{"meta":{"query_hash":"210e62fba4a8","filters":{"venue":"Software Testing Verification and Reliability"},"cohort_total":28,"direct_labels_cover":0,"predictions_cover":28,"exported":28,"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/210e62fba4a8","api":"https://metacan.xera.ac/api/v1/cohort?venue=Software+Testing+Verification+and+Reliability"},"results":[{"id":"W1516438975","doi":"10.1002/stvr.1573","title":"Anomaly detection in performance regression testing by transaction profile estimation","year":2015,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software System Performance and Reliability","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":"York University","funders":"Science Foundation Ireland","keywords":"Computer science; Regression testing; Workload; Anomaly detection; Data mining; Software performance testing; Software regression; Regression analysis; Regression; Non-regression testing; Software; Machine learning; Software quality; Statistics; Software system; Operating system; Software development","score_opus":0.02938539619239948,"score_gpt":0.2524789475198476,"score_spread":0.22309355132744813,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1516438975","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":[0.70728105,0.00008577246,0.29098785,0.00011276387,0.00030304422,0.00041934228,0.0000023848168,0.00067209726,0.00013570212],"genre_scores_gemma":[0.87253284,0.0000075704597,0.1272496,0.000025377085,0.000027671518,0.00009068862,0.000011754308,0.00001160483,0.00004291932],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9978537,0.0001943199,0.0005750292,0.0007147665,0.00036383286,0.00029837113],"domain_scores_gemma":[0.9981153,0.00038225393,0.00026083604,0.00062866317,0.00045256148,0.00016039335],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001929767,0.00022085928,0.00023238319,0.00012839446,0.00027217928,0.00011106956,0.00027562559,0.00019017131,0.0000018057129],"category_scores_gemma":[0.0032999215,0.00018654921,0.000029628076,0.0011521002,0.00010506162,0.0012351539,0.00005406051,0.00029529457,0.000020343958],"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":[0.00003637037,0.00015054783,0.3654985,0.00029514887,0.000002232373,7.6472145e-7,0.00074173155,0.0028548848,0.0018829434,0.0000057138905,0.000094833114,0.6284363],"study_design_scores_gemma":[0.00061781803,0.00027375304,0.35870698,0.0001997264,0.0000067280416,0.000034215795,0.000055033124,0.63324827,0.0058652535,0.000501832,0.00018249745,0.0003078927],"about_ca_topic_score_codex":0.0003442207,"about_ca_topic_score_gemma":0.0000069323164,"teacher_disagreement_score":0.6303934,"about_ca_system_score_codex":0.0002848464,"about_ca_system_score_gemma":0.00017081865,"threshold_uncertainty_score":0.7607256},"labels":[],"label_agreement":null},{"id":"W1522741019","doi":"10.1002/stvr.1572","title":"Coverage‐based regression test case selection, minimization and prioritization: a case study on an industrial system","year":2015,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":95,"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":"Canadian Network for Research and Innovation in Machining Technology, Natural Sciences and Engineering Research Council of Canada; Fonds National de la Recherche Luxembourg","keywords":"Regression testing; Minification; Selection (genetic algorithm); Computer science; Prioritization; Fault detection and isolation; Suite; Reliability engineering; Test suite; Regression; Fault (geology); Regression analysis; Data mining; Machine learning; Test case; Artificial intelligence; Statistics; Engineering; Software; Mathematics; Software system","score_opus":0.08433280682020698,"score_gpt":0.31182946332980477,"score_spread":0.22749665650959777,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1522741019","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":[0.7037818,0.000020362579,0.28737822,0.00009623001,0.00019354532,0.0007524516,0.000006784828,0.007744317,0.000026282263],"genre_scores_gemma":[0.87458813,4.5006996e-7,0.12517887,0.000050924118,0.00008381901,0.00006288978,0.000009478461,0.000016568996,0.000008839556],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99776596,0.00047230677,0.00043641796,0.00082336547,0.00030869368,0.00019328434],"domain_scores_gemma":[0.99650633,0.0014762147,0.00027058314,0.000704745,0.00076851295,0.0002736364],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0015285356,0.00023509393,0.00022375357,0.00015430523,0.00061562215,0.00031899294,0.00016125344,0.00016969464,4.8967865e-7],"category_scores_gemma":[0.011790175,0.00020701981,0.000018318608,0.00081449153,0.00008525265,0.00041012297,0.000080516904,0.0002431698,0.000001325107],"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":[0.00003835779,0.00091036095,0.93376935,0.00010608967,0.000005535681,0.0012100006,0.002284081,0.0010059907,0.000014891988,0.00008063306,0.00088078744,0.059693955],"study_design_scores_gemma":[0.004431493,0.005818601,0.04676836,0.00065484695,0.00010357954,0.03263719,0.0017116934,0.90412277,0.0002554952,0.0021055646,0.00017875119,0.001211632],"about_ca_topic_score_codex":0.0007774481,"about_ca_topic_score_gemma":0.000013085286,"teacher_disagreement_score":0.9031168,"about_ca_system_score_codex":0.00017623932,"about_ca_system_score_gemma":0.00026344787,"threshold_uncertainty_score":0.99653393},"labels":[],"label_agreement":null},{"id":"W1527815941","doi":"10.1002/stvr.1576","title":"Automatic fault localization for client‐side JavaScript","year":2015,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":23,"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 British Columbia","funders":"Natural Sciences and Engineering Research Council of Canada; Intel Corporation","keywords":"JavaScript; Unobtrusive JavaScript; Computer science; Ajax; Scripting language; Web application; Debugging; Programming language; Rich Internet application; Document Object Model; Code (set theory); Source code; Dynamic web page; Tracing; Operating system; World Wide Web; Web service; Web page","score_opus":0.05771615534496858,"score_gpt":0.29183010284763583,"score_spread":0.23411394750266726,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1527815941","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":[0.016833642,0.000109932014,0.96812874,0.0003743882,0.0003077101,0.0004925627,0.000004652573,0.013597507,0.00015083698],"genre_scores_gemma":[0.34872708,0.0000021833878,0.6508001,0.00026037943,0.00003719624,0.00009942232,0.000015824158,0.000013491225,0.000044320637],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99836427,0.00011206006,0.00040655717,0.0005941288,0.00026422864,0.00025875363],"domain_scores_gemma":[0.9969961,0.0010269108,0.00019976265,0.0008111243,0.00077068474,0.00019541418],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0014853051,0.00017511778,0.00019726812,0.00008647987,0.00026148843,0.000176076,0.0003952323,0.00011449634,0.0000010503294],"category_scores_gemma":[0.021419846,0.00016034866,0.00004435334,0.0004928673,0.00011718638,0.0003331118,0.000118968564,0.00010960354,0.000009527867],"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":[0.0000195952,0.00034817532,0.20885241,0.00050391175,0.000014954735,0.0000027411018,0.0024314234,0.0018039865,0.00008382417,0.0034301486,0.04190099,0.74060786],"study_design_scores_gemma":[0.00046121085,0.00022080827,0.017421933,0.00010617373,0.000016985961,0.000025150468,0.000022037197,0.8372286,0.00024125293,0.13964836,0.004269002,0.0003384624],"about_ca_topic_score_codex":0.00012120922,"about_ca_topic_score_gemma":0.0000013549005,"teacher_disagreement_score":0.8354246,"about_ca_system_score_codex":0.00010671016,"about_ca_system_score_gemma":0.00017254466,"threshold_uncertainty_score":0.98682314},"labels":[],"label_agreement":null},{"id":"W1969864380","doi":"10.1002/stvr.241","title":"Editorial: <i>Mutation 2000—A Symposium on Mutation Testing</i>","year":2001,"lang":"en","type":"editorial","venue":"Software Testing Verification and Reliability","topic":"Genetics, Bioinformatics, and Biomedical Research","field":"Biochemistry, Genetics and Molecular Biology","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":"Bell (Canada)","funders":"","keywords":"Citation; Mutation; Library science; Computer science; World Wide Web; Information retrieval; Genetics; Biology","score_opus":0.014995523560344589,"score_gpt":0.281196696684783,"score_spread":0.26620117312443836,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1969864380","genre_codex":"editorial","genre_gemma":"editorial","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"editorial","genre_consensus":"editorial","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0065368526,0.00024813908,0.0033274542,0.0002959601,0.9860494,0.0011149952,0.00039603803,0.00029042273,0.0017406947],"genre_scores_gemma":[0.0025086494,0.0005582625,0.01314227,0.000116700736,0.97674274,0.000118394535,0.0049451133,0.00008544831,0.0017824154],"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","domain_scores_codex":[0.99643093,0.000197885,0.00075844687,0.0009612472,0.001174972,0.00047649906],"domain_scores_gemma":[0.99512553,0.0012316029,0.00047138133,0.000824809,0.002058951,0.00028770574],"candidate_categories":["metaresearch","metaepi_narrow","research_integrity"],"consensus_categories":[],"category_scores_codex":[0.0015286271,0.0004352254,0.00034928357,0.0001191997,0.00032832424,0.00016665063,0.00036243725,0.0015087985,0.000008319182],"category_scores_gemma":[0.046178278,0.0003994057,0.0000947703,0.0003580548,0.00040568408,0.000013495456,0.00014217435,0.00076721044,0.000048424634],"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":[0.00011551918,0.00014808387,0.00038511975,0.00040695036,0.000019992323,0.0000021654982,0.000071739094,0.00014394813,0.0023491743,9.1214724e-7,0.98051494,0.015841454],"study_design_scores_gemma":[0.0007127103,0.0012437965,0.0006522395,0.00018246897,0.00005625954,0.000007841997,0.00003897732,0.00042837762,0.00095666014,0.00031428615,0.9948747,0.00053164654],"about_ca_topic_score_codex":0.0001632216,"about_ca_topic_score_gemma":0.000008477906,"teacher_disagreement_score":0.04464965,"about_ca_system_score_codex":0.00015203466,"about_ca_system_score_gemma":0.0010599823,"threshold_uncertainty_score":0.9998458},"labels":[],"label_agreement":null},{"id":"W1982395839","doi":"10.1002/1099-1689(200009)10:3<149::aid-stvr206>3.0.co;2-t","title":"State generation and automated class testing","year":2000,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":36,"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 Victoria","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Container (type theory); Computer science; White-box testing; Tree (set theory); Class (philosophy); Block (permutation group theory); Java; Reliability (semiconductor); Black box; Code coverage; Reliability engineering; Data mining; Programming language; Engineering; Artificial intelligence; Software; Software development","score_opus":0.03457682983096203,"score_gpt":0.2649357170513301,"score_spread":0.2303588872203681,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1982395839","genre_codex":"empirical","genre_gemma":"methods","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":[0.74745286,0.00018054826,0.20586921,0.00046941874,0.00012226416,0.00037308186,0.000006595544,0.04511643,0.00040956726],"genre_scores_gemma":[0.48002267,0.000012752895,0.5196669,0.0001569893,0.000029466233,0.000022949607,0.0000062540003,0.000010935836,0.00007108366],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9983387,0.0001421461,0.00036543998,0.0006952132,0.00019707308,0.0002614226],"domain_scores_gemma":[0.99780685,0.0009914916,0.0001302211,0.0006429853,0.0002921439,0.0001363296],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008853673,0.0001928843,0.00017702035,0.00006933148,0.0004361436,0.0002682875,0.00024686955,0.00009280637,0.000005619912],"category_scores_gemma":[0.005012273,0.00018022265,0.000019652422,0.00058471883,0.00015607888,0.0003601401,0.00008716468,0.00017891,0.000011654979],"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":[0.0000046620507,0.00007217468,0.116514646,0.0000742427,0.000004572283,0.000003922684,0.00044217033,0.00063666515,0.00088983885,0.000118504955,0.0022621308,0.87897646],"study_design_scores_gemma":[0.00018394992,0.00011316361,0.19686529,0.00006690897,0.0000076105794,0.000067144974,0.0000017956635,0.78770846,0.00027710735,0.013871395,0.0005394077,0.00029778611],"about_ca_topic_score_codex":0.0002536048,"about_ca_topic_score_gemma":0.000002054479,"teacher_disagreement_score":0.8786787,"about_ca_system_score_codex":0.0000475142,"about_ca_system_score_gemma":0.00007980576,"threshold_uncertainty_score":0.73492664},"labels":[],"label_agreement":null},{"id":"W2035512953","doi":"10.1002/stvr.418","title":"Fault‐driven stress testing of distributed real‐time software based on UML models","year":2009,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software System Performance and Reliability","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 Calgary","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Stress testing (software); Unified Modeling Language; Test case; Test (biology); Stress test; Node (physics); Stress (linguistics); Cover (algebra); Real-time computing; Software; Reliability engineering; Programming language; Engineering; Machine learning","score_opus":0.02393212963172253,"score_gpt":0.24975384158457228,"score_spread":0.22582171195284975,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2035512953","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":[0.3717332,0.000041938056,0.6242728,0.0004387101,0.00019524548,0.00076498557,0.00014114879,0.0020610173,0.00035093824],"genre_scores_gemma":[0.73741835,0.0000057872594,0.26233247,0.00008841745,0.000037714482,0.000023724355,0.00006146114,0.000013590257,0.00001845688],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9968071,0.00023920661,0.0008695753,0.0010428755,0.0005887711,0.00045244675],"domain_scores_gemma":[0.9947598,0.0018328286,0.00052078796,0.0016934458,0.00097382773,0.00021933064],"candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0011057814,0.00036372786,0.00050586445,0.00013696749,0.0003952302,0.00010637381,0.0007609984,0.0002449694,0.0000054996726],"category_scores_gemma":[0.0065594525,0.00030900878,0.00011243396,0.0011872682,0.00022462614,0.00053753034,0.00010696759,0.00031468846,0.000013861306],"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":[0.000107089254,0.0013041609,0.45401913,0.00087047386,0.000019819607,0.000007773583,0.00069605414,0.31867334,0.0016232894,0.0004889934,0.00037706312,0.22181281],"study_design_scores_gemma":[0.0005145411,0.0004828806,0.28451404,0.00040573903,0.000018729641,0.0000061687374,0.000013410837,0.7087105,0.0008167946,0.004091346,0.00003099579,0.0003948712],"about_ca_topic_score_codex":0.00015156524,"about_ca_topic_score_gemma":9.2653505e-7,"teacher_disagreement_score":0.39003715,"about_ca_system_score_codex":0.00015569759,"about_ca_system_score_gemma":0.00027913606,"threshold_uncertainty_score":0.9999362},"labels":[],"label_agreement":null},{"id":"W2053464286","doi":"10.1002/stvr.1476","title":"Status and Awards","year":2012,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Labor Movements and Unions","field":"Social Sciences","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":"Business","score_opus":0.03552312542783721,"score_gpt":0.314535663699376,"score_spread":0.2790125382715388,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2053464286","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":[0.99504,0.00030541475,0.0004948131,0.00052178645,0.00020266305,0.00015474627,0.000009401919,0.00019494652,0.003076207],"genre_scores_gemma":[0.9907029,0.00011346935,0.008589012,0.00016607002,0.000093762625,0.000008450692,0.0000043793475,0.0000044578123,0.00031749412],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.999285,0.00009763855,0.000109551765,0.00014214445,0.00012853982,0.00023713322],"domain_scores_gemma":[0.9992614,0.000255426,0.000047056088,0.00013247014,0.00011324277,0.00019042332],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008452899,0.000054518845,0.000063130276,0.000017592778,0.00054224074,0.000043054046,0.00003948348,0.000052785446,0.000036906706],"category_scores_gemma":[0.003282905,0.000046772177,0.000009750552,0.00015998252,0.00025612608,0.0001886394,0.000024922098,0.000067669294,0.0000018056909],"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":[0.0000014150717,0.000044636592,0.94627815,0.0000074476525,0.0000015813906,2.7015401e-8,0.0031258536,5.937089e-7,0.0000341232,0.017168107,0.00012444692,0.033213627],"study_design_scores_gemma":[0.00006780568,0.000009119533,0.9009945,0.000005334679,0.00000620746,1.5373959e-7,0.00055778027,0.0000071970007,0.000008831574,0.0014187789,0.09686091,0.000063433414],"about_ca_topic_score_codex":0.0026073614,"about_ca_topic_score_gemma":0.000056422658,"teacher_disagreement_score":0.09673646,"about_ca_system_score_codex":0.000042679105,"about_ca_system_score_gemma":0.00006500125,"threshold_uncertainty_score":0.41705322},"labels":[],"label_agreement":null},{"id":"W2084111010","doi":"10.1002/stvr.458","title":"An approach for testing pointcut descriptors in AspectJ","year":2011,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Advanced Software Engineering Methodologies","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":"RealNetworks (Canada)","funders":"","keywords":"AspectJ; Computer science; Aspect-oriented programming; Modularity (biology); Set (abstract data type); Programming language; Compile time; Compiler; Software engineering; Base (topology); Test case; Software","score_opus":0.13691961179041653,"score_gpt":0.29543681318094034,"score_spread":0.1585172013905238,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2084111010","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":[0.042773586,0.00005067754,0.9550318,0.0000143898005,0.00015710978,0.000385086,0.0000025213055,0.001447366,0.00013745353],"genre_scores_gemma":[0.27360728,0.0000015922818,0.7262425,0.00002365504,0.000019788395,0.00008151251,0.0000037025898,0.000012927644,0.0000070262213],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983699,0.00015423427,0.00034310936,0.00071110216,0.000116564435,0.00030512552],"domain_scores_gemma":[0.99717724,0.0015869297,0.00012632435,0.0007664767,0.00024112934,0.00010188961],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0016849767,0.00017694753,0.00020615774,0.000104711435,0.00014749533,0.00005282319,0.00049823604,0.00010285217,7.8767107e-7],"category_scores_gemma":[0.020288514,0.00016893832,0.00002796144,0.0006020815,0.00011406073,0.0005339528,0.00008291402,0.00019405955,8.584961e-7],"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":[0.00006173898,0.00084218575,0.5698472,0.0007871458,0.000013052436,0.000005617818,0.0071463115,0.039905466,0.0062539596,0.010635129,0.000036968384,0.3644652],"study_design_scores_gemma":[0.0005087358,0.00037084753,0.71242225,0.00006695178,0.0000102517415,0.000027112306,0.00018263242,0.18556105,0.0021302153,0.09806064,0.00003648475,0.0006228291],"about_ca_topic_score_codex":0.00008584629,"about_ca_topic_score_gemma":9.1351495e-7,"teacher_disagreement_score":0.36384237,"about_ca_system_score_codex":0.00008272285,"about_ca_system_score_gemma":0.000066214874,"threshold_uncertainty_score":0.98796403},"labels":[],"label_agreement":null},{"id":"W2087103218","doi":"10.1002/stvr.295","title":"Eight maxims for software inspectors","year":2004,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Engineering 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":"Royal Military College of Canada","funders":"","keywords":"Cover (algebra); Set (abstract data type); Software; Engineering; Engineering ethics; Software inspection; Computer science; Software engineering; Management science; Software development; Software quality; Mechanical engineering","score_opus":0.027674755534598493,"score_gpt":0.2663009924304416,"score_spread":0.23862623689584309,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2087103218","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":[0.14301471,0.00014085643,0.8537269,0.0005100871,0.00031985686,0.0005179575,0.000017156806,0.0017334513,0.000018969002],"genre_scores_gemma":[0.3684474,0.000007709378,0.631273,0.00006222145,0.000057745438,0.00006908581,0.000016628574,0.000020099913,0.00004611246],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99821025,0.000037817244,0.00031358385,0.0007449359,0.00030494333,0.00038845165],"domain_scores_gemma":[0.99638736,0.0019778528,0.000090209454,0.00093952875,0.0004127153,0.00019230405],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0007521953,0.00019716282,0.00019128759,0.0001146737,0.00031922173,0.00015634288,0.0005889853,0.0001242919,0.0000033727342],"category_scores_gemma":[0.019281983,0.00018373832,0.000063054635,0.0006616428,0.00014311573,0.00034807614,0.00015836624,0.00022245696,0.00002050055],"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":[0.000063135,0.000823482,0.698514,0.001477995,0.000056851084,0.000017505296,0.0027237318,0.014833197,0.0017287809,0.018693658,0.0017136633,0.259354],"study_design_scores_gemma":[0.0023936837,0.00056910666,0.8579366,0.0002305575,0.000025063984,0.00007209215,0.000037110705,0.0059670033,0.0042542284,0.117071725,0.01022953,0.0012132522],"about_ca_topic_score_codex":0.000056507808,"about_ca_topic_score_gemma":0.0000013779514,"teacher_disagreement_score":0.25814074,"about_ca_system_score_codex":0.00020101493,"about_ca_system_score_gemma":0.0002078274,"threshold_uncertainty_score":0.98897904},"labels":[],"label_agreement":null},{"id":"W2111584600","doi":"10.1002/stvr.452","title":"On reducing test length for FSMs with extra states","year":2011,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":38,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Computer Research Institute of Montréal","funders":"","keywords":"Tree traversal; Test suite; Computer science; Reduction (mathematics); Set (abstract data type); Algorithm; Finite-state machine; Test set; Implementation; Test case; Theoretical computer science; Mathematics; Programming language; Artificial intelligence","score_opus":0.04516024558319632,"score_gpt":0.25357394988629844,"score_spread":0.20841370430310213,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2111584600","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":[0.090549774,0.000049612332,0.8972841,0.00018308325,0.00010752066,0.00053192227,0.000008930456,0.010986928,0.00029814022],"genre_scores_gemma":[0.47337148,0.000003107327,0.5263883,0.00009615089,0.000015383539,0.000076873635,0.0000044211783,0.00001359803,0.000030689],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99844545,0.000056993096,0.00029059686,0.00073439995,0.00018151289,0.00029102273],"domain_scores_gemma":[0.9950766,0.0033433703,0.00018554128,0.0008974437,0.00037549902,0.00012155132],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009024243,0.0002079537,0.00018568126,0.00008255292,0.00034507614,0.00009825994,0.00039291842,0.00008361241,0.000002647087],"category_scores_gemma":[0.0084015615,0.00016298954,0.00003429518,0.00035545928,0.00016536728,0.00022571425,0.000064685424,0.00017057738,0.0000043782893],"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":[0.0001802217,0.0015033619,0.36240864,0.00077054097,0.0000318474,0.000012116967,0.0074775824,0.0004601003,0.00041875162,0.00894751,0.006991797,0.6107975],"study_design_scores_gemma":[0.001380897,0.0046967454,0.38658276,0.0010885539,0.000061680636,0.00011713521,0.000054643457,0.10468331,0.0062984596,0.49275807,0.00069672824,0.0015810237],"about_ca_topic_score_codex":0.00019462156,"about_ca_topic_score_gemma":0.0000014272562,"teacher_disagreement_score":0.6092165,"about_ca_system_score_codex":0.000050453356,"about_ca_system_score_gemma":0.00008996765,"threshold_uncertainty_score":0.99995106},"labels":[],"label_agreement":null},{"id":"W2145149095","doi":"10.1002/stvr.461","title":"Regression test suite prioritization using system models","year":2011,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":39,"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":"Regression testing; Test suite; Computer science; Prioritization; Reliability engineering; Test (biology); Test case; Fault detection and isolation; System under test; Test Management Approach; Empirical research; Overhead (engineering); Suite; Regression analysis; Data mining; Machine learning; Artificial intelligence; Engineering; Software system; Statistics; Software; Programming language","score_opus":0.08413484429999112,"score_gpt":0.2648815431189903,"score_spread":0.18074669881899919,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2145149095","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":[0.0458025,0.000106346764,0.9375282,0.000027807388,0.00018786987,0.0003104212,0.0000041162916,0.015486811,0.00054594467],"genre_scores_gemma":[0.51320344,0.0000032658622,0.48671088,0.000022997827,0.00001998205,0.000013243524,0.0000027276828,0.000011049815,0.000012406562],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99827534,0.00012220259,0.00041464617,0.00067572395,0.00025118692,0.00026087253],"domain_scores_gemma":[0.9973971,0.00075215305,0.00027050922,0.0009540851,0.0004930947,0.00013300909],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00094563304,0.00020617987,0.00020435303,0.00011136093,0.0004111127,0.00011467861,0.0004239031,0.00014288144,0.0000014010749],"category_scores_gemma":[0.004143094,0.00017951576,0.00003872257,0.00060351857,0.00011724546,0.0005927942,0.0001847756,0.000172393,0.000004381627],"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":[0.000026582828,0.0005799589,0.8201281,0.0011878352,0.0000141636165,0.000024605102,0.004711619,0.0007955329,0.0018947173,0.013030899,0.0009083586,0.15669763],"study_design_scores_gemma":[0.00022535263,0.0001320338,0.073737025,0.00063932355,0.000022274293,0.00012584185,0.000029024925,0.86453813,0.0010828301,0.058967676,0.000034196262,0.0004662886],"about_ca_topic_score_codex":0.0003525542,"about_ca_topic_score_gemma":6.3134104e-7,"teacher_disagreement_score":0.8637426,"about_ca_system_score_codex":0.000129563,"about_ca_system_score_gemma":0.00010569442,"threshold_uncertainty_score":0.73204404},"labels":[],"label_agreement":null},{"id":"W2155388066","doi":"10.1002/stvr.210","title":"A rigorous method for test templates generation from object‐oriented specifications","year":2001,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":3,"is_retracted":false,"has_abstract":true,"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":"Template; Computer science; Formal specification; Programming language; Focus (optics); Test case; White-box testing; Object-oriented programming; Formal methods; Extension (predicate logic); Specification language; Test (biology); Software engineering; Software; Software development; Software construction","score_opus":0.07291649913695267,"score_gpt":0.30554752011573033,"score_spread":0.23263102097877766,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2155388066","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":[0.020001836,0.00009917531,0.96895957,0.00093385356,0.00023576576,0.0005860906,0.000041481246,0.009017361,0.00012487387],"genre_scores_gemma":[0.18461071,0.000016579317,0.814645,0.00016344404,0.00015691521,0.00020486566,0.00009233974,0.000018466822,0.000091686125],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9979649,0.00012702466,0.00048565672,0.0009225514,0.00021465696,0.00028516588],"domain_scores_gemma":[0.99299747,0.00483635,0.00026337238,0.001105716,0.00065842,0.00013867972],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0010273404,0.00022751692,0.00023516612,0.00009744792,0.000591554,0.00020158249,0.00041330294,0.00014497325,0.000007394136],"category_scores_gemma":[0.014852289,0.00021512365,0.00006315443,0.0006720645,0.00009358101,0.00031531148,0.00009457287,0.00016649939,0.000012584791],"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":[0.00003357093,0.00085294986,0.5234631,0.00008308635,0.000032002317,0.0000060220837,0.0018574463,0.00045712004,0.011207268,0.006298813,0.020853627,0.43485504],"study_design_scores_gemma":[0.0008003493,0.00037048108,0.26755366,0.00010814966,0.000058242334,0.00008491831,0.000029396902,0.56032366,0.0032253067,0.14821608,0.01836441,0.00086532417],"about_ca_topic_score_codex":0.0007342249,"about_ca_topic_score_gemma":0.000013870559,"teacher_disagreement_score":0.55986655,"about_ca_system_score_codex":0.0000925909,"about_ca_system_score_gemma":0.000102074075,"threshold_uncertainty_score":0.99344605},"labels":[],"label_agreement":null},{"id":"W2196997841","doi":"","title":"Eight maxims for software inspectors: Research Articles","year":2004,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Engineering 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":"Royal Military College of Canada","funders":"","keywords":"Cover (algebra); Set (abstract data type); Software; Software inspection; Engineering; Engineering ethics; Computer science; Software engineering; Engineering management; Sociology; Software development; Software quality; Mechanical engineering","score_opus":0.07757743155948835,"score_gpt":0.3266650500084117,"score_spread":0.24908761844892333,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2196997841","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":[0.3133171,0.00020335254,0.6832892,0.00093112467,0.00020174353,0.00066792965,0.000014102557,0.0013557404,0.000019771693],"genre_scores_gemma":[0.51514584,0.000009636501,0.48456377,0.000032548818,0.00007042922,0.00010142462,0.0000095867235,0.00002016871,0.000046588913],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99744,0.00012060345,0.00035931272,0.00087810884,0.0005948228,0.0006071428],"domain_scores_gemma":[0.9927427,0.0047252118,0.00007117191,0.0011850348,0.0010245439,0.00025134208],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0025932852,0.00018573353,0.00019464902,0.00022331979,0.00057281053,0.00027352475,0.00075933733,0.000141001,0.000004400768],"category_scores_gemma":[0.040881757,0.00017168255,0.000057562105,0.0013219595,0.0003104099,0.00042660974,0.00027293258,0.00041120622,0.000038341863],"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":[0.00011479476,0.0013814104,0.59398204,0.0017255828,0.000059603066,0.000027948343,0.0057453937,0.008603259,0.0053446395,0.029877216,0.0031910932,0.349947],"study_design_scores_gemma":[0.0024060481,0.0008830105,0.65824074,0.00035430957,0.000016307105,0.00006473493,0.00013658291,0.0074971677,0.011654833,0.30894276,0.008719991,0.0010834909],"about_ca_topic_score_codex":0.00012133724,"about_ca_topic_score_gemma":0.0000032866851,"teacher_disagreement_score":0.3488635,"about_ca_system_score_codex":0.00028644246,"about_ca_system_score_gemma":0.0002991193,"threshold_uncertainty_score":0.9671973},"labels":[],"label_agreement":null},{"id":"W2461407631","doi":"10.1002/stvr.1609","title":"Prioritizing manual test cases in rapid release environments","year":2016,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":35,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Polytechnique Montréal; University of Manitoba","funders":"Mitacs; Lunds Universitet; University of Manitoba","keywords":"Computer science; Prioritization; Test suite; Unit testing; Agile software development; Test (biology); Code coverage; Suite; Test case; Embedded system; Software engineering; Operating system; Software; Engineering; Machine learning","score_opus":0.02512928207923182,"score_gpt":0.25991368093417877,"score_spread":0.23478439885494695,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2461407631","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":[0.5501346,0.00031900458,0.4361752,0.0015228908,0.00023062665,0.0005999396,0.00001925774,0.010783203,0.00021528397],"genre_scores_gemma":[0.74813664,0.000035932982,0.2515905,0.0000995043,0.000024891955,0.000036647565,0.0000017732972,0.000011297164,0.0000628204],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9982302,0.00012675597,0.00039106605,0.0007185573,0.0002251409,0.000308287],"domain_scores_gemma":[0.9944291,0.0043637524,0.00015102168,0.0008578129,0.000066982095,0.00013134786],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0009180242,0.00018708984,0.00018170635,0.000111350724,0.00017683029,0.000073004376,0.00037493042,0.00010277154,0.000003838044],"category_scores_gemma":[0.021620242,0.00014324764,0.00003234283,0.00035179956,0.00018260283,0.00036759113,0.00018984168,0.00014596601,0.00002054876],"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":[0.000005126714,0.00014634035,0.563756,0.000031202086,0.0000013230953,0.000023772809,0.00013167475,0.0000023958376,0.0016283767,0.00011523919,0.00036488016,0.4337937],"study_design_scores_gemma":[0.00060799374,0.00028475976,0.97283983,0.0004605944,0.000007802855,0.00015752301,0.000008868476,0.0026526884,0.0013076002,0.01956175,0.0015865638,0.00052399625],"about_ca_topic_score_codex":0.00013654062,"about_ca_topic_score_gemma":0.0000023668174,"teacher_disagreement_score":0.4332697,"about_ca_system_score_codex":0.00013633772,"about_ca_system_score_gemma":0.00006676585,"threshold_uncertainty_score":0.9866211},"labels":[],"label_agreement":null},{"id":"W2806282111","doi":"10.1002/stvr.1669","title":"MuMonDE: A framework for evaluating model clone detectors using model mutation analysis","year":2018,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Engineering Research","field":"Computer Science","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Queen's University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"clone (Java method); Preprocessor; Mutation; Computer science; Mutation testing; Data mining; Detector; Precision and recall; Software engineering; Artificial intelligence; Genetics; Biology","score_opus":0.10310354306044087,"score_gpt":0.3755641998763948,"score_spread":0.2724606568159539,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2806282111","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":[0.3954381,0.000022273109,0.6036503,0.000044631506,0.000060492464,0.0002779367,0.000007986967,0.000495792,0.0000025213328],"genre_scores_gemma":[0.48952848,6.732877e-7,0.51035637,0.00002122538,0.000027485112,0.00004111289,0.0000056907925,0.00001132039,0.000007650186],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99793446,0.00008509376,0.00039967406,0.00080153276,0.00042183488,0.00035737923],"domain_scores_gemma":[0.99484915,0.0027706933,0.00018313853,0.000939868,0.0011102059,0.00014694378],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0018944081,0.00018484134,0.00023762259,0.00024707516,0.0004810793,0.00017029507,0.00041455118,0.00015661195,0.0000018397532],"category_scores_gemma":[0.029484497,0.00018490189,0.00009184403,0.0017354087,0.0001550966,0.00033812606,0.0001381276,0.0002145131,0.0000025081617],"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":[0.000017335815,0.00005473328,0.02515528,0.00010695961,0.00004430519,3.1340917e-7,0.001149725,0.9471109,0.002028612,0.0009891633,0.000010585517,0.023332084],"study_design_scores_gemma":[0.00014696817,0.000084643885,0.011497557,0.000029380339,0.00007682997,0.0000023484615,0.0000105422,0.94740266,0.00056404556,0.039987233,0.000001082486,0.00019668721],"about_ca_topic_score_codex":0.000048147078,"about_ca_topic_score_gemma":0.0000017232172,"teacher_disagreement_score":0.09409037,"about_ca_system_score_codex":0.00017786091,"about_ca_system_score_gemma":0.00023643661,"threshold_uncertainty_score":0.97869056},"labels":[],"label_agreement":null},{"id":"W2809981234","doi":"10.1002/stvr.1665","title":"P<scp>esto</scp>: Automated migration of DOM‐based Web tests towards the visual approach","year":2018,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":52,"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":"Computer science; Automation; Web application; Test (biology); Visual Basic; Visualization; Software engineering; Artificial intelligence; Programming language; Software; World Wide Web; Engineering","score_opus":0.033805008270209574,"score_gpt":0.28879608613740154,"score_spread":0.254991077867192,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2809981234","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":[0.45325324,0.00008767116,0.5252904,0.00029232076,0.00017165196,0.0005097213,0.000008175788,0.019831605,0.0005551884],"genre_scores_gemma":[0.624447,0.0000020355706,0.37528732,0.00013805326,0.000045345976,0.000040938274,0.00001029075,0.000011471501,0.00001751686],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99787563,0.00027136545,0.0005006887,0.0006619146,0.0003997045,0.00029068592],"domain_scores_gemma":[0.9952805,0.0022356766,0.00037691428,0.0010822135,0.00091984967,0.00010485825],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0017636828,0.00023218367,0.00024368387,0.00011812059,0.0004341105,0.00014539718,0.0006659018,0.000158467,0.0000011841654],"category_scores_gemma":[0.015660029,0.00016901582,0.000062636434,0.0011242189,0.00060256314,0.00022704934,0.00015941178,0.00021237879,0.0000070869874],"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":[0.000028340512,0.0015668614,0.79991406,0.0007867217,0.000041919702,0.0000032282176,0.004518598,0.00070470106,0.007510026,0.0013641166,0.043706942,0.13985448],"study_design_scores_gemma":[0.0002232311,0.00029149026,0.31080624,0.000080668025,0.000017104152,0.000016864497,0.000016216107,0.6807459,0.0025955704,0.0043788934,0.00070303486,0.00012479036],"about_ca_topic_score_codex":0.00032062488,"about_ca_topic_score_gemma":0.0000049222926,"teacher_disagreement_score":0.6800412,"about_ca_system_score_codex":0.000057545072,"about_ca_system_score_gemma":0.00032530635,"threshold_uncertainty_score":0.9926315},"labels":[],"label_agreement":null},{"id":"W2979327180","doi":"10.1002/stvr.1713","title":"An end‐user‐centric test generation methodology for performance evaluation of mobile networked applications","year":2019,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software System Performance and Reliability","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 Waterloo","funders":"","keywords":"Computer science; Computer network; Test suite; Wireless network; End user; Wireless; Test case; Machine learning; Operating system","score_opus":0.06774374964768788,"score_gpt":0.3249485911742777,"score_spread":0.25720484152658984,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2979327180","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":[0.5904573,0.00016877147,0.40661377,0.00003184286,0.00030974488,0.0021272744,0.000009762835,0.00024620662,0.000035294688],"genre_scores_gemma":[0.7562675,0.00002419404,0.24260867,0.000028207376,0.00009814756,0.0008767645,0.000061494204,0.000010676374,0.000024334488],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99752045,0.0003930608,0.0006559382,0.0007765924,0.0003968841,0.00025706258],"domain_scores_gemma":[0.99511564,0.001466109,0.0004234387,0.001294108,0.0016052693,0.00009541725],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0057799853,0.00018139751,0.00030652396,0.00010367735,0.0002597064,0.000055028457,0.00045418178,0.00017550499,0.00001501628],"category_scores_gemma":[0.0019243952,0.00016018307,0.000060161594,0.0006897963,0.00011699354,0.0005740107,0.000060382907,0.00012731532,0.000011767437],"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":[0.00001605227,0.0003323393,0.48479882,0.0004300596,0.000009851371,1.9306706e-8,0.00047011743,0.03208456,0.006148008,0.00038034856,0.000058343518,0.47527146],"study_design_scores_gemma":[0.0006921017,0.0004894845,0.21116276,0.000032843374,0.000045704022,0.0000057317475,0.00004196344,0.78227097,0.003074805,0.0008424615,0.0010955456,0.00024563348],"about_ca_topic_score_codex":0.000033939512,"about_ca_topic_score_gemma":0.000002124606,"teacher_disagreement_score":0.7501864,"about_ca_system_score_codex":0.0001263453,"about_ca_system_score_gemma":0.00030077674,"threshold_uncertainty_score":0.6532076},"labels":[],"label_agreement":null},{"id":"W2994711059","doi":"10.1002/stvr.1721","title":"Leveraging metamorphic testing to automatically detect inconsistencies in code generator families","year":2019,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","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":"McGill University","funders":"Agence Nationale de la Recherche","keywords":"Computer science; Software quality; Oracle; Software; Leverage (statistics); Regression testing; Code (set theory); Code coverage; Unreachable code; Software development; Code generation; Programming language; Redundant code; Software construction; Set (abstract data type); Operating system; Artificial intelligence","score_opus":0.03641849049665075,"score_gpt":0.2574023251371866,"score_spread":0.22098383464053584,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2994711059","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":[0.7547442,0.00008471105,0.23526841,0.00038113023,0.00016791717,0.0005106825,0.0000030206397,0.008630551,0.00020937761],"genre_scores_gemma":[0.50786984,0.0000014617078,0.4917551,0.00028268035,0.0000131211345,0.000045170735,0.0000012303383,0.000012219283,0.000019204748],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.997443,0.00021551858,0.0006381865,0.00093896553,0.00034719243,0.00041714573],"domain_scores_gemma":[0.99470073,0.0033361125,0.00019749338,0.001136562,0.0004458889,0.00018323639],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.002001292,0.00027230816,0.0003881939,0.00027132416,0.00022338553,0.00023293273,0.0006089333,0.00012298918,0.0000038561643],"category_scores_gemma":[0.024946714,0.00025259287,0.000046260135,0.0014874926,0.000105920924,0.00032764202,0.00035318424,0.00030406215,0.000045181037],"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":[0.000007048522,0.000078500525,0.88073,0.00023353094,0.000006837165,0.000006950198,0.0012438924,0.0015692114,0.0038560685,0.00022167045,0.00023586607,0.111810446],"study_design_scores_gemma":[0.0003309694,0.00023742297,0.7676047,0.00041724675,0.000012669134,0.000049865535,0.000046804143,0.20910786,0.0012360069,0.02008365,0.0002381691,0.00063465076],"about_ca_topic_score_codex":0.0004311439,"about_ca_topic_score_gemma":0.000011888509,"teacher_disagreement_score":0.25648668,"about_ca_system_score_codex":0.00014585207,"about_ca_system_score_gemma":0.0002463042,"threshold_uncertainty_score":0.9999926},"labels":[],"label_agreement":null},{"id":"W3013655954","doi":"10.1002/stvr.380","title":"Automated discovery of state transitions and their functions in source code","year":2007,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":29,"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":"Engineering and Physical Sciences Research Council; National Aeronautics and Space Administration","keywords":"Computer science; State (computer science); Source code; Finite-state machine; Reverse engineering; Set (abstract data type); Code (set theory); Programming language; Software; Software engineering; Transition (genetics); Theoretical computer science","score_opus":0.01937952920259313,"score_gpt":0.2572724153819903,"score_spread":0.2378928861793972,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3013655954","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":[0.4071725,0.000054395514,0.58806074,0.00011540234,0.000034394398,0.00011615948,0.000010982271,0.004406615,0.000028808592],"genre_scores_gemma":[0.8353387,0.0000047557355,0.16456345,0.000039681803,0.000005933712,0.000009539333,0.000005903872,0.0000072614794,0.000024780464],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99886286,0.0000750759,0.00037848795,0.0003899701,0.000102669735,0.00019090525],"domain_scores_gemma":[0.9978559,0.0013469597,0.00012968562,0.00044336656,0.00015958372,0.00006451261],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001241302,0.00012769033,0.00017047602,0.00013914544,0.00014784058,0.000060747214,0.00016119127,0.00006681889,4.1818248e-7],"category_scores_gemma":[0.0020752957,0.000109974084,0.000025610096,0.00061478675,0.00020314983,0.00028849873,0.00006322138,0.00015185826,5.8377856e-7],"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":[0.000037908394,0.00045624166,0.7707582,0.00030824143,0.0000128546435,0.0000047334092,0.0052740173,0.0011326972,0.003989526,0.0004088314,0.00045020794,0.21716653],"study_design_scores_gemma":[0.00031227717,0.00013051937,0.872484,0.00017741801,0.0000069535768,0.000043270804,0.000081899314,0.10380101,0.0014406458,0.0211267,0.00012997494,0.00026530167],"about_ca_topic_score_codex":0.00031383842,"about_ca_topic_score_gemma":0.00003433544,"teacher_disagreement_score":0.4281662,"about_ca_system_score_codex":0.000035946614,"about_ca_system_score_gemma":0.00005275707,"threshold_uncertainty_score":0.4484613},"labels":[],"label_agreement":null},{"id":"W3040858112","doi":"10.1002/stvr.1745","title":"TimelyRep: Timing deterministic replay for Android web applications","year":2020,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Green IT and Sustainability","field":"Engineering","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":"McGill University","funders":"","keywords":"Debugging; Computer science; Android (operating system); Web application; Touchscreen; Backward compatibility; Mobile device; Operating system; Embedded system; Event (particle physics); World Wide Web","score_opus":0.030880898994668088,"score_gpt":0.24372867710065063,"score_spread":0.21284777810598254,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3040858112","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":[0.51584274,0.00033745886,0.4758656,0.0008163847,0.00018705954,0.0024170699,0.00012979226,0.0034080015,0.0009959221],"genre_scores_gemma":[0.95086664,0.000006680102,0.048557743,0.000083967105,0.00008690196,0.0002949939,0.000036869264,0.00002673766,0.00003949605],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9989527,0.000025482877,0.00032555044,0.00039943913,0.000090499605,0.00020634645],"domain_scores_gemma":[0.9987327,0.00048115934,0.000053628286,0.00036664025,0.00020622616,0.00015966651],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00031390323,0.00014939299,0.0001806377,0.000024771587,0.00020974444,0.000044521857,0.00011991831,0.00009085902,0.000011853077],"category_scores_gemma":[0.0027695254,0.00015160123,0.000047875317,0.00023886253,0.000100776604,0.000086073975,0.000026739724,0.00013636946,0.000010879079],"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":[0.00021392982,0.00035922386,0.39725742,0.012203981,0.00009232587,0.0000066516227,0.0051091397,0.02417345,0.016976977,0.0010068112,0.00886202,0.5337381],"study_design_scores_gemma":[0.0010676298,0.00026419206,0.08159151,0.000054036245,0.000121675475,0.000017093245,0.00035912995,0.8360751,0.0009993715,0.006387408,0.07212888,0.0009339581],"about_ca_topic_score_codex":0.000004755047,"about_ca_topic_score_gemma":8.8670447e-7,"teacher_disagreement_score":0.8119017,"about_ca_system_score_codex":0.000059461818,"about_ca_system_score_gemma":0.00005530176,"threshold_uncertainty_score":0.61821187},"labels":[],"label_agreement":null},{"id":"W3092453472","doi":"10.1002/stvr.1751","title":"BUGSJS: a benchmark and taxonomy of JavaScript bugs","year":2020,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":23,"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 British Columbia","funders":"European Social Fund; European Commission; Natural Sciences and Engineering Research Council of Canada; Advanced Remanufacturing and Technology Centre; National Research, Development and Innovation Office; Innovációs és Technológiai Minisztérium","keywords":"JavaScript; Computer science; Unobtrusive JavaScript; Benchmark (surveying); Unit testing; Software bug; Debugging; Taxonomy (biology); Programming language; Web application; Software; Software testing; Test case; Software engineering; Rich Internet application; World Wide Web; Machine learning","score_opus":0.05055143313862166,"score_gpt":0.23784943200647207,"score_spread":0.18729799886785042,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3092453472","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":[0.124593884,0.00054497825,0.8630589,0.0019847378,0.00013086564,0.00056961796,0.000009754038,0.008673445,0.00043377094],"genre_scores_gemma":[0.54054844,0.000010078064,0.45913547,0.00024750977,0.000019037556,0.000025943158,0.0000021795634,0.0000056964336,0.0000056553286],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9986395,0.000085982894,0.00035745735,0.0005738101,0.00017191388,0.00017132537],"domain_scores_gemma":[0.9980181,0.00083582854,0.00019820414,0.00052324944,0.00025544318,0.00016916887],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005511987,0.00015355772,0.00022766028,0.00004720159,0.00014149604,0.00007274273,0.0003338023,0.00008280217,0.0000032313762],"category_scores_gemma":[0.008090468,0.00014215313,0.00003347906,0.00045286145,0.00021070485,0.0002230042,0.00021912188,0.00016038513,0.0000022264348],"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":[0.000017914788,0.000119171025,0.7038515,0.00057644106,0.000013297227,0.0000033010003,0.0015688188,0.000034944882,0.0009246173,0.001602984,0.0066559464,0.28463104],"study_design_scores_gemma":[0.0012397835,0.001254531,0.759297,0.00050596613,0.00007543569,0.000098193894,0.00009084362,0.145851,0.0037990736,0.06938388,0.01696659,0.0014377114],"about_ca_topic_score_codex":0.00013743389,"about_ca_topic_score_gemma":4.4565866e-7,"teacher_disagreement_score":0.41595456,"about_ca_system_score_codex":0.000017798677,"about_ca_system_score_gemma":0.00007726569,"threshold_uncertainty_score":0.9685631},"labels":[],"label_agreement":null},{"id":"W3206885924","doi":"10.1002/stvr.1799","title":"A mutation framework for evaluating security analysis tools in IoT applications","year":2021,"lang":"en","type":"preprint","venue":"Software Testing Verification and Reliability","topic":"Advanced Malware Detection Techniques","field":"Computer Science","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Toronto Metropolitan University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Taint checking; Internet of Things; Sensitivity (control systems); Precision and recall; Context (archaeology); Security analysis; Set (abstract data type); Data mining; Process (computing); Domain (mathematical analysis); Information flow; Information retrieval; Computer security; Software; Engineering; Programming language","score_opus":0.06271620279564884,"score_gpt":0.3700882404693309,"score_spread":0.3073720376736821,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3206885924","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":[0.04529462,0.00022901474,0.95101213,0.00022897092,0.00014785484,0.0018389844,0.000036374415,0.0011976928,0.000014369782],"genre_scores_gemma":[0.3583991,0.000016500675,0.6394122,0.000051351173,0.000029304128,0.0019640238,0.00011152614,0.000012026418,0.000003971987],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.99696267,0.00026617586,0.00076223566,0.0014491543,0.00030543748,0.0002543067],"domain_scores_gemma":[0.9940093,0.0025619841,0.00062798214,0.001584116,0.001126082,0.00009054028],"candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0018265599,0.00027307734,0.00046691103,0.0003219348,0.00025565655,0.0004734725,0.00054041564,0.00043277253,0.0000040200807],"category_scores_gemma":[0.015530457,0.0003100929,0.0001717055,0.0019681582,0.000092357775,0.00030574054,0.00048443562,0.00071498833,8.283541e-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":[0.000027958497,0.0004973632,0.027068784,0.0019287204,0.00014249326,0.0000034006775,0.0035606786,0.065459795,0.0003384553,0.008099604,0.000013664935,0.8928591],"study_design_scores_gemma":[0.00020375848,0.00008629293,0.065011226,0.00026472198,0.00019311669,0.000005645298,0.000163021,0.41293758,0.0011351184,0.5192478,0.00013468388,0.00061706203],"about_ca_topic_score_codex":0.00010888648,"about_ca_topic_score_gemma":0.000023469656,"teacher_disagreement_score":0.892242,"about_ca_system_score_codex":0.0002989965,"about_ca_system_score_gemma":0.00029902192,"threshold_uncertainty_score":0.9999351},"labels":[],"label_agreement":null},{"id":"W3207636857","doi":"10.1002/stvr.1796","title":"GPU acceleration of finite state machine input execution: Improving scale and performance","year":2021,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","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":"McGill University","funders":"Engineering and Physical Sciences Research Council; University of Edinburgh","keywords":"Computer science; Scalability; Parallel computing; Kernel (algebra); Speedup; Multi-core processor; Finite-state machine; Acceleration; Process (computing); CUDA; Computer engineering; Algorithm; Programming language; Operating system","score_opus":0.02022802077216418,"score_gpt":0.24044934009781616,"score_spread":0.22022131932565198,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3207636857","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":[0.51943827,0.00024883018,0.47778985,0.00018138725,0.0000809946,0.00012013571,0.000005788275,0.0020752547,0.000059524788],"genre_scores_gemma":[0.64260423,0.00003921373,0.3572326,0.000055684584,0.000011148833,0.000010686182,0.000008495219,0.000005668596,0.000032248878],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9987197,0.00009393516,0.00035433102,0.00049428793,0.00017758865,0.00016017054],"domain_scores_gemma":[0.99797016,0.0006885423,0.00019949874,0.000579562,0.00048739443,0.000074854215],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007053636,0.00013362714,0.00017419337,0.000059036487,0.00024365824,0.000113113936,0.00017449155,0.00006722315,0.000002223478],"category_scores_gemma":[0.0035113932,0.00012826225,0.00002296494,0.00044751746,0.0001336393,0.00036737171,0.00019679168,0.00015816758,0.0000011128177],"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":[0.000010294186,0.00008814393,0.4916123,0.00038675862,0.0000038230337,0.0000028275679,0.0006835577,0.00021086747,0.0015464773,0.00007788542,0.0000908866,0.5052862],"study_design_scores_gemma":[0.00034531287,0.00021780662,0.5767328,0.00021187906,0.000015017811,0.00009899086,0.000010020012,0.3983274,0.012983621,0.010556875,0.0001501467,0.00035008727],"about_ca_topic_score_codex":0.00013587145,"about_ca_topic_score_gemma":0.0000038548988,"teacher_disagreement_score":0.5049361,"about_ca_system_score_codex":0.000031143707,"about_ca_system_score_gemma":0.000121214776,"threshold_uncertainty_score":0.52303827},"labels":[],"label_agreement":null},{"id":"W4230255398","doi":"10.1002/stvr.385","title":"Error‐preserving reductions on communication protocols","year":2007,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Formal Methods in Verification","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 Windsor","funders":"","keywords":"Reachability; Computer science; Model checking; Variety (cybernetics); State (computer science); Process (computing); Transformation (genetics); Formal methods; Communications protocol; Distributed computing; Theoretical computer science; Programming language; Artificial intelligence; Computer network","score_opus":0.08466286065561494,"score_gpt":0.36484616857315927,"score_spread":0.2801833079175443,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4230255398","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":[0.041641265,0.000032473195,0.94832,0.000691749,0.00020680334,0.0037769796,0.0000030306473,0.0012785719,0.0040491433],"genre_scores_gemma":[0.17539272,0.000004008882,0.82364696,0.00011166229,0.00003678574,0.0006820834,0.0000089186915,0.000011336694,0.00010549683],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99815774,0.00027473396,0.00048277422,0.0005389791,0.00028715085,0.00025859967],"domain_scores_gemma":[0.99645007,0.00088012894,0.00027340036,0.0018941066,0.00037826673,0.00012401667],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003952032,0.00015328437,0.00014035041,0.00011775692,0.00057658955,0.00013607602,0.00076365634,0.00013524354,0.000007374735],"category_scores_gemma":[0.0075218724,0.0001470414,0.00003192825,0.00072521507,0.00019187856,0.0005704486,0.0001829533,0.00035703342,0.000024304416],"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":[0.00013968717,0.0011756764,0.07472285,0.00047977088,0.000022635833,0.0000025735694,0.0043220185,0.0021719742,0.0074796546,0.2047545,0.0019441284,0.70278454],"study_design_scores_gemma":[0.0006498436,0.00038941923,0.8761221,0.00034120752,0.000015078303,0.000047826252,0.00024582053,0.049648408,0.009363617,0.040048793,0.022383604,0.00074425194],"about_ca_topic_score_codex":0.000057630423,"about_ca_topic_score_gemma":0.0000025100048,"teacher_disagreement_score":0.8013993,"about_ca_system_score_codex":0.00012850095,"about_ca_system_score_gemma":0.00006642528,"threshold_uncertainty_score":0.9004927},"labels":[],"label_agreement":null},{"id":"W4236669998","doi":"10.1002/stvr.401","title":"Modelling methods for web application verification and testing: state of the art","year":2008,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":46,"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; Web testing; Software engineering; Software; Web application; Verification and validation; Web modeling; Software testing; Data mining; World Wide Web; Web service; Programming language; Engineering","score_opus":0.06046944475577618,"score_gpt":0.3077181642846764,"score_spread":0.24724871952890023,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4236669998","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":[0.07067727,0.00013485651,0.9257042,0.0002511447,0.00007943813,0.00064611656,0.0000066484413,0.0024484845,0.000051875882],"genre_scores_gemma":[0.4010731,0.000011054049,0.5987321,0.000038156006,0.000011939508,0.00009715554,0.0000034914858,0.000009967124,0.000022988323],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99837625,0.0001803384,0.00046594584,0.0006160024,0.00017079915,0.00019066363],"domain_scores_gemma":[0.99502796,0.0028333687,0.00040004932,0.0009932833,0.0006676403,0.0000776787],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0016378337,0.00017090191,0.00021293046,0.00006684853,0.00049587607,0.000046678484,0.0004133434,0.00008222745,1.8064014e-7],"category_scores_gemma":[0.0077291825,0.0001376734,0.000045436278,0.0006429994,0.00030700176,0.00019746932,0.00013606843,0.00015547434,8.8409774e-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":[0.000031420102,0.00025748755,0.1890419,0.00056025997,0.000016252208,2.9768145e-7,0.0013672521,0.012082109,0.012464524,0.0016863748,0.00097145984,0.78152066],"study_design_scores_gemma":[0.00017120209,0.000072272094,0.052994452,0.000062790525,0.000013103117,0.000028182058,0.0000024857288,0.88363796,0.0022894575,0.05975158,0.00078827,0.00018823259],"about_ca_topic_score_codex":0.000082020626,"about_ca_topic_score_gemma":4.429564e-7,"teacher_disagreement_score":0.87155586,"about_ca_system_score_codex":0.000041880834,"about_ca_system_score_gemma":0.00015200269,"threshold_uncertainty_score":0.92531115},"labels":[],"label_agreement":null},{"id":"W4248647026","doi":"10.1002/stvr.410","title":"Improving the coverage criteria of UML state machines using data flow analysis","year":2009,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","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":"Carleton University","funders":"","keywords":"Computer science; Unified Modeling Language; Guard (computer science); Finite-state machine; Data mining; Data flow diagram; Test suite; Context (archaeology); State (computer science); Control flow; Data-flow analysis; Tree (set theory); Test case; Algorithm; Machine learning; Database; Programming language; Mathematics; Software","score_opus":0.04962212616320409,"score_gpt":0.3113156837427992,"score_spread":0.2616935575795951,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4248647026","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":[0.10421861,0.00011800311,0.8926222,0.00025071763,0.00007623665,0.0001703985,0.000035088204,0.0024894767,0.000019304935],"genre_scores_gemma":[0.5423722,0.000004415052,0.45748457,0.00009513962,0.000016086562,0.000001967961,0.000016352255,0.0000042731854,0.000004936439],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99824095,0.0001821735,0.00045151537,0.00065828755,0.00025869315,0.00020835357],"domain_scores_gemma":[0.99602854,0.0009673744,0.0003361685,0.002271335,0.00033010935,0.000066441986],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001839679,0.00016834734,0.00026240782,0.00011894391,0.0003132579,0.0001804482,0.0010653164,0.000057439996,0.0000026565374],"category_scores_gemma":[0.007380912,0.0001226545,0.000055037694,0.0012314703,0.00013452508,0.00040474415,0.00033608152,0.00016642248,5.669326e-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":[0.00001184047,0.00012726072,0.15469313,0.00009100228,0.000036770856,0.0000022724764,0.0005318355,0.0032084605,0.0015679789,0.00007493391,0.00033072068,0.8393238],"study_design_scores_gemma":[0.00007157257,0.00004548904,0.16671772,0.000027230859,0.00006940669,0.000008101148,0.0000015063233,0.8212285,0.00018616911,0.01148616,0.000019759174,0.00013835994],"about_ca_topic_score_codex":0.00078329147,"about_ca_topic_score_gemma":0.0000042126453,"teacher_disagreement_score":0.8391854,"about_ca_system_score_codex":0.000032877742,"about_ca_system_score_gemma":0.00009130535,"threshold_uncertainty_score":0.8836174},"labels":[],"label_agreement":null},{"id":"W4251245562","doi":"10.1002/stvr.396","title":"Transition covering tests for systems with queues","year":2008,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","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":"Computer Research Institute of Montréal; Nortel (Canada)","funders":"","keywords":"Context (archaeology); Computer science; Queue; Metric (unit); Concurrency; Cover (algebra); Model-based testing; Test (biology); Test case; Reliability engineering; Distributed computing; Programming language; Engineering; Operations management; Mechanical engineering","score_opus":0.0432607937311662,"score_gpt":0.25965007470477164,"score_spread":0.21638928097360544,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4251245562","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":[0.109989874,0.00013108655,0.8794623,0.00017946641,0.000102019076,0.00047828717,0.000006738712,0.009595705,0.00005452879],"genre_scores_gemma":[0.587909,0.000007565455,0.41185686,0.000055155764,0.000030004601,0.000105118816,0.000006471044,0.000010872967,0.000018956809],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99871415,0.0000613941,0.00026663113,0.00054173206,0.00019221171,0.00022387886],"domain_scores_gemma":[0.9975461,0.0012199004,0.00013732674,0.00058126025,0.00042227752,0.000093146075],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005395967,0.00016619277,0.00019157518,0.00006603947,0.0004639795,0.00009845254,0.0002446733,0.000083183935,3.41911e-7],"category_scores_gemma":[0.0023167229,0.00013866829,0.000031008298,0.0003443139,0.0001514578,0.00031228794,0.00003372233,0.00011273407,0.000001716273],"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":[0.000298989,0.001251404,0.76139593,0.0035941964,0.000089000954,0.00006255285,0.012119901,0.022376474,0.0023207418,0.0064689494,0.015539887,0.17448194],"study_design_scores_gemma":[0.0019489042,0.0018419396,0.3484124,0.0010584894,0.000063582476,0.0013146644,0.00005963991,0.6089415,0.0014322492,0.03019264,0.002964646,0.0017693316],"about_ca_topic_score_codex":0.00022968517,"about_ca_topic_score_gemma":0.0000012514162,"teacher_disagreement_score":0.586565,"about_ca_system_score_codex":0.000056758847,"about_ca_system_score_gemma":0.00010952802,"threshold_uncertainty_score":0.5654729},"labels":[],"label_agreement":null},{"id":"W4322489317","doi":"10.1002/stvr.1842","title":"An investigation of distributed computing for combinatorial testing","year":2023,"lang":"en","type":"article","venue":"Software Testing Verification and Reliability","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Chicoutimi","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Hypergraph; Computer science; Graph; Theoretical computer science; Vertex (graph theory); Algorithm; Parallel computing; Mathematics; Discrete mathematics","score_opus":0.0456312420782707,"score_gpt":0.2955049233390974,"score_spread":0.2498736812608267,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4322489317","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":[0.46735862,0.000010726057,0.52071446,0.00011009374,0.00023509686,0.00032408713,0.00001917196,0.011221113,0.000006618931],"genre_scores_gemma":[0.57909185,6.0408337e-7,0.42075267,0.00002110248,0.000041079602,0.000022895096,0.00005871048,0.000009815505,0.0000012963166],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983119,0.00013975463,0.00048515107,0.00057761476,0.00022062354,0.00026496066],"domain_scores_gemma":[0.9937956,0.0041784034,0.00033891905,0.0007160111,0.00084511255,0.00012597124],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0020192964,0.00016146236,0.0002278625,0.000099981255,0.0003682111,0.00010201705,0.00044427149,0.00011378238,2.6427693e-7],"category_scores_gemma":[0.016846806,0.00016118179,0.00003771846,0.001451455,0.00017320743,0.00028349046,0.00012450464,0.00013078662,0.0000016025073],"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":[0.000011889842,0.00008093044,0.92058694,0.00036442018,0.00000634468,9.222547e-7,0.00076356245,0.0012075987,0.002893696,0.002519441,0.001064881,0.07049934],"study_design_scores_gemma":[0.00029366234,0.00025587884,0.42577252,0.00011730203,0.000009659437,0.0000052985047,0.000012747297,0.42430222,0.0012387368,0.14775926,0.00002303252,0.0002096751],"about_ca_topic_score_codex":0.00011762573,"about_ca_topic_score_gemma":3.5803518e-7,"teacher_disagreement_score":0.49481446,"about_ca_system_score_codex":0.000044349927,"about_ca_system_score_gemma":0.00013024734,"threshold_uncertainty_score":0.9914347},"labels":[],"label_agreement":null}]}