score":0.8864824,"about_ca_system_score_codex":0.0003796296,"about_ca_system_score_gemma":0.00038486012,"threshold_uncertainty_score":0.99994856},"labels":[],"label_agreement":null}]}