{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":996,"total_is_capped":false,"direct_labels_cover":2,"predictions_cover":996,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"query_hash":"100aa09bc2bb","filters":{"topic":"Software Testing and Debugging Techniques"}},"results":[{"id":"W1577404745","doi":"10.1007/s10009-007-0044-z","title":"The software model checker Blast","year":2007,"lang":"en","type":"article","venue":"International Journal on Software Tools for Technology Transfer","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":594,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Simon Fraser University","funders":"","keywords":"Computer science; Predicate abstraction; Predicate (mathematical logic); Programming language; Memory safety; Model checking; Programmer; Program analysis; Symbolic execution; Static analysis; Software","retraction":null,"screen_n_in":null,"score":{"opus":0.0292661240647697,"gpt":0.3019749021356067,"spread":0.272708778070837,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001304751,0.0002882028,0.000227223,0.0005495513,0.0006265812,0.0006346113,0.003376931,0.0003183575,0.00000516752],"category_scores_gemma":[0.002479609,0.0002097764,0.0002735305,0.0003686378,0.0002213255,0.0006114863,0.0001441744,0.0008166314,0.00002634934],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002116833,"about_ca_system_score_gemma":0.0001578328,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00000148654,"about_ca_topic_score_gemma":0.000006861332,"domain_scores_codex":[0.9975857,0.00002261766,0.0006160554,0.0004322928,0.0007136144,0.0006297166],"domain_scores_gemma":[0.9971406,0.001131021,0.0001172259,0.0005876943,0.0008975687,0.0001259258],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002173323,0.0001852705,0.001616816,0.000006059486,0.0002118439,0.0001423882,0.000166019,0.00056613,0.000144486,0.1697166,0.01446239,0.8125647],"study_design_scores_gemma":[0.002718662,0.001110028,0.001134573,0.0003471548,0.0000483385,0.002425137,0.00003723467,0.02204656,0.02035416,0.839809,0.1088631,0.001106037],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.007489565,0.0001519657,0.9785917,0.004840612,0.001246469,0.0002508382,0.00002076642,0.007317055,0.00009095945],"genre_scores_gemma":[0.262687,0.00009126679,0.7343988,0.002031089,0.0002855171,0.00008012202,0.000006807586,0.00005315768,0.00036629],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.8114586,"threshold_uncertainty_score":0.8554432,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2106072155","doi":"10.1145/2568225.2568271","title":"Coverage is not strongly correlated with test suite effectiveness","year":2014,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":398,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"","keywords":"Test suite; Suite; Computer science; Proxy (statistics); Test (biology); Confounding; Test case; Statistics; Data mining; Machine learning; Mathematics; Regression analysis; Geology; Geography","retraction":null,"screen_n_in":null,"score":{"opus":0.008925411068723877,"gpt":0.22585647833184,"spread":0.2169310672631161,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004132118,0.0001506866,0.0001635363,0.00006937492,0.0001027322,0.0001540903,0.0005177393,0.00005887808,0.0000121126],"category_scores_gemma":[0.0003361534,0.0001105226,0.00003111686,0.0003145589,0.00004042397,0.0002265606,0.0001370558,0.0001359738,0.00006890502],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000299135,"about_ca_system_score_gemma":0.00003591802,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001534613,"about_ca_topic_score_gemma":0.000003454783,"domain_scores_codex":[0.9989675,0.00007549429,0.0001139788,0.0003755177,0.000224317,0.0002431359],"domain_scores_gemma":[0.9974527,0.001697931,0.00005836347,0.00060923,0.0001044926,0.00007723255],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00009665208,0.0004770368,0.7536665,0.0001311378,0.00007387555,0.00007403986,0.0007159901,0.0006895232,0.001485468,0.03355272,0.03380033,0.1752367],"study_design_scores_gemma":[0.002469382,0.003221383,0.4012133,0.0004887831,0.00003335571,0.0002173164,0.000002145874,0.5158896,0.04486172,0.02579502,0.004464668,0.001343297],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04433278,0.00000509963,0.9402823,0.0001525712,0.00008539899,0.0001382851,0.000001430962,0.01046235,0.004539736],"genre_scores_gemma":[0.9325945,6.007951e-7,0.06644562,0.0005680937,0.00001979924,0.00001512123,0.00000132208,0.00001244459,0.0003424466],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8882618,"threshold_uncertainty_score":0.4506979,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4391974543","doi":"10.1109/tse.2024.3368208","title":"Software Testing With Large Language Models: Survey, Landscape, and Vision","year":2024,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":394,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"York University","funders":"National Natural Science Foundation of China","keywords":"Computer science; Software engineering; Software","retraction":null,"screen_n_in":null,"score":{"opus":0.01712040667441351,"gpt":0.2519216021540697,"spread":0.2348011954796562,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004307948,0.0002758799,0.0002043153,0.0003335048,0.0001739915,0.000356657,0.0002940583,0.000100362,0.000004650301],"category_scores_gemma":[0.0001307121,0.0002414727,0.00004782404,0.0008946389,0.00001908114,0.0006385835,0.00001028966,0.0004129281,0.00001037041],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004373556,"about_ca_system_score_gemma":0.00005627131,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007347211,"about_ca_topic_score_gemma":0.00001836886,"domain_scores_codex":[0.9985636,0.00003535745,0.0001922715,0.0005519171,0.0002780079,0.0003788106],"domain_scores_gemma":[0.9979014,0.001423628,0.00002564287,0.0004460401,0.00007669147,0.0001265968],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00004930575,0.0003161851,0.004695001,0.001078558,0.0002237394,0.0007250738,0.004010785,0.4715207,0.0002554365,0.0005504697,0.002971538,0.5136032],"study_design_scores_gemma":[0.0002335854,0.0002505039,0.001153659,0.0008959645,0.00002088192,0.0001627627,0.000007736919,0.9957085,0.0006930251,0.0002500827,0.0001439881,0.0004793389],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.009641844,0.0006335574,0.965713,0.00001924298,0.0003717317,0.0001313164,0.00004195171,0.02343378,0.00001352697],"genre_scores_gemma":[0.5931725,0.00001074565,0.4066329,0.00003383341,0.00002607447,0.00003106284,0.000003327154,0.00004493156,0.0000446815],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5835306,"threshold_uncertainty_score":0.9846971,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4367672983","doi":"10.1016/j.jss.2023.111734","title":"GitHub Copilot AI pair programmer: Asset or Liability?","year":2023,"lang":"en","type":"article","venue":"Journal of Systems and Software","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":340,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"York University; Polytechnique Montréal","funders":"Fonds de recherche du Québec; Natural Sciences and Engineering Research Council of Canada; Canadian Institute for Advanced Research","keywords":"Programmer; Liability; Asset (computer security); Computer science; Computer security; Software engineering; Business; Programming language; Finance","retraction":null,"screen_n_in":null,"score":{"opus":0.0401045157013695,"gpt":0.3012254088637339,"spread":0.2611208931623644,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001574341,0.0001533978,0.0003745779,0.0002257547,0.0001475814,0.0004063418,0.0005187945,0.00009184073,0.000002090503],"category_scores_gemma":[0.001203567,0.0001018429,0.00009060877,0.0006106605,0.00005098329,0.0004502911,0.0001671575,0.0002499911,0.00001071821],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004102232,"about_ca_system_score_gemma":0.0001692993,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006388364,"about_ca_topic_score_gemma":0.000002201473,"domain_scores_codex":[0.9983994,0.0001156986,0.0005481407,0.000222527,0.0004299803,0.0002841927],"domain_scores_gemma":[0.9980904,0.0006223114,0.0003392537,0.0003743999,0.0003871284,0.0001865151],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00005067293,0.0001811112,0.3830893,0.0008204465,0.0001137595,0.0007929514,0.001290335,0.0001311827,0.00002099889,0.001226787,0.4864551,0.1258274],"study_design_scores_gemma":[0.00529501,0.01362432,0.2095333,0.009799542,0.0002130666,0.01530536,0.0006704932,0.06383058,0.0003097507,0.07838491,0.5998933,0.003140385],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.118728,0.00224751,0.8568392,0.004014459,0.002888534,0.0007184925,0.00001811884,0.01447877,0.0000668486],"genre_scores_gemma":[0.8570566,0.00008884296,0.1411578,0.0003965842,0.0003848065,0.00003106478,0.000002584753,0.00003328644,0.0008484633],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7383286,"threshold_uncertainty_score":0.4153034,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3043761819","doi":"10.1145/3395363.3397369","title":"CoCoNuT: combining context-aware neural translation models using ensemble for program repair","year":2020,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":329,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"Natural Sciences and Engineering Research Council of Canada; Facebook","keywords":"Computer science; Context (archaeology); Machine translation; Artificial intelligence; Programming language; Machine learning; Software engineering; Natural language processing","retraction":null,"screen_n_in":null,"score":{"opus":0.1833401154001179,"gpt":0.3325084569739183,"spread":0.1491683415738004,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001805115,0.0001366371,0.000179894,0.00004110545,0.0001579996,0.0001380927,0.0003324439,0.00006354201,0.000001430622],"category_scores_gemma":[0.00007683725,0.0001300539,0.0001175421,0.0002328315,0.00002564121,0.0004916535,0.00006444344,0.00009399345,0.000001084708],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001717911,"about_ca_system_score_gemma":0.00004839165,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006914186,"about_ca_topic_score_gemma":0.000003362115,"domain_scores_codex":[0.9989389,0.00003713517,0.0002394886,0.0003852293,0.0001484367,0.0002507653],"domain_scores_gemma":[0.9991911,0.0002672384,0.00007263412,0.0002475728,0.0001175555,0.0001038826],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00006990842,0.0001977543,0.001301208,0.0002222097,0.00005415046,0.00002219029,0.005685476,0.01064379,0.0005098095,0.02273963,0.03017729,0.9283766],"study_design_scores_gemma":[0.0003341573,0.0002931858,0.000007047338,0.00002929189,0.000008178854,0.000009978152,0.00001423356,0.9816617,0.0005568761,0.01662295,0.000293676,0.0001686952],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.002881715,0.00007764616,0.9391558,0.001076358,0.00007386635,0.0006532814,0.000001567186,0.05584811,0.0002316877],"genre_scores_gemma":[0.5110445,3.971691e-7,0.4880148,0.0008539909,0.00002475025,0.00004303488,0.000002771699,0.000009462401,0.000006262771],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.971018,"threshold_uncertainty_score":0.5303442,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1536265389","doi":"10.1007/3-540-46423-9_2","title":"Optimizing Java Bytecode Using the Soot Framework: Is It Feasible?","year":2000,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":313,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Bytecode; Java bytecode; Computer science; Java; Programming language; Java annotation; Class (philosophy); Generics in Java; Program optimization; Java applet; Operating system; Compiler; Artificial intelligence","retraction":null,"screen_n_in":null,"score":{"opus":0.05289934935660727,"gpt":0.3031957885134227,"spread":0.2502964391568154,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","scholarly_communication","open_science"],"consensus_categories":[],"category_scores_codex":[0.00128173,0.0007173932,0.0005900614,0.0007478898,0.0007985505,0.001353028,0.006140446,0.0005461646,0.00007436929],"category_scores_gemma":[0.000268002,0.0005573237,0.0002172527,0.0009954231,0.0009063488,0.0005604887,0.001808911,0.001712616,0.00004536924],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003368097,"about_ca_system_score_gemma":0.0006746717,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001000898,"about_ca_topic_score_gemma":0.00001288505,"domain_scores_codex":[0.9952644,0.00006079028,0.0006299246,0.001813889,0.001232889,0.0009980998],"domain_scores_gemma":[0.9949201,0.001574353,0.0003616393,0.002741683,0.0002226227,0.0001796267],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000007423636,0.00003897763,0.0001187875,0.00004073413,0.00002489269,0.0001340492,0.004689309,0.06349618,0.00002923421,0.008647664,0.001194181,0.9215786],"study_design_scores_gemma":[0.00006533026,0.0000614891,0.0000107832,0.0009310442,0.00001010738,0.0001603907,9.893617e-8,0.4985569,0.0004522634,0.497696,0.001489595,0.0005659456],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.00003298611,0.0008121623,0.9894285,0.002632916,0.001101413,0.0003807457,0.000004392977,0.002415881,0.003190956],"genre_scores_gemma":[0.02900715,0.00008981725,0.9614568,0.008648426,0.0005217352,0.000008803256,0.000001241307,0.00005785926,0.0002081996],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.9210126,"threshold_uncertainty_score":0.9996879,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2025618033","doi":"10.1007/s10270-002-0004-8","title":"A UML-Based Approach to System Testing","year":2002,"lang":"en","type":"article","venue":"Software & Systems Modeling","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":303,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Computer science; Unified Modeling Language; Sequence diagram; Class diagram; Object Constraint Language; Software engineering; Applications of UML; Testability; Model-based testing; Programming language; UML tool; Context (archaeology); Test case; Test Management Approach; Systems Modeling Language; Activity diagram; System testing; Software development; Reliability engineering; Software; Software construction","retraction":null,"screen_n_in":null,"score":{"opus":0.09020794658056883,"gpt":0.2392422866175647,"spread":0.1490343400369959,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009811245,0.0003869236,0.0004816003,0.0003797993,0.0004285632,0.0005823767,0.001422891,0.000161883,6.658205e-7],"category_scores_gemma":[0.001130128,0.0003706245,0.0001172616,0.001226306,0.00001928672,0.0003293241,0.0002815387,0.0002723805,0.0001118952],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002764631,"about_ca_system_score_gemma":0.00006921546,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005637488,"about_ca_topic_score_gemma":5.240738e-7,"domain_scores_codex":[0.996735,0.000164919,0.000681472,0.0009957913,0.0006867924,0.0007360195],"domain_scores_gemma":[0.9972215,0.0004530426,0.0001822278,0.001433205,0.0003928039,0.0003172652],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000002866236,0.0001230427,0.002449024,0.0006973682,0.00001995374,0.00003637305,0.0008334672,0.9679515,0.00003220628,0.002499116,0.004524813,0.02083027],"study_design_scores_gemma":[0.0001732372,0.00006874737,0.000005336015,0.0009278898,0.00001008086,0.0001378448,0.00003010248,0.9978403,0.00001105766,0.0002206707,0.0001173619,0.0004573236],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.001712759,0.0005651311,0.9526383,0.00004528631,0.000542793,0.0005968832,0.000004098613,0.04176677,0.002128015],"genre_scores_gemma":[0.5439675,2.148012e-7,0.4555323,0.0001202785,0.0001246779,0.0001683323,0.000001524739,0.00003598197,0.00004925389],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5422547,"threshold_uncertainty_score":0.9998746,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3004493192","doi":"10.1016/j.jss.2020.110542","title":"On testing machine learning programs","year":2020,"lang":"en","type":"article","venue":"PolyPublie (École Polytechnique de Montréal)","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":198,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Computer science; Transformative learning; Domain (mathematical analysis); Software testing; Software engineering; Test strategy; Field (mathematics); Software; Witness; Data science; Artificial intelligence; Machine learning; Psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.02772335947957308,"gpt":0.238025687806026,"spread":0.2103023283264529,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0006225286,0.0003966389,0.0003570227,0.0002407508,0.0003784891,0.0004866751,0.001447818,0.0001923467,0.000008407723],"category_scores_gemma":[0.003084274,0.0003855487,0.00013793,0.001408991,0.00006685615,0.0003735695,0.0005781389,0.000893072,0.00005395525],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001680288,"about_ca_system_score_gemma":0.0001408759,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002345767,"about_ca_topic_score_gemma":0.0000608046,"domain_scores_codex":[0.9973204,0.0001822147,0.0004258977,0.0007830232,0.0004956172,0.0007928479],"domain_scores_gemma":[0.9977366,0.0005628861,0.0002461784,0.0008289374,0.0001443587,0.000481016],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00005925542,0.0003617396,0.1346812,0.00008152946,0.00004391948,0.0003630303,0.000920746,0.004382005,0.002162282,0.07154898,0.007626771,0.7777686],"study_design_scores_gemma":[0.0002777474,0.00102416,0.003028185,0.000115515,0.000009612141,0.0001086604,0.000004947976,0.9702724,0.002385322,0.02046677,0.001792184,0.0005145049],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.01382433,0.0003209152,0.9335827,0.005797466,0.00006663262,0.0004934,0.000003068971,0.04488453,0.00102702],"genre_scores_gemma":[0.5403733,0.000008199279,0.4547417,0.004517992,0.00008211557,0.000161815,0.000007172563,0.00004429712,0.00006343784],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9658904,"threshold_uncertainty_score":0.9998596,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1928770023","doi":"10.1109/icsm.1989.65194","title":"Insights into regression testing (software testing)","year":2003,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":184,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"","keywords":"Regression testing; Testability; Test plan; Computer science; Non-regression testing; Test (biology); Plan (archaeology); Risk-based testing; Regression analysis; Test Management Approach; Manual testing; Test strategy; Set (abstract data type); Reliability engineering; Regression; Keyword-driven testing; White-box testing; Machine learning; Data mining; Measure (data warehouse); Software; Statistics; Programming language; Mathematics; Engineering; Software system","retraction":null,"screen_n_in":null,"score":{"opus":0.04500660842911066,"gpt":0.2756500233323609,"spread":0.2306434149032503,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.0002913292,0.0002189742,0.0001820918,0.0001642668,0.0003529463,0.0001903551,0.0007287529,0.00009479612,0.000006437092],"category_scores_gemma":[0.01334153,0.0001658811,0.00003897485,0.001260775,0.00004779996,0.0004000069,0.0002453153,0.0001977177,0.00006231046],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005214937,"about_ca_system_score_gemma":0.0001345515,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001390156,"about_ca_topic_score_gemma":0.000003432881,"domain_scores_codex":[0.998454,0.0001089048,0.0002715451,0.0005431021,0.0003099147,0.0003125634],"domain_scores_gemma":[0.9968427,0.001715278,0.0001250384,0.0008639247,0.0003051295,0.0001479522],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000002985186,0.0002534201,0.1566916,0.00009658474,0.00001522048,0.0002184515,0.001277282,0.00009087018,0.004193676,0.05923986,0.05940317,0.7185169],"study_design_scores_gemma":[0.0003681232,0.0003924211,0.007015646,0.0006252732,0.00000889057,0.0002215448,0.00001032305,0.03331197,0.02134519,0.9265258,0.009211374,0.0009633941],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.01244072,0.0002664774,0.934166,0.00008708886,0.0002647283,0.0001325527,7.642689e-8,0.03376392,0.01887846],"genre_scores_gemma":[0.3363511,7.832864e-7,0.6630763,0.0003309466,0.00002473192,0.00001352776,3.790482e-7,0.00001290463,0.000189339],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.867286,"threshold_uncertainty_score":0.9949695,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2125727889","doi":"10.1145/1368088.1368136","title":"Sufficient mutation operators for measuring test effectiveness","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":181,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Mutant; Mutation testing; Test suite; Mutation; Measure (data warehouse); Computer science; Process (computing); Mathematics; Test case; Data mining; Genetics; Biology; Machine learning; Programming language; Gene","retraction":null,"screen_n_in":null,"score":{"opus":0.04459589575011752,"gpt":0.266663017612815,"spread":0.2220671218626974,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003897777,0.00007824453,0.00008648329,0.00007145019,0.0002127296,0.00004785206,0.0002816533,0.00002801187,9.072599e-7],"category_scores_gemma":[0.0008400108,0.00006674838,0.00003643898,0.0002194003,0.00002215729,0.0001623255,0.0000593292,0.0000374531,0.000008091025],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004183252,"about_ca_system_score_gemma":0.00005000261,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003773022,"about_ca_topic_score_gemma":9.309523e-7,"domain_scores_codex":[0.9993442,0.00002928458,0.0001009487,0.0002368239,0.0001353491,0.0001533943],"domain_scores_gemma":[0.998567,0.0009861529,0.00002654075,0.0002384273,0.0001408167,0.0000410248],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00007287495,0.002582564,0.5893288,0.0006406818,0.00009822672,0.0003648295,0.009866153,0.008115077,0.0199405,0.08927833,0.04993952,0.2297725],"study_design_scores_gemma":[0.001327283,0.001204617,0.15019,0.0002148883,0.0000114545,0.0005266875,0.00001316913,0.4772036,0.3468841,0.02074648,0.0007483973,0.000929252],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1494028,0.0000189024,0.8436256,0.00004291695,0.0001072248,0.0002138001,4.773085e-7,0.006269137,0.0003191885],"genre_scores_gemma":[0.7854305,6.428422e-7,0.2143798,0.0000666054,0.00001809448,0.00006721666,7.210958e-7,0.000005602853,0.00003084149],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6360278,"threshold_uncertainty_score":0.272192,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4385187421","doi":"10.1109/sp46215.2023.10179420","title":"Examining Zero-Shot Vulnerability Repair with Large Language Models","year":2023,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":178,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Zero (linguistics); Computer science; Vulnerability (computing); Ground zero; Computer security; Physics; Linguistics; Nuclear physics","retraction":null,"screen_n_in":null,"score":{"opus":0.07209621342546572,"gpt":0.3021107687630803,"spread":0.2300145553376146,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009948318,0.0001296098,0.000143628,0.0001156917,0.0001391435,0.00008586352,0.0004978678,0.00004947151,0.00001165176],"category_scores_gemma":[0.0002059448,0.00009701772,0.00003848615,0.0007605345,0.0000283925,0.00035713,0.000327572,0.0001368062,0.0000725284],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002494656,"about_ca_system_score_gemma":0.00004374569,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009363167,"about_ca_topic_score_gemma":0.00001274869,"domain_scores_codex":[0.998686,0.00006724424,0.0001524341,0.000471103,0.0002635185,0.0003596862],"domain_scores_gemma":[0.9984936,0.0003756176,0.00003979551,0.0009539632,0.0000590954,0.00007792358],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00004721493,0.000579318,0.1628423,0.0002489634,0.0001389828,0.001235702,0.02600465,0.005154631,0.0009278814,0.1769191,0.5435123,0.08238888],"study_design_scores_gemma":[0.0002963881,0.0002355253,0.006326729,0.00006950757,0.000005681428,0.00003877035,0.00007172536,0.9233274,0.001046405,0.06776404,0.0004249555,0.0003928248],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1048523,0.0000192845,0.8097978,0.0001183022,0.00005523905,0.0001050803,0.000001715579,0.07664462,0.008405714],"genre_scores_gemma":[0.7341148,0.000001052392,0.26456,0.0002745136,0.0000171299,0.0000254457,0.000003347113,0.00001105306,0.0009926218],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9181728,"threshold_uncertainty_score":0.3956268,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2092198121","doi":"10.1016/j.jss.2012.12.051","title":"A survey of software testing practices in Canada","year":2013,"lang":"en","type":"article","venue":"Journal of Systems and Software","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":168,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Calgary","funders":"","keywords":"Software testing; Test (biology); Survey research; Software; Engineering; Computer science; Psychology; Applied psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.06050471719038154,"gpt":0.2706308278705806,"spread":0.2101261106801991,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001349842,0.0001197479,0.0003929638,0.000157857,0.00004843632,0.0001292752,0.0004773212,0.00005004222,0.000002033711],"category_scores_gemma":[0.01138015,0.00009395617,0.00002471042,0.0004303693,0.00002183862,0.0006035301,0.0001085529,0.0002184117,6.81901e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001191104,"about_ca_system_score_gemma":0.001164578,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9302604,"about_ca_topic_score_gemma":0.1654199,"domain_scores_codex":[0.9983648,0.0001948367,0.0007088761,0.0001521562,0.0003833287,0.0001959964],"domain_scores_gemma":[0.9939373,0.003075397,0.001666303,0.0002547333,0.0009500952,0.0001161762],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00000205179,0.00001646426,0.9744677,0.0001109158,0.00001080404,0.00003895886,0.0001284361,0.0001026935,0.000009777504,0.000008198695,0.00356105,0.02154298],"study_design_scores_gemma":[0.0002262893,0.0001554923,0.9914261,0.0009114385,0.000005297481,0.0004271615,0.00005987324,0.005858697,0.00001720224,0.0006879078,0.00007059934,0.0001539886],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8908774,0.004114165,0.1037441,0.0001132414,0.0005673674,0.0002255096,0.000007511494,0.0003291726,0.00002147282],"genre_scores_gemma":[0.8934345,0.00001485155,0.1064603,0.00003791907,0.00002966103,0.000004266717,3.215849e-7,0.000007731529,0.00001042039],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7648406,"threshold_uncertainty_score":0.9969474,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2092245922","doi":"10.1016/j.infsof.2013.03.004","title":"Graphical user interface (GUI) testing: Systematic mapping and repository","year":2013,"lang":"en","type":"article","venue":"Information and Software Technology","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":167,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Graphical user interface testing; Graphical user interface; Computer science; User interface; Software; Interface (matter); Information retrieval; Data mining; Software engineering; User interface design; Programming language; Operating system","retraction":null,"screen_n_in":null,"score":{"opus":0.01087614462494123,"gpt":0.2151209586309208,"spread":0.2042448140059796,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000231702,0.0001457953,0.0002234803,0.0005060312,0.0001923013,0.000322007,0.0003902412,0.000198549,0.000001328569],"category_scores_gemma":[0.001733837,0.0001243478,0.00001877112,0.0005458827,0.0001474254,0.00118313,0.0003876874,0.0002132116,0.00002598417],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002045196,"about_ca_system_score_gemma":0.00002397969,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004982732,"about_ca_topic_score_gemma":4.382356e-7,"domain_scores_codex":[0.9989988,0.0000311173,0.0004463728,0.0001812859,0.0001333508,0.0002091277],"domain_scores_gemma":[0.9987386,0.0003429595,0.0001997411,0.0004310494,0.0002182394,0.00006938627],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000006374947,0.0001081391,0.405191,0.01417862,0.0001347352,0.00003192841,0.006435069,0.00001789,0.0006026089,0.07983512,0.02198742,0.4714711],"study_design_scores_gemma":[0.002363859,0.001406915,0.1010688,0.0147573,0.00008532058,0.009974144,0.001259719,0.22585,0.003659748,0.6306054,0.005721265,0.003247482],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.06584708,0.0003449679,0.9139868,0.0008685787,0.0001219507,0.0004564641,4.218467e-7,0.01812021,0.0002535185],"genre_scores_gemma":[0.7004635,0.000008577076,0.2990792,0.000300546,0.000006234316,0.0001091135,6.821765e-7,0.000004918179,0.00002724103],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6346164,"threshold_uncertainty_score":0.5070755,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2136921053","doi":"10.1109/tse.2010.58","title":"The Effects of Time Constraints on Test Case Prioritization: A Series of Controlled Experiments","year":2010,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":166,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"","keywords":"Regression testing; Prioritization; Computer science; Context (archaeology); Risk-based testing; Reliability engineering; Process (computing); Software; Regression analysis; Data mining; Risk analysis (engineering); Machine learning; Software development; Engineering; Management science; Software construction","retraction":null,"screen_n_in":null,"score":{"opus":0.005484845015205005,"gpt":0.2193811498363502,"spread":0.2138963048211452,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001921477,0.0001673611,0.0002514268,0.0001303256,0.0001570345,0.00004284748,0.0002972658,0.00007926135,0.000005816336],"category_scores_gemma":[0.0007819458,0.0001354089,0.0001009785,0.0002579572,0.0001058704,0.0001401755,0.000003675227,0.0002420503,0.000003967388],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001855343,"about_ca_system_score_gemma":0.00004599855,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000008528981,"about_ca_topic_score_gemma":0.000001421779,"domain_scores_codex":[0.9990958,0.00002404995,0.0002913794,0.0001966999,0.0002119601,0.0001801088],"domain_scores_gemma":[0.9957479,0.003481106,0.0001004667,0.0004915614,0.0001182754,0.00006072753],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"bench_or_experimental","study_design_scores_codex":[0.0006458235,0.003760255,0.001004239,0.001708043,0.001130773,0.001813411,0.007200509,0.04914959,0.3938009,0.006745785,0.002264252,0.5307764],"study_design_scores_gemma":[0.002701499,0.001191204,0.0001592826,0.0005791385,0.00005098992,0.0009648604,0.000009760905,0.04585632,0.9476696,0.0002782524,0.00008208014,0.0004569596],"study_design_candidate":"bench_or_experimental","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.03091574,0.00003299535,0.9655684,0.0000192605,0.0007684826,0.0003418628,0.000008414631,0.0023229,0.00002193386],"genre_scores_gemma":[0.9067792,0.00000693201,0.09302247,0.00001311689,0.00001906415,0.0000864086,2.829476e-7,0.00001729519,0.00005519011],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8758635,"threshold_uncertainty_score":0.5521815,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2045837563","doi":"10.1007/s10664-012-9219-7","title":"Static test case prioritization using topic models","year":2012,"lang":"en","type":"article","venue":"Empirical Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":156,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Test suite; Computer science; Code coverage; Test case; Source code; Black box; Test Management Approach; Programming language; White-box testing; Test (biology); Test data; Operating system; Software; Software development; Artificial intelligence; Machine learning","retraction":null,"screen_n_in":null,"score":{"opus":0.06432137932201937,"gpt":0.3036697114581389,"spread":0.2393483321361196,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002420302,0.0001730436,0.0001598542,0.0001191105,0.0001122495,0.00009755384,0.0002551256,0.00008151158,0.000003972741],"category_scores_gemma":[0.001400123,0.0001739218,0.0000486347,0.0004492367,0.00001219212,0.0007973905,0.0001690105,0.0001706461,0.000008227141],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001260916,"about_ca_system_score_gemma":0.00003353302,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003682193,"about_ca_topic_score_gemma":2.968668e-7,"domain_scores_codex":[0.9988658,0.00002193053,0.0002247727,0.0002323451,0.0002037142,0.0004514626],"domain_scores_gemma":[0.9985216,0.0008052508,0.00004636206,0.0003727608,0.00005943031,0.0001945704],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000002932833,0.0005641402,0.7631552,0.0004284631,0.00005890921,0.001289744,0.005828251,0.1719298,0.0002808286,0.003240478,0.009944203,0.04327707],"study_design_scores_gemma":[0.0001116512,0.00003663657,0.003398017,0.00007515762,0.00001100808,0.001471306,0.000002664606,0.9907523,0.000249052,0.002996732,0.0005388826,0.0003565747],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.07237958,0.0001909417,0.9180328,0.00006440358,0.0002831485,0.00009874723,0.000001736448,0.008934137,0.00001452531],"genre_scores_gemma":[0.5370809,0.000001168149,0.4626541,0.0001249468,0.0001053682,0.000007784058,0.000001127405,0.00001541916,0.000009191915],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8188225,"threshold_uncertainty_score":0.7092326,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2116368564","doi":"10.1145/1985793.1985870","title":"Automated cross-browser compatibility testing","year":2011,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":147,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Compatibility (geochemistry); Computer science; Web browser; Client-side scripting; Emulation; World Wide Web; Backward compatibility; Web-based simulation; Pairwise comparison; Web navigation; Software engineering; Information retrieval; Web page; The Internet; Web API; Operating system; Engineering; Artificial intelligence","retraction":null,"screen_n_in":null,"score":{"opus":0.1194683359133998,"gpt":0.3234254279076075,"spread":0.2039570919942077,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004233979,0.0001121967,0.0001152379,0.0000567705,0.0001280424,0.0001194859,0.0007575767,0.00004925479,0.00003876772],"category_scores_gemma":[0.0007413682,0.00009359887,0.00003169259,0.0004100059,0.00006653432,0.000334033,0.0002831185,0.00009055826,0.00008923702],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002205077,"about_ca_system_score_gemma":0.0000408713,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0007095289,"about_ca_topic_score_gemma":0.000004488732,"domain_scores_codex":[0.9990091,0.00004026972,0.0002130348,0.0003458297,0.0001478514,0.0002439367],"domain_scores_gemma":[0.9986652,0.0002801137,0.00006547882,0.0007273449,0.0001834313,0.00007847267],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00000314521,0.0001663031,0.9475484,0.00001935571,0.000007932865,0.00002462002,0.0007728066,0.000009620508,0.0002086726,0.008610988,0.01577791,0.02685026],"study_design_scores_gemma":[0.0001092966,0.00009283931,0.5505617,0.00002354917,0.000001725152,0.00003209077,0.000001042807,0.3794353,0.005081506,0.06438921,0.00005305656,0.0002186619],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1576886,0.00001211968,0.6194055,0.00003068669,0.0001808734,0.0001248195,5.935973e-7,0.1705699,0.05198683],"genre_scores_gemma":[0.5242279,4.400581e-8,0.4755566,0.0001381072,0.000007498592,0.000004525715,2.49964e-7,0.000004061234,0.0000609331],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.3969867,"threshold_uncertainty_score":0.3816851,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2158798798","doi":"10.1145/1368088.1368118","title":"Calysto","year":2008,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":146,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Computer science; Scalability; Static analysis; Source lines of code; Automation; Static program analysis; Context (archaeology); Software; Software bug; Model checking; Code (set theory); Computer engineering; Programming language; Software development; Database; Set (abstract data type)","retraction":null,"screen_n_in":null,"score":{"opus":0.03481987681072068,"gpt":0.236430062464799,"spread":0.2016101856540783,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00005568577,0.00003430307,0.00003956861,0.00003226493,0.00006834885,0.000012685,0.0003119666,0.00001486988,0.000007914276],"category_scores_gemma":[0.00006141039,0.00002742811,0.00001744367,0.000139677,0.00001643445,0.00009293381,0.00007734033,0.0000300464,0.0000879103],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000006648134,"about_ca_system_score_gemma":0.00001798175,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005092239,"about_ca_topic_score_gemma":5.220795e-7,"domain_scores_codex":[0.9996661,0.000007812047,0.00004933847,0.0001102929,0.00008090577,0.00008553695],"domain_scores_gemma":[0.9995939,0.00005573565,0.000009987131,0.0002868031,0.00002203847,0.00003158068],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[6.020704e-7,0.00004841759,0.02795497,0.000002684585,0.000004571134,0.0001389104,0.0005086487,0.000005345461,0.0001088077,0.06476039,0.8576487,0.04881797],"study_design_scores_gemma":[0.0006915327,0.0004399863,0.09208687,0.00004987613,0.000006053776,0.003760757,0.000004047221,0.184661,0.02694485,0.5803423,0.1095257,0.001486949],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.0044194,0.00001970616,0.9543152,0.0003021186,0.00005315264,0.00001766678,3.05064e-8,0.01980166,0.02107108],"genre_scores_gemma":[0.5422674,0.000001508985,0.4564103,0.0004208136,0.00001142633,0.000002231345,6.648552e-8,0.000001432058,0.000884847],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7481229,"threshold_uncertainty_score":0.1129938,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1607747749","doi":"","title":"Specification-based regression test selection with risk analysis","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":143,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Regression testing; Computer science; Test suite; Test Management Approach; Test case; Regression analysis; Traceability; Risk-based testing; Data mining; Unified Modeling Language; Reliability engineering; Machine learning; Software engineering; Programming language; Software; Engineering; Software development","retraction":null,"screen_n_in":null,"score":{"opus":0.01913616712242833,"gpt":0.228583529516784,"spread":0.2094473623943557,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001540663,0.00008893486,0.00009844838,0.0002520496,0.0001590686,0.0001043117,0.0002853804,0.00003387842,0.0000894202],"category_scores_gemma":[0.0002027447,0.00005960299,0.00004383565,0.001982719,0.0000198075,0.0001325799,0.00001934905,0.00008629404,0.00003942101],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000312556,"about_ca_system_score_gemma":0.00001080988,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001178093,"about_ca_topic_score_gemma":0.00002533035,"domain_scores_codex":[0.9992281,0.00003755043,0.0001129967,0.0002986861,0.0001982327,0.0001244049],"domain_scores_gemma":[0.9988723,0.0004054323,0.0001035304,0.0004632091,0.0001071272,0.00004834772],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001910548,0.0001768923,0.9091217,0.00000305048,0.00003889639,0.000003385313,0.00008358416,0.001185973,0.00006006388,0.0005285446,0.03939553,0.04940046],"study_design_scores_gemma":[0.00008677049,0.0001133119,0.05080521,0.0000118685,0.00004456945,0.000003488539,6.87823e-7,0.9452366,0.002105612,0.001004083,0.0004625999,0.0001251845],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.002357331,0.00001810465,0.9851276,0.0005153825,0.00001425662,0.00005559767,5.394043e-7,0.008956606,0.002954592],"genre_scores_gemma":[0.5906422,0.00000366282,0.408815,0.00009377509,0.00001583267,0.000008142538,0.000001147545,0.000003448908,0.0004168358],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9440507,"threshold_uncertainty_score":0.2430539,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1483143768","doi":"10.1007/3-540-45441-1_15","title":"A UML-Based Approach to System Testing","year":2001,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":137,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Computer science; Testability; Unified Modeling Language; Sequence diagram; Software engineering; Class diagram; Test case; Model-based testing; Programming language; Reliability engineering; Software; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.0394705976459676,"gpt":0.2527500202852448,"spread":0.2132794226392772,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001443631,0.0006620615,0.0006460708,0.001373527,0.0003553962,0.0007988262,0.004627149,0.0003378873,0.000001592175],"category_scores_gemma":[0.0006367979,0.0006145979,0.0001237917,0.001790195,0.0003199836,0.0002849417,0.001333161,0.0007867332,0.00004921692],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006179955,"about_ca_system_score_gemma":0.0007316036,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00008936403,"about_ca_topic_score_gemma":0.000004844915,"domain_scores_codex":[0.9951694,0.00004645368,0.000588636,0.002108128,0.00118843,0.0008989609],"domain_scores_gemma":[0.9956309,0.001151465,0.0002925079,0.002174639,0.00043091,0.0003195431],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000003848422,0.00004988409,0.0003090445,0.0001552794,0.000006084849,0.0001519454,0.0002983454,0.08856788,0.00002392487,0.01133069,0.0002457205,0.8988574],"study_design_scores_gemma":[0.0001376434,0.000198264,0.00006119987,0.001439772,0.000006545478,0.0002639338,3.805614e-8,0.942509,0.0002019571,0.0538101,0.0005683819,0.0008031727],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.00001068266,0.0001154849,0.9621968,0.0002134155,0.0007661197,0.000554162,0.000002615377,0.008670566,0.02747018],"genre_scores_gemma":[0.1364232,6.379529e-7,0.8615652,0.001482136,0.0003420378,0.00003338743,0.000002374253,0.00004727607,0.0001037103],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.8980542,"threshold_uncertainty_score":0.9996305,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2102418068","doi":"10.1109/tse.2002.1049402","title":"Timed Wp-method: testing real-time systems","year":2002,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":135,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Concordia University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Nondeterministic algorithm; Correctness; Timed automaton; Automaton; System under test; Finite-state machine; Fault coverage; Real-time operating system; Test case; Real-time computing; Distributed computing; Algorithm; Theoretical computer science; Embedded system","retraction":null,"screen_n_in":null,"score":{"opus":0.02672061398918014,"gpt":0.2342479938867207,"spread":0.2075273798975405,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0003782933,0.0003092425,0.0003082942,0.0003753713,0.0002229118,0.0002047865,0.0006200568,0.0001412799,0.0000297764],"category_scores_gemma":[0.0002562958,0.0003291331,0.0001115182,0.0009361383,0.00001723565,0.0003573619,0.000006264544,0.0003745235,0.0002462812],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001278952,"about_ca_system_score_gemma":0.00001997557,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001036231,"about_ca_topic_score_gemma":2.323379e-7,"domain_scores_codex":[0.9982241,0.00005960484,0.0003645822,0.0005223097,0.0003500289,0.0004793927],"domain_scores_gemma":[0.9972542,0.00164602,0.00008248465,0.0007287334,0.0001139025,0.0001746809],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000004428416,0.0002325413,0.0001208326,0.0001807929,0.00009352344,0.0001175525,0.0006021577,0.8072634,0.005019221,0.0001839124,0.007396425,0.1787852],"study_design_scores_gemma":[0.0001599194,0.0001058971,0.00006467586,0.0002474296,0.00001506102,0.0001634677,0.000001141796,0.9947463,0.003706865,0.00005587337,0.0003263582,0.0004070195],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.0007546978,0.00006639904,0.9630008,0.00002651513,0.0007475171,0.0001954917,0.000007005068,0.0349737,0.000227834],"genre_scores_gemma":[0.1373837,0.00001150655,0.8617237,0.00003760563,0.00008093422,0.00009678226,6.460476e-7,0.00005885443,0.0006063022],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.1874829,"threshold_uncertainty_score":0.9999161,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2163289685","doi":"10.1109/icsm.2002.1167775","title":"Automating impact analysis and regression test selection based on UML designs","year":2003,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":131,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Unified Modeling Language; Regression testing; Selection (genetic algorithm); Test (biology); Regression analysis; Data mining; Programming language; Software engineering; Artificial intelligence; Machine learning; Software; Software development","retraction":null,"screen_n_in":null,"score":{"opus":0.02588388994970477,"gpt":0.3075896533459878,"spread":0.2817057633962831,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004726553,0.0001130452,0.0001283853,0.0003291656,0.0001679115,0.0001502484,0.0001325255,0.00004736594,0.00001869194],"category_scores_gemma":[0.001084707,0.00007894233,0.00006048604,0.001297574,0.00001272197,0.0001372398,0.00002048352,0.00008079298,0.000003323829],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004780489,"about_ca_system_score_gemma":0.00004610143,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009052583,"about_ca_topic_score_gemma":0.0000051023,"domain_scores_codex":[0.9991733,0.00009550049,0.0001218881,0.0002808133,0.000162784,0.0001657218],"domain_scores_gemma":[0.9987966,0.0007335833,0.00007000822,0.0002775821,0.00005163908,0.00007056337],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001961741,0.00009987145,0.9341925,0.000007848954,0.00003078224,0.000004960274,0.00007111642,0.001229183,0.001194068,0.0005710341,0.003156422,0.05944028],"study_design_scores_gemma":[0.00007860749,0.0002114335,0.08177032,0.00003203756,0.00002215126,0.000008593113,4.781967e-7,0.9082828,0.007714846,0.001757766,0.000005217329,0.0001158053],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.02545508,0.000007718044,0.966273,0.00005648414,0.0000164341,0.00005674849,2.594054e-7,0.006776038,0.001358312],"genre_scores_gemma":[0.6156365,4.040677e-7,0.3842144,0.0001027301,0.000004144526,0.000003780026,5.036206e-7,0.000003381552,0.00003411871],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9070536,"threshold_uncertainty_score":0.3219175,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2741328617","doi":"10.1145/3106237.3106274","title":"Better test cases for better automated program repair","year":2017,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":126,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Computer science; Overfitting; Metric (unit); Artificial intelligence; Benchmark (surveying); Test (biology); Machine learning; Software bug; Pattern recognition (psychology); Data mining; Software; Programming language; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.04420587928543526,"gpt":0.3430140819929544,"spread":0.2988082027075191,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000250762,0.0001362871,0.0001378065,0.00005396146,0.0004880643,0.0005911202,0.001016095,0.0000678399,0.000005371109],"category_scores_gemma":[0.001568299,0.0001083177,0.0001021147,0.00005320075,0.00007351606,0.0003918796,0.0003047973,0.00006688286,0.0000263465],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001674226,"about_ca_system_score_gemma":0.0000289107,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001565495,"about_ca_topic_score_gemma":0.00001037937,"domain_scores_codex":[0.9990013,0.00001268941,0.0001675401,0.0003815381,0.000132806,0.0003041333],"domain_scores_gemma":[0.997699,0.0005960267,0.0001126811,0.001408649,0.0001162633,0.00006740628],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001280816,0.00009917173,0.05822735,0.00001840596,0.000008991991,0.0000449235,0.00002861679,1.675205e-7,0.00001453746,0.0002222653,0.8445455,0.09678876],"study_design_scores_gemma":[0.0003348007,0.0006621883,0.05058042,0.00007898876,0.00001325127,0.0001397533,6.732137e-7,0.9022577,0.001256721,0.0161507,0.02810263,0.0004221803],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"software","genre_gemma":"methods","genre_scores_codex":[0.0201278,0.00002473986,0.11176,0.009821402,0.0004534515,0.001149699,0.00000896188,0.8526846,0.003969321],"genre_scores_gemma":[0.1464303,3.479909e-7,0.8512427,0.001576481,0.00009914507,0.0002193799,0.000002456325,0.00001293409,0.0004161889],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9022576,"threshold_uncertainty_score":0.5700185,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1967187747","doi":"10.1007/s10515-011-0093-0","title":"Prioritizing test cases with string distances","year":2011,"lang":"en","type":"article","venue":"Automated Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":125,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Computer Research Institute of Montréal","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Test suite; String (physics); Computer science; Random testing; Code (set theory); Test case; Algorithm; Property (philosophy); Suite; Test (biology); Data mining; Mathematics; Machine learning; Programming language","retraction":null,"screen_n_in":null,"score":{"opus":0.01989486036119929,"gpt":0.2132109271948698,"spread":0.1933160668336705,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00015651,0.0002463646,0.0002019428,0.0001675568,0.0001326558,0.000134231,0.0005948676,0.00006828119,0.000002465036],"category_scores_gemma":[0.0009920719,0.0002221566,0.00003908033,0.0005788426,0.00002847995,0.0005427253,0.0001444359,0.0001764498,0.00001052801],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005773294,"about_ca_system_score_gemma":0.00004668412,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001073039,"about_ca_topic_score_gemma":0.000004500158,"domain_scores_codex":[0.9987245,0.00001180847,0.000219857,0.000388863,0.0002242166,0.0004307446],"domain_scores_gemma":[0.998471,0.0007322209,0.00007672774,0.0005236892,0.0000810785,0.0001152544],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00005477021,0.0009697861,0.785522,0.00127523,0.0004710796,0.01231908,0.01477371,0.004442904,0.002822207,0.03892102,0.01044968,0.1279785],"study_design_scores_gemma":[0.001555195,0.00183567,0.294448,0.004205625,0.0001075921,0.00470699,0.00006854574,0.6601063,0.02324806,0.003551501,0.001605045,0.00456149],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.06265202,0.0002790896,0.8099237,0.000007692455,0.0001662012,0.0001249204,0.000004552677,0.1267603,0.00008151012],"genre_scores_gemma":[0.5588182,0.000003591921,0.4410919,0.00001680769,0.00001979683,0.00001707644,0.000001588924,0.00002338822,0.000007682995],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6556634,"threshold_uncertainty_score":0.9059281,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2157056800","doi":"10.1145/2535838.2535857","title":"Symbolic optimization with SMT solvers","year":2014,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":122,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Computer science; Satisfiability modulo theories; Programming language; Symbolic execution; Satisfiability; Relation (database); Theoretical computer science; Software; Subtyping; Concolic testing; Data mining","retraction":null,"screen_n_in":null,"score":{"opus":0.005850794322010276,"gpt":0.1925467586558487,"spread":0.1866959643338385,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001349322,0.00005935559,0.00005843321,0.00004977782,0.00005793329,0.00008118449,0.0002979082,0.00002200819,0.000007793046],"category_scores_gemma":[0.00006535051,0.00004301586,0.00001166715,0.000183158,0.00001770587,0.0001605926,0.00005479838,0.00003621585,0.00001403338],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000009100006,"about_ca_system_score_gemma":0.000015914,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004048452,"about_ca_topic_score_gemma":0.000001448842,"domain_scores_codex":[0.9995344,0.0000181894,0.00005792857,0.0001651806,0.0001093367,0.0001150149],"domain_scores_gemma":[0.9994745,0.00008600506,0.00002592112,0.0003294195,0.00004486422,0.00003930481],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0000144748,0.0001933727,0.03939256,0.00004317884,0.00003995641,0.0000146081,0.001488058,0.1283201,0.00007385982,0.3961122,0.14832,0.2859876],"study_design_scores_gemma":[0.0001117956,0.0001097303,0.0004532817,0.00001692641,0.000001985479,0.00001506417,9.496511e-7,0.9891642,0.0004318613,0.008719659,0.0008511132,0.0001234018],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.0004055155,0.000002907893,0.9736347,0.0003045857,0.00003765128,0.00003713823,3.387635e-8,0.01118444,0.01439307],"genre_scores_gemma":[0.325068,6.111603e-7,0.6743377,0.0004269491,0.00001506253,0.000004151344,4.513333e-7,0.000003789221,0.0001432837],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.8608441,"threshold_uncertainty_score":0.1754136,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3196239222","doi":"10.1145/3468264.3468591","title":"A comprehensive study of deep learning compiler bugs","year":2021,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":112,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"National Key Research and Development Program of China; National Natural Science Foundation of China","keywords":"Compiler; Computer science; Programming language; Boosting (machine learning); Deep learning; Context (archaeology); Code (set theory); Code generation; Parallel computing; Artificial intelligence; Operating system","retraction":null,"screen_n_in":null,"score":{"opus":0.03173090442745961,"gpt":0.283983725789503,"spread":0.2522528213620434,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00007800895,0.0000830554,0.000171662,0.00005736224,0.00007637463,0.0000512642,0.0003457524,0.00002566531,0.00002372636],"category_scores_gemma":[0.0001399718,0.00007575578,0.0000345037,0.0003802598,0.00001956192,0.00009115512,0.0003850565,0.0001272768,0.00001196421],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000009669976,"about_ca_system_score_gemma":0.00002264069,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001015481,"about_ca_topic_score_gemma":0.000009364899,"domain_scores_codex":[0.9991022,0.0001231756,0.0001773851,0.0002602906,0.0002027551,0.0001341524],"domain_scores_gemma":[0.9988021,0.0003762879,0.00006418298,0.0004276053,0.0002909381,0.0000389421],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00002062044,0.004323571,0.4403572,0.0001577613,0.0002639222,0.0007527735,0.02967495,0.003232036,0.005186387,0.01674093,0.01756378,0.4817261],"study_design_scores_gemma":[0.003602179,0.003839559,0.3778014,0.0002166713,0.00006194551,0.0004472567,0.004212062,0.5353408,0.01995663,0.04398524,0.008983401,0.001552788],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.330675,0.0001411585,0.6591977,0.00009361526,0.0001102752,0.00011171,1.096487e-7,0.006527221,0.003143224],"genre_scores_gemma":[0.8362119,0.000001869784,0.1634878,0.0001396571,0.00001024636,0.00000532713,6.100282e-7,0.000004834847,0.0001376912],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.5321088,"threshold_uncertainty_score":0.3089231,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2162032875","doi":"10.1109/real.1998.739748","title":"Timed test cases generation based on state characterization technique","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":112,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Concordia University; Université de Montréal","funders":"","keywords":"Computer science; Nondeterministic algorithm; Correctness; Test suite; Automaton; Timed automaton; Finite-state machine; Set (abstract data type); System under test; State (computer science); Granularity; Algorithm; Test case; Programming language; Theoretical computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.0420348950945121,"gpt":0.2411333431361208,"spread":0.1990984480416087,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001702445,0.0001130308,0.00008306289,0.0001456255,0.0001051599,0.0001423449,0.0002579917,0.00004389209,0.00006173774],"category_scores_gemma":[0.00038569,0.00009931971,0.00002494976,0.0003021116,0.00001533707,0.0002284614,0.0000361135,0.00007051438,0.00007486744],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003538471,"about_ca_system_score_gemma":0.00001408372,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001886291,"about_ca_topic_score_gemma":0.000001354813,"domain_scores_codex":[0.9991993,0.00004196449,0.0001515366,0.000276569,0.0001778294,0.0001527469],"domain_scores_gemma":[0.9990884,0.0002758241,0.00006516473,0.0004485521,0.00007316704,0.00004888123],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000007270382,0.001185616,0.01047243,0.00004034793,0.000009212687,0.0002247146,0.0003402577,0.0004850362,0.2905959,0.002083527,0.1266578,0.5678979],"study_design_scores_gemma":[0.00007421912,0.0002308653,0.0008861456,0.00002328323,0.000001213476,0.00002289175,6.750255e-8,0.9055538,0.09186279,0.0004576246,0.0007323535,0.0001547422],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.002520254,0.0000029103,0.9840991,0.0006655232,0.00006551842,0.0002195226,0.000003957762,0.01119341,0.001229757],"genre_scores_gemma":[0.7468537,0.000003100286,0.2508786,0.001386646,0.0000478364,0.0000860668,0.00001936034,0.00001074584,0.0007139798],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9050688,"threshold_uncertainty_score":0.405014,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2151635819","doi":"10.1145/1276958.1277172","title":"Automatic mutation test input data generation via ant colony","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":109,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Polytechnique Montréal","funders":"Natural Sciences and Engineering Research Council of Canada; Canada Research Chairs","keywords":"Computer science; Ant colony optimization algorithms; Mutation; Context (archaeology); Genetic algorithm; Java; Code coverage; Test data; Data mining; Machine learning; Artificial intelligence; Software","retraction":null,"screen_n_in":null,"score":{"opus":0.05954161236665662,"gpt":0.3142915271483908,"spread":0.2547499147817342,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0009898989,0.00008177898,0.00007704211,0.00009719279,0.0001064327,0.0001341966,0.0007502942,0.00004536868,0.00001051456],"category_scores_gemma":[0.0007845871,0.0000721352,0.00001134626,0.0003104967,0.00001933808,0.000492656,0.0002727173,0.00005995757,0.00004701846],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003770458,"about_ca_system_score_gemma":0.00004418172,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001317391,"about_ca_topic_score_gemma":0.00004339989,"domain_scores_codex":[0.9990789,0.00002185551,0.0002144192,0.0003133523,0.0002001849,0.0001712519],"domain_scores_gemma":[0.9983753,0.0004928101,0.00007525138,0.0009323946,0.00007248006,0.00005178159],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[5.872096e-7,0.0001591945,0.008895942,0.00001841641,0.000007619296,0.00007453882,0.0003833946,0.00002425812,0.004370254,0.002004207,0.1269421,0.8571195],"study_design_scores_gemma":[0.00006859501,0.00007455061,0.006543418,0.00001086149,0.00000303574,0.00007165445,9.820931e-7,0.9830382,0.004728157,0.005041384,0.0003086346,0.0001105527],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.01708328,0.00001906659,0.9714015,0.0003114029,0.0001661489,0.000112919,0.000001237695,0.01026739,0.0006370358],"genre_scores_gemma":[0.5041803,5.755647e-7,0.4953302,0.000340082,0.00006872648,0.000002573323,0.00002383102,0.000003885746,0.00004977384],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9830139,"threshold_uncertainty_score":0.2941588,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2036896594","doi":"10.1016/j.infsof.2013.02.006","title":"A systematic mapping study of web application testing","year":2013,"lang":"en","type":"article","venue":"Information and Software Technology","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":108,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia; University of Calgary","funders":"","keywords":"Bibliometrics; Computer science; Data science; Field (mathematics); Schema (genetic algorithms); Information retrieval; Systematic review; World Wide Web; MEDLINE","retraction":null,"screen_n_in":null,"score":{"opus":0.01289810841019044,"gpt":0.2246790883964861,"spread":0.2117809799862957,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002722041,0.0001012583,0.0002157072,0.0005718073,0.00009949078,0.00008455176,0.0004393714,0.00009718306,9.101914e-7],"category_scores_gemma":[0.001494459,0.00008865416,0.00001368049,0.0008966145,0.00004224016,0.000827916,0.0002348153,0.0001062668,0.00003050473],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001892925,"about_ca_system_score_gemma":0.00002701849,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001010941,"about_ca_topic_score_gemma":0.00000182875,"domain_scores_codex":[0.9990491,0.00002503709,0.0005211834,0.0001258561,0.0001471393,0.0001316557],"domain_scores_gemma":[0.9985707,0.0002604692,0.0003469009,0.0004829166,0.000309302,0.00002967074],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001777753,0.000277513,0.2528964,0.006278553,0.00007019582,0.000002131798,0.006678794,0.00004921345,0.0002943448,0.01588373,0.001162663,0.7164047],"study_design_scores_gemma":[0.00246202,0.001817023,0.05631425,0.004173302,0.00005612393,0.0005048336,0.004144008,0.6778311,0.0005615389,0.2506283,0.0002367825,0.001270683],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.2325947,0.00004918363,0.7547653,0.0001396313,0.00003240189,0.0009764613,4.026655e-7,0.01131455,0.0001273146],"genre_scores_gemma":[0.8093665,0.000001162086,0.190235,0.00007028227,0.000002547609,0.0003183191,8.027576e-7,0.000002957172,0.000002434379],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.715134,"threshold_uncertainty_score":0.3615212,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2043839920","doi":"10.1016/j.jss.2014.01.010","title":"Web application testing: A systematic literature review","year":2014,"lang":"en","type":"article","venue":"Journal of Systems and Software","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":108,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Computer science; Web testing; Dependability; Systematic review; Field (mathematics); Web application; World Wide Web; Empirical research; Data science; Set (abstract data type); Web engineering; Software engineering; Information retrieval; Web application security; The Internet; Web development","retraction":null,"screen_n_in":null,"score":{"opus":0.01622432630285369,"gpt":0.2516204591044873,"spread":0.2353961328016336,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001912791,0.0001307524,0.0005026064,0.0001265166,0.00008645634,0.0002508042,0.0004623088,0.00006635509,1.230194e-7],"category_scores_gemma":[0.003068886,0.00008831288,0.00006840118,0.0004153416,0.00001588944,0.000292077,0.00006576414,0.0001816141,0.00000275472],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002424431,"about_ca_system_score_gemma":0.00004613293,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004104212,"about_ca_topic_score_gemma":1.459642e-7,"domain_scores_codex":[0.9985767,0.0001817472,0.0006417306,0.0001601422,0.0003156138,0.0001240886],"domain_scores_gemma":[0.9973716,0.0007071591,0.0007854306,0.0004139043,0.0006111374,0.0001107968],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"systematic_review","study_design_gemma":"systematic_review","study_design_scores_codex":[0.0000095531,0.0002014838,0.02752645,0.7771004,0.0001747088,0.0002028885,0.001212888,0.00008333251,0.0003258532,0.01916789,0.09112327,0.08287123],"study_design_scores_gemma":[0.0006432261,0.0008241173,0.001258937,0.8988911,0.0001772035,0.01500354,0.00001373881,0.06318143,0.00001567392,0.01331745,0.006019725,0.0006538676],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.0004094358,0.146252,0.8512502,0.0002430003,0.0002713471,0.000356931,0.000001005014,0.00114653,0.00006955572],"genre_scores_gemma":[0.6251572,0.004557907,0.3684712,0.001049951,0.000537314,0.00007638586,0.000001388861,0.00003214914,0.0001164101],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6247478,"threshold_uncertainty_score":0.3673965,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4254309703","doi":"10.1007/s10270-002-8208-5","title":"A UML-Based Approach to System Testing","year":2002,"lang":"en","type":"article","venue":"Software & Systems Modeling","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":107,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Computer science; Unified Modeling Language; Sequence diagram; Class diagram; Object Constraint Language; Software engineering; Applications of UML; Testability; Model-based testing; Context (archaeology); Test case; Programming language; UML tool; Test Management Approach; System testing; Activity diagram; Software development; Reliability engineering; Software; Software construction","retraction":null,"screen_n_in":null,"score":{"opus":0.09020794658056883,"gpt":0.2392422866175647,"spread":0.1490343400369959,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0009811245,0.0003869236,0.0004816003,0.0003797993,0.0004285632,0.0005823767,0.001422891,0.000161883,6.658205e-7],"category_scores_gemma":[0.001130128,0.0003706245,0.0001172616,0.001226306,0.00001928672,0.0003293241,0.0002815387,0.0002723805,0.0001118952],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002764631,"about_ca_system_score_gemma":0.00006921546,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005637488,"about_ca_topic_score_gemma":5.240738e-7,"domain_scores_codex":[0.996735,0.000164919,0.000681472,0.0009957913,0.0006867924,0.0007360195],"domain_scores_gemma":[0.9972215,0.0004530426,0.0001822278,0.001433205,0.0003928039,0.0003172652],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000002866236,0.0001230427,0.002449024,0.0006973682,0.00001995374,0.00003637305,0.0008334672,0.9679515,0.00003220628,0.002499116,0.004524813,0.02083027],"study_design_scores_gemma":[0.0001732372,0.00006874737,0.000005336015,0.0009278898,0.00001008086,0.0001378448,0.00003010248,0.9978403,0.00001105766,0.0002206707,0.0001173619,0.0004573236],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.001712759,0.0005651311,0.9526383,0.00004528631,0.000542793,0.0005968832,0.000004098613,0.04176677,0.002128015],"genre_scores_gemma":[0.5439675,2.148012e-7,0.4555323,0.0001202785,0.0001246779,0.0001683323,0.000001524739,0.00003598197,0.00004925389],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5422547,"threshold_uncertainty_score":0.9998746,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2550120381","doi":"10.48550/arxiv.1611.01855","title":"Neuro-Symbolic Program Synthesis","year":2016,"lang":"en","type":"article","venue":"arXiv (Cornell University)","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":105,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Correctness; Computer science; Construct (python library); Program synthesis; Set (abstract data type); Representation (politics); Artificial neural network; Domain (mathematical analysis); Task (project management); String (physics); Theoretical computer science; Regular expression; Algorithm; Artificial intelligence; Programming language; Mathematics","retraction":null,"screen_n_in":null,"score":{"opus":0.06483697021880241,"gpt":0.1871020113852399,"spread":0.1222650411664375,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001409772,0.0001272149,0.0001202595,0.0001334061,0.0001100225,0.00005198935,0.000990942,0.00005721904,0.00001435502],"category_scores_gemma":[0.0003739074,0.0001012972,0.00007586819,0.0005535071,0.00008193262,0.0003775705,0.0002697953,0.00006300011,0.0001536139],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004665439,"about_ca_system_score_gemma":0.00003840972,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002924522,"about_ca_topic_score_gemma":0.000002251924,"domain_scores_codex":[0.9990078,0.00006966562,0.00008397154,0.0004977114,0.0000559267,0.0002848999],"domain_scores_gemma":[0.9984072,0.0006050533,0.00006097014,0.0007453828,0.00006531374,0.0001161232],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00002545938,0.0004003957,0.07438581,0.00002376182,0.00005075348,0.0007508154,0.0001243302,0.0001254837,0.0006361675,0.366272,0.02110923,0.5360959],"study_design_scores_gemma":[0.001424251,0.0008415693,0.04335064,0.0004836104,0.0001078472,0.0001569274,0.00001205078,0.2005673,0.01236432,0.7075378,0.03102504,0.002128674],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1155533,0.00001035763,0.8623358,0.0005175672,0.000142874,0.000165901,0.000001703949,0.01844469,0.002827866],"genre_scores_gemma":[0.9824035,0.00001681512,0.01686344,0.0001290557,0.00002661394,0.000002363495,1.120435e-7,0.00000999783,0.0005480882],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8668503,"threshold_uncertainty_score":0.4130781,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2129718707","doi":"10.1016/j.infsof.2006.11.002","title":"A state-based approach to integration testing based on UML models","year":2006,"lang":"en","type":"article","venue":"Information and Software Technology","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":103,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Unified Modeling Language; Computer science; Integration testing; Test case; Class diagram; Reliability engineering; Sequence diagram; Data mining; Software; Software engineering; Programming language; Engineering; Machine learning","retraction":null,"screen_n_in":null,"score":{"opus":0.01904630375427704,"gpt":0.2255805552280354,"spread":0.2065342514737583,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002762242,0.0001677324,0.0001512245,0.0009256242,0.0001655003,0.0001901549,0.000403656,0.0001484796,5.507846e-7],"category_scores_gemma":[0.0007224776,0.000152567,0.00002387329,0.001051595,0.0000538751,0.0006603693,0.00008832583,0.0001896377,0.00001579927],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006384207,"about_ca_system_score_gemma":0.00008715895,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001126639,"about_ca_topic_score_gemma":0.000003291826,"domain_scores_codex":[0.9989631,0.00002104367,0.0003347663,0.0002230308,0.0002144874,0.0002435384],"domain_scores_gemma":[0.9989276,0.0002004066,0.0001334273,0.0004559792,0.0002308201,0.00005178299],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00001356123,0.00009124882,0.003367295,0.00003102967,0.000002075945,0.000001737605,0.0001691019,0.07236279,0.00001467604,0.02648069,0.003750496,0.8937153],"study_design_scores_gemma":[0.0002563268,0.0001831853,0.0005385278,0.00005626344,0.000001671318,0.000009399037,0.000004310181,0.9234101,0.0007707534,0.07425653,0.0003271239,0.00018581],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.002767819,0.000008449934,0.9765669,0.0006522191,0.00004166648,0.0002680848,0.000004669102,0.01794267,0.001747568],"genre_scores_gemma":[0.496047,1.446141e-7,0.5028376,0.001018798,0.000004649957,0.00006753212,0.00001448043,0.000004463706,0.000005377579],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.8935295,"threshold_uncertainty_score":0.6221499,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2145733341","doi":"10.1145/2786805.2786858","title":"Assertions are strongly correlated with test suite effectiveness","year":2015,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":101,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"Natural Sciences and Engineering Research Council of Canada","keywords":"Test suite; Assertion; Test (biology); Computer science; Suite; Java; Test case; Code coverage; Code (set theory); Unit testing; Programming language; Algorithm; Machine learning; Software; Regression analysis","retraction":null,"screen_n_in":null,"score":{"opus":0.02982942736725199,"gpt":0.2597681029273936,"spread":0.2299386755601416,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003682772,0.0001142475,0.0001228927,0.00007546842,0.0000772631,0.0001287764,0.0004254556,0.00004759713,0.000002123939],"category_scores_gemma":[0.0005842808,0.00008292763,0.00002066541,0.0004637451,0.00003416269,0.0002840222,0.0001113083,0.0001182816,0.0000463842],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005493632,"about_ca_system_score_gemma":0.00009886685,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001538013,"about_ca_topic_score_gemma":0.0000199481,"domain_scores_codex":[0.9991595,0.00006187771,0.00009397655,0.0002705797,0.0002054326,0.0002085899],"domain_scores_gemma":[0.9984511,0.0006471912,0.00005980289,0.000463305,0.0002463234,0.0001322884],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00001915286,0.0002742446,0.9667841,0.00001801161,0.00002291397,0.00008241671,0.0002379866,0.0007005297,0.0000801573,0.006631892,0.01993169,0.005216961],"study_design_scores_gemma":[0.002093531,0.002233465,0.7202278,0.000524271,0.00003593955,0.0003983375,0.00005341108,0.2298511,0.00452725,0.03787209,0.001116012,0.001066779],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.04869111,0.00002165154,0.9318488,0.0001972951,0.0001050733,0.0001632435,0.000001100768,0.01599734,0.002974396],"genre_scores_gemma":[0.8883801,1.845376e-7,0.1112871,0.00005986833,0.00001300402,0.00002637597,0.000001672044,0.000008917734,0.0002228036],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.839689,"threshold_uncertainty_score":0.338169,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1959797194","doi":"10.1007/978-3-642-31057-7_30","title":"Application-Only Call Graph Construction","year":2012,"lang":"en","type":"book-chapter","venue":"Lecture notes in computer science","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":100,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"","keywords":"Call graph; Computer science; Graph; Theoretical computer science; Programming language","retraction":null,"screen_n_in":null,"score":{"opus":0.0139956444234305,"gpt":0.2447250528338945,"spread":0.230729408410464,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0007889556,0.0004402425,0.0003874561,0.0008011725,0.0002639613,0.0003333879,0.002620505,0.0003384859,0.000007042305],"category_scores_gemma":[0.0001203727,0.000418575,0.0001128833,0.0006548959,0.0008043799,0.000515126,0.0008561618,0.0006855351,0.00005968016],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002129603,"about_ca_system_score_gemma":0.0004217963,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005576017,"about_ca_topic_score_gemma":0.00001500402,"domain_scores_codex":[0.9969831,0.00002550986,0.0004331319,0.001208587,0.0007501396,0.0005995514],"domain_scores_gemma":[0.9970881,0.0005371711,0.0003104748,0.001617839,0.0002565235,0.0001898319],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000001036344,0.0000135882,0.0009539499,0.00001655896,0.000004614132,0.000007038409,0.0001515949,0.0004898781,0.00002244386,0.03367247,0.0001311977,0.9645357],"study_design_scores_gemma":[0.0001114221,0.00007607608,0.0004831036,0.0002234213,0.00001029964,0.0003973564,1.492318e-8,0.07660013,0.0004480644,0.9179512,0.003000119,0.000698761],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.00001316416,0.0006003428,0.9916472,0.0003577781,0.001124258,0.0003278751,0.000002921241,0.003805654,0.002120826],"genre_scores_gemma":[0.07829445,0.00003861966,0.9201715,0.0008731374,0.0004774899,0.000018501,0.000005464871,0.00003069966,0.00009014781],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.9638368,"threshold_uncertainty_score":0.9998266,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4394015350","doi":"10.1016/j.infsof.2024.107468","title":"Effective test generation using pre-trained Large Language Models and mutation testing","year":2024,"lang":"en","type":"article","venue":"Information and Software Technology","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":97,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Polytechnique Montréal","funders":"Fonds de recherche du Québec; Natural Sciences and Engineering Research Council of Canada; Canada Research Chairs; Canadian Institute for Advanced Research","keywords":"Test (biology); Mutation; Computer science; Natural language processing; Genetics; Biology; Gene","retraction":null,"screen_n_in":null,"score":{"opus":0.01684748034253352,"gpt":0.2775876674362418,"spread":0.2607401870937083,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003156947,0.000124388,0.0001131727,0.0004882875,0.0001850509,0.0003160165,0.0001331004,0.0001542559,7.683826e-7],"category_scores_gemma":[0.001276847,0.0001164734,0.00001425631,0.0005956185,0.00005256283,0.00173682,0.0001542251,0.0001613955,0.000003502106],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004281879,"about_ca_system_score_gemma":0.000044141,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003839129,"about_ca_topic_score_gemma":0.000003532882,"domain_scores_codex":[0.9992799,0.00001894615,0.0002233511,0.0001913958,0.0001083097,0.00017811],"domain_scores_gemma":[0.9992214,0.0003773888,0.0000708986,0.0001859578,0.0001093665,0.00003500283],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001362817,0.00001321039,0.002531554,0.0001250111,0.00001047724,0.00001567116,0.004026331,0.0002925196,0.0006906574,0.01359391,0.0002477223,0.9784516],"study_design_scores_gemma":[0.0001460956,0.0001014752,0.0007259908,0.00009629407,0.00000750205,0.0003075525,0.00003666068,0.9643289,0.0007541211,0.03326168,0.00009088668,0.0001428841],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.1505017,0.0004333632,0.835888,0.0001577461,0.00007520317,0.0002267382,0.000007201062,0.01266615,0.00004392647],"genre_scores_gemma":[0.7107661,0.000006381202,0.289024,0.0001269255,0.0000200501,0.0000343907,0.00001165395,0.000005605948,0.0000049166],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9783087,"threshold_uncertainty_score":0.4749646,"prediction_status":"machine_predicted_unvalidated"},"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,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"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","retraction":null,"screen_n_in":null,"score":{"opus":0.08433280682020698,"gpt":0.3118294633298048,"spread":0.2274966565095978,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.001528536,0.0002350939,0.0002237536,0.0001543052,0.0006156221,0.0003189929,0.0001612534,0.0001696946,4.896787e-7],"category_scores_gemma":[0.01179017,0.0002070198,0.00001831861,0.0008144915,0.00008525265,0.000410123,0.0000805169,0.0002431698,0.000001325107],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001762393,"about_ca_system_score_gemma":0.0002634479,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0007774481,"about_ca_topic_score_gemma":0.00001308529,"domain_scores_codex":[0.997766,0.0004723068,0.000436418,0.0008233655,0.0003086937,0.0001932843],"domain_scores_gemma":[0.9965063,0.001476215,0.0002705831,0.000704745,0.000768513,0.0002736364],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00003835779,0.0009103609,0.9337693,0.0001060897,0.000005535681,0.001210001,0.002284081,0.001005991,0.00001489199,0.00008063306,0.0008807874,0.05969395],"study_design_scores_gemma":[0.004431493,0.005818601,0.04676836,0.000654847,0.0001035795,0.03263719,0.001711693,0.9041228,0.0002554952,0.002105565,0.0001787512,0.001211632],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7037818,0.00002036258,0.2873782,0.00009623001,0.0001935453,0.0007524516,0.000006784828,0.007744317,0.00002628226],"genre_scores_gemma":[0.8745881,4.5007e-7,0.1251789,0.00005092412,0.00008381901,0.00006288978,0.000009478461,0.000016569,0.000008839556],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9031168,"threshold_uncertainty_score":0.9965339,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2099441126","doi":"10.1109/tse.2003.1214327","title":"General test result checking with log file analysis","year":2003,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":95,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University","funders":"","keywords":"Computer science; Random testing; Formalism (music); Programming language; Unit testing; Operating system; Software; Test case","retraction":null,"screen_n_in":null,"score":{"opus":0.01114718760434826,"gpt":0.215001202784079,"spread":0.2038540151797307,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001922093,0.0002528651,0.0002522806,0.0005018831,0.0001844781,0.0001355094,0.0003802889,0.00009058866,0.00007920685],"category_scores_gemma":[0.0002134777,0.0002372688,0.0001436497,0.001882013,0.00002207403,0.0002567336,0.00000238204,0.0003181418,0.00001739912],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008885792,"about_ca_system_score_gemma":0.00005345914,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004145673,"about_ca_topic_score_gemma":0.000007751677,"domain_scores_codex":[0.9986122,0.00002543399,0.0002293232,0.0004713261,0.0002861109,0.0003756013],"domain_scores_gemma":[0.998276,0.0007479516,0.00005811886,0.0007035306,0.00008917283,0.0001252535],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000006637182,0.0002065263,0.00191652,0.00003514298,0.0003575011,0.000076666,0.0003438007,0.9820529,0.0002106727,0.0002887506,0.002587913,0.01191697],"study_design_scores_gemma":[0.001179208,0.0007925229,0.006258056,0.0004341477,0.0005773937,0.0003197308,0.00001048547,0.9119093,0.06935214,0.0004664742,0.006286966,0.002413544],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.002097336,0.00002534465,0.9867371,0.00001593493,0.0001798991,0.00009430626,0.00003921776,0.01072938,0.00008148545],"genre_scores_gemma":[0.4250256,0.000002955168,0.5746192,0.00005397941,0.00001700989,0.00005709665,0.000003760957,0.00002411502,0.0001962879],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.4229282,"threshold_uncertainty_score":0.9675543,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2952903800","doi":"10.1109/icst.2019.00019","title":"BugsJS: a Benchmark of JavaScript Bugs","year":2019,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":94,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"JavaScript; Unobtrusive JavaScript; Computer science; Benchmark (surveying); Programming language; Test case; Software bug; Unit testing; Rich Internet application; Operating system; Database; Software; Machine learning","retraction":null,"screen_n_in":null,"score":{"opus":0.004992930306730839,"gpt":0.2021300423044523,"spread":0.1971371119977214,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002333794,0.00007520448,0.0001282834,0.00008144735,0.00001817591,0.00003204594,0.0006472999,0.00003796847,0.00008091523],"category_scores_gemma":[0.0000950806,0.00006206841,0.00004992602,0.0002561047,0.00001950653,0.000178754,0.0001981764,0.00005767579,0.00009907159],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001055945,"about_ca_system_score_gemma":0.00003579499,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001530691,"about_ca_topic_score_gemma":9.568513e-7,"domain_scores_codex":[0.9992791,0.00001986787,0.0001544409,0.0002158117,0.0001820731,0.0001487426],"domain_scores_gemma":[0.9990231,0.000150157,0.00005727322,0.0006584922,0.00007230979,0.00003865157],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000009027037,0.0002456314,0.1716663,0.0001394588,0.00004434886,0.00001276474,0.00124866,0.00002521184,0.00544961,0.3021294,0.3956679,0.1233618],"study_design_scores_gemma":[0.001326701,0.001481453,0.05134641,0.00048681,0.00002030144,0.0001378466,0.00003691667,0.1305384,0.0793455,0.694615,0.03922489,0.001439716],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.07177263,0.0001738506,0.8034472,0.00052005,0.0006268471,0.0002478593,9.780554e-7,0.01359485,0.1096157],"genre_scores_gemma":[0.675952,0.000001874927,0.3225239,0.0002777603,0.00001057279,0.000003368389,4.246464e-7,0.000003838832,0.001226235],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6041794,"threshold_uncertainty_score":0.2531076,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2535681473","doi":"10.1109/mutation.2006.10","title":"Mutation Operators for Concurrent Java (J2SE 5.0)","year":2006,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":94,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Computer science; Concurrency; Java; Programming language; Java concurrency; Operating system; Semaphore; Real time Java","retraction":null,"screen_n_in":null,"score":{"opus":0.0165711930763524,"gpt":0.2721297957116077,"spread":0.2555586026352553,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0001495649,0.00006711866,0.00006521577,0.00004925234,0.00007490453,0.0001087115,0.0002476667,0.00002662433,0.000004130716],"category_scores_gemma":[0.00008297707,0.00005651817,0.0000318969,0.0001277071,0.00001409885,0.000157555,0.00004365598,0.00002789916,0.00001389808],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001990918,"about_ca_system_score_gemma":0.00002871727,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001329986,"about_ca_topic_score_gemma":0.000006544504,"domain_scores_codex":[0.9994544,0.00001235146,0.0001228891,0.0001875268,0.00008982212,0.0001330482],"domain_scores_gemma":[0.9995083,0.0001571358,0.00003116949,0.0001908444,0.00008713292,0.00002544901],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001641451,0.0001000159,0.002961689,0.0000196092,0.000004932604,0.000006942371,0.0001349215,0.0001217124,0.0002475627,0.3477556,0.4461718,0.2024735],"study_design_scores_gemma":[0.0009688252,0.000413792,0.004825321,0.0000583685,0.00001130324,0.00005672081,0.000005914409,0.5467756,0.03799754,0.3704988,0.03770189,0.000685969],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.00474806,0.00004147199,0.9854017,0.000312955,0.0002488926,0.0001572422,0.000001133531,0.00762329,0.001465245],"genre_scores_gemma":[0.6316637,3.391351e-7,0.3676842,0.0002024118,0.00005369726,0.00004021733,0.000004860961,0.000004381169,0.0003461534],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6269157,"threshold_uncertainty_score":0.2304744,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2156296121","doi":"10.1109/issre.2007.31","title":"Using Machine Learning to Support Debugging with Tarantula","year":2007,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":93,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Debugging; Computer science; Statement (logic); Ranking (information retrieval); Test (biology); Software bug; Decision tree; Fault (geology); Fault tree analysis; Machine learning; Algorithmic program debugging; Test case; Tree (set theory); Artificial intelligence; Programming language; Reliability engineering; Software; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.04028565877204957,"gpt":0.3031118466118101,"spread":0.2628261878397605,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008205763,0.0001398957,0.0001345976,0.0002147207,0.0001872077,0.0001215013,0.0004515474,0.000036299,0.00001716903],"category_scores_gemma":[0.0001293159,0.0001100003,0.00002736587,0.0005658355,0.00001922978,0.0002177709,0.0002015568,0.0001621035,0.00002427046],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0000462293,"about_ca_system_score_gemma":0.00005342454,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005358825,"about_ca_topic_score_gemma":0.00006371414,"domain_scores_codex":[0.9987996,0.00002186699,0.0001747512,0.0003370428,0.000260125,0.0004066349],"domain_scores_gemma":[0.9991971,0.0001745251,0.00005208602,0.0003510232,0.00007426747,0.0001509867],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00009153032,0.0001828445,0.5309679,0.00005408053,0.00006803007,0.001131016,0.004210137,0.005116822,0.007486703,0.01093811,0.005978835,0.433774],"study_design_scores_gemma":[0.0009138199,0.002010933,0.01045004,0.0003160243,0.00003458273,0.001899984,0.00004107774,0.9130669,0.04639611,0.009263732,0.01381276,0.001793988],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.0389224,0.00001624022,0.9450706,0.0001299992,0.0000578415,0.00009301672,1.576248e-7,0.01298946,0.002720283],"genre_scores_gemma":[0.4775219,2.193925e-7,0.521827,0.0004597327,0.00002048648,0.000001175454,5.788862e-7,0.000009648068,0.0001591679],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.9079501,"threshold_uncertainty_score":0.4485683,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2160197556","doi":"10.1109/tc.2006.80","title":"Optimizing the length of checking sequences","year":2006,"lang":"en","type":"article","venue":"IEEE Transactions on Computers","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":93,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"Engineering and Physical Sciences Research Council; Natural Sciences and Engineering Research Council of Canada; Leverhulme Trust","keywords":"Sequence (biology); State (computer science); Set (abstract data type); Algorithm; Finite-state machine; Mathematics; Computer science; Combinatorics","retraction":null,"screen_n_in":null,"score":{"opus":0.02136507433847591,"gpt":0.2432458928702451,"spread":0.2218808185317692,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0002332699,0.0001209377,0.0001278822,0.0001235352,0.0002223661,0.00008805269,0.0007217317,0.00004024867,0.000001945252],"category_scores_gemma":[0.00000342837,0.00009124092,0.00009819723,0.0003848692,0.00009115342,0.0001823127,0.000004340002,0.0001654071,0.000004210985],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003251772,"about_ca_system_score_gemma":0.00003605157,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003169932,"about_ca_topic_score_gemma":0.000008777031,"domain_scores_codex":[0.9990833,0.00005721845,0.0002159092,0.0002372578,0.0002218553,0.0001844964],"domain_scores_gemma":[0.9989073,0.0004717497,0.0000916722,0.0004495221,0.00005392726,0.00002581921],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00001072086,0.0002736437,0.0001267535,0.00003093435,0.00006268505,0.00001960825,0.001607505,0.6442675,0.001232777,0.004339761,0.005367792,0.3426603],"study_design_scores_gemma":[0.0003225889,0.00024368,0.0003081401,0.0002583339,0.00002374824,0.00007464711,0.00001722564,0.9018099,0.0851272,0.01099886,0.0004234641,0.0003922301],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.00400988,0.00003941211,0.991817,0.0004519572,0.0005847838,0.00009260842,0.000001339382,0.00269092,0.000312097],"genre_scores_gemma":[0.7150267,0.000003952051,0.2847749,0.0001342676,0.00002735963,0.000008387994,1.636106e-7,0.000005925484,0.00001839591],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.7110168,"threshold_uncertainty_score":0.3720697,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1920048997","doi":"10.1109/ase.1998.732614","title":"Testing using log file analysis: tools, methods, and issues","year":2002,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":93,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Western University","funders":"Natural Sciences and Engineering Research Council of Canada; University of British Columbia","keywords":"Computer science; Scope (computer science); Software; Software testing; Programming language; Database; Software engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.1628634645043925,"gpt":0.373331725608945,"spread":0.2104682611045524,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004371672,0.0001149443,0.0001958263,0.0002000114,0.0001452328,0.0003162381,0.0003452349,0.00004620666,0.0001960239],"category_scores_gemma":[0.001514282,0.00009910834,0.00004471734,0.001238023,0.00003276108,0.0003538232,0.0002488905,0.00008079648,0.000008951569],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001414803,"about_ca_system_score_gemma":0.000006979721,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003782385,"about_ca_topic_score_gemma":0.000002379764,"domain_scores_codex":[0.9990399,0.0001001694,0.0001745131,0.0003537142,0.0001212034,0.0002105349],"domain_scores_gemma":[0.9981249,0.001212386,0.0000672125,0.000453807,0.00007521314,0.00006646885],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[4.240184e-7,0.00005706492,0.01875751,0.00001748709,0.0001308421,0.00002693788,0.000551754,0.0001840478,0.0007506427,0.00180724,0.07736847,0.9003476],"study_design_scores_gemma":[0.00003521761,0.00003101115,0.003934338,0.00002281113,0.00004942971,0.00003505333,0.000004646913,0.9840805,0.0005648592,0.009908004,0.001151504,0.0001826242],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.001308302,0.0007130514,0.9870311,0.0001874937,0.00003345377,0.00004921899,0.000002780481,0.00781964,0.002854944],"genre_scores_gemma":[0.03167396,0.000005051592,0.9677196,0.0002437295,0.00003078339,0.000004709048,0.000001312917,0.000005961815,0.0003149186],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.9838964,"threshold_uncertainty_score":0.4041521,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1880492488","doi":"10.1120/jacmp.v14i1.4052","title":"Automated IMRT planning with regional optimization using planning scripts","year":2013,"lang":"en","type":"article","venue":"Journal of Applied Clinical Medical Physics","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":92,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"London Health Sciences Centre; Western University","funders":"Cancer Care Ontario","keywords":"Scripting language; Radiation treatment planning; Computer science; Medical physics; Medicine; Radiation therapy; Radiology; Programming language","retraction":null,"screen_n_in":null,"score":{"opus":0.09085317414621082,"gpt":0.3731214052892368,"spread":0.282268231143026,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001170006,0.0001755597,0.0004499117,0.00007778707,0.0001215826,0.0001335182,0.0008166115,0.0001895848,0.00001124982],"category_scores_gemma":[0.0005982689,0.0001258092,0.0001060154,0.0003921838,0.0001847723,0.0004144172,0.0001496864,0.0008070896,0.000006205971],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004439814,"about_ca_system_score_gemma":0.0003870436,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000007779799,"about_ca_topic_score_gemma":2.26324e-8,"domain_scores_codex":[0.9973465,0.00007689585,0.0009353395,0.0002615228,0.00109493,0.0002848757],"domain_scores_gemma":[0.9970124,0.001152035,0.0008140148,0.0002874399,0.000331796,0.0004022533],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0004176142,0.001792908,0.03351011,0.0001290635,0.0005794613,0.0007785934,0.001744131,0.5567521,0.0001524004,0.008316708,0.1989261,0.1969007],"study_design_scores_gemma":[0.0008139862,0.000323058,0.001629194,0.0006334143,0.00002562814,0.0001693426,0.00001394935,0.9699625,0.00004405787,0.02609339,0.00008985399,0.0002016516],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.03149569,0.00003821914,0.9655402,0.0005005991,0.0002711295,0.0001070516,1.524527e-7,0.001722728,0.0003241832],"genre_scores_gemma":[0.4371929,0.000004603128,0.5606228,0.001538773,0.0006182088,0.000003284814,0.00000113777,0.00001657638,0.000001723121],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.4132103,"threshold_uncertainty_score":0.5130351,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2056488032","doi":"10.1145/1391984.1391985","title":"Unit-level test adequacy criteria for visual dataflow languages and a testing methodology","year":2008,"lang":"en","type":"article","venue":"ACM Transactions on Software Engineering and Methodology","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":92,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Dalhousie University","funders":"","keywords":"Computer science; Dataflow; Context (archaeology); Unit testing; Visual programming language; Empirical research; Software engineering; Programming language; Software","retraction":null,"screen_n_in":null,"score":{"opus":0.3232977272659157,"gpt":0.4163952044641103,"spread":0.09309747719819456,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.001389497,0.0002790582,0.00043149,0.0003456153,0.0003306694,0.00004497692,0.0004408362,0.0002002918,0.00000319793],"category_scores_gemma":[0.02163374,0.0002790802,0.00005670954,0.0003642591,0.0001129384,0.0002116091,0.00005969472,0.0003249873,0.000001302852],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001775438,"about_ca_system_score_gemma":0.00005596422,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001040593,"about_ca_topic_score_gemma":0.000003202286,"domain_scores_codex":[0.9983016,0.0002996254,0.0002859674,0.0005947637,0.0001002221,0.0004178481],"domain_scores_gemma":[0.9676278,0.03147611,0.00007014266,0.0005655641,0.0001147325,0.0001456635],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00006858348,0.0002459822,0.004847866,0.0003812986,0.0001450411,0.0001097678,0.002743813,0.001295612,0.01026787,0.0008979223,0.000571897,0.9784244],"study_design_scores_gemma":[0.01043469,0.01488964,0.1200833,0.001556643,0.000845295,0.02888515,0.0004930165,0.607833,0.1213767,0.07241214,0.01319636,0.007994033],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.01587246,0.0003192727,0.9784707,0.0001978415,0.0002826361,0.0001928241,0.0000373943,0.004624269,0.000002568135],"genre_scores_gemma":[0.07256598,0.00006208932,0.9268639,0.0002353064,0.0000650708,0.0001087564,0.000007245993,0.0000357654,0.00005589013],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.9704303,"threshold_uncertainty_score":0.9999661,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4378942602","doi":"10.1145/3597926.3598135","title":"How Effective Are Neural Networks for Fixing Security Vulnerabilities","year":2023,"lang":"en","type":"preprint","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":90,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta; York University","funders":"National Science Foundation","keywords":"Computer science; Task (project management); Vulnerability (computing); Code (set theory); Software security assurance; Secure coding; Computer security; Source code; Software; Artificial neural network; Software bug; Automation; Programming language; Software engineering; Artificial intelligence; Information security; Engineering; Security service","retraction":null,"screen_n_in":null,"score":{"opus":0.03806044719484671,"gpt":0.2861846731068167,"spread":0.24812422591197,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","scholarly_communication"],"consensus_categories":[],"category_scores_codex":[0.0008726011,0.0003804102,0.0004805297,0.0001597974,0.0002324434,0.00105574,0.001366067,0.000350977,8.193198e-7],"category_scores_gemma":[0.001835534,0.0003427662,0.0002822644,0.0002115168,0.00006919151,0.0002087888,0.002401987,0.0007886444,0.000001501152],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001089059,"about_ca_system_score_gemma":0.00004347981,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002831731,"about_ca_topic_score_gemma":0.00003578175,"domain_scores_codex":[0.99783,0.0001847783,0.0002308338,0.00102829,0.0002370749,0.0004890027],"domain_scores_gemma":[0.9945963,0.003608647,0.0002430858,0.001194374,0.0002655447,0.00009203493],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00007131168,0.0002843323,0.04231357,0.00434546,0.0004618696,0.0001486389,0.006339779,0.1051589,0.000005788671,0.02755996,0.6546026,0.1587078],"study_design_scores_gemma":[0.00007181149,0.00005991511,0.0007383786,0.0002034637,0.000008915395,0.000004326102,0.00001342485,0.6826349,0.00006673793,0.3157921,0.00009767858,0.000308345],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.002840188,0.0002005022,0.9511022,0.002611427,0.001986894,0.001179431,0.00001729459,0.03998609,0.0000759638],"genre_scores_gemma":[0.8588312,0.000007329737,0.1383414,0.0002650172,0.0005549084,0.001341586,0.00002313561,0.0000493311,0.0005861693],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8559909,"threshold_uncertainty_score":0.9999813,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2067617772","doi":"10.1109/tse.2014.2363479","title":"Instance Generator and Problem Representation to Improve Object Oriented Code Coverage","year":2014,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":86,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Polytechnique Montréal","funders":"","keywords":"Computer science; Unit testing; Test case; Programming language; Code generation; Java; Code coverage; Object-oriented programming; Data structure; Source code; Generator (circuit theory); Software; Theoretical computer science; Machine learning; Operating system","retraction":null,"screen_n_in":null,"score":{"opus":0.009565537506992744,"gpt":0.2305431465418212,"spread":0.2209776090348285,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.000193143,0.0001883476,0.0001693547,0.0002036883,0.0001434902,0.0001174171,0.0002299115,0.00006788442,0.000001826329],"category_scores_gemma":[0.0001341209,0.0002002719,0.00004380495,0.0004937935,0.00001232054,0.0002990924,0.000006658844,0.0001918172,0.000009034404],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007021408,"about_ca_system_score_gemma":0.00002471203,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003066585,"about_ca_topic_score_gemma":0.000005638365,"domain_scores_codex":[0.9988242,0.00002890763,0.0002084279,0.00047642,0.0002088967,0.0002531047],"domain_scores_gemma":[0.9989578,0.0002948789,0.00004215111,0.0004843687,0.00007982164,0.0001410306],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00005124285,0.0002213241,0.001116819,0.0002302875,0.00009319634,0.00002463707,0.002454906,0.461641,0.0184052,0.002393533,0.002146558,0.5112213],"study_design_scores_gemma":[0.001247289,0.0008910585,0.001470014,0.0004589043,0.00003635258,0.00008067838,0.000006343374,0.7613114,0.2257856,0.001511534,0.005852932,0.001347887],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.02414405,0.00001725788,0.96736,0.00005623537,0.0005070698,0.0002449994,0.00001103311,0.007642321,0.00001706783],"genre_scores_gemma":[0.5338054,0.000005414764,0.4658477,0.0001481981,0.00003208351,0.00009117211,7.806121e-7,0.00002046834,0.00004871766],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.5098734,"threshold_uncertainty_score":0.8166853,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2121504314","doi":"10.1145/2642937.2642991","title":"Leveraging existing tests in automated test generation for web applications","year":2014,"lang":"en","type":"article","venue":"","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":82,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"Natural Sciences and Engineering Research Council of Canada; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung","keywords":"Test suite; Computer science; Web crawler; Crawling; Code coverage; Test (biology); Web testing; Test harness; Random testing; Test Management Approach; Test case; Domain (mathematical analysis); Generator (circuit theory); Software engineering; Event (particle physics); Data mining; Machine learning; Web page; Programming language; World Wide Web; Software; Power (physics); Web development; Software development; Web application security","retraction":null,"screen_n_in":null,"score":{"opus":0.0607930062329493,"gpt":0.3199788099738178,"spread":0.2591858037408685,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004804717,0.00008124665,0.0000855808,0.0001182698,0.0001264737,0.0001258574,0.0003149301,0.00003702282,8.884593e-7],"category_scores_gemma":[0.0008937853,0.00007904707,0.00001938739,0.0003257483,0.00001133007,0.0001473851,0.00006693566,0.00004871958,0.000008549421],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003549852,"about_ca_system_score_gemma":0.0000366145,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003996296,"about_ca_topic_score_gemma":0.00001538103,"domain_scores_codex":[0.9992519,0.00002180227,0.0001817215,0.0002862342,0.00008277517,0.0001755705],"domain_scores_gemma":[0.998632,0.0008548592,0.00005564177,0.0003434405,0.00008012092,0.00003390316],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000001638647,0.000380416,0.1007584,0.0001121381,0.000008116989,0.000002506377,0.000785768,0.001580962,0.06699625,0.0931227,0.1000134,0.6362377],"study_design_scores_gemma":[0.0001201052,0.00002736064,0.002896338,0.00002101139,9.807825e-7,0.000004844397,7.753231e-7,0.9812879,0.001033332,0.01287226,0.001625161,0.0001099231],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.004105115,0.00001081198,0.9724594,0.0003036219,0.00004147868,0.0002593396,5.439671e-7,0.02098165,0.001838094],"genre_scores_gemma":[0.5279391,2.6175e-7,0.4715877,0.0001922497,0.00005423291,0.0001731809,0.0000030877,0.000005098973,0.00004511232],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.9797069,"threshold_uncertainty_score":0.3223446,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2142266413","doi":"10.5555/998675.999415","title":"Using simulation to empirically investigate test coverage criteria based on statechart","year":2004,"lang":"en","type":"article","venue":"International Conference on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":81,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Computer science; Code coverage; Class (philosophy); Test (biology); Test case; Reliability engineering; Fault detection and isolation; Test strategy; Data mining; Machine learning; Artificial intelligence; Programming language; Software; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.08716524489143154,"gpt":0.3492094312045713,"spread":0.2620441863131397,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0001952904,0.000237063,0.0001586258,0.0003922532,0.00007476439,0.0003230111,0.0007081759,0.00007306846,0.0000305462],"category_scores_gemma":[0.002710468,0.0002507923,0.00005241195,0.0003105636,0.00001552453,0.0002932125,0.0001228254,0.000212447,0.00005060355],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003003419,"about_ca_system_score_gemma":0.0001488598,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003593791,"about_ca_topic_score_gemma":9.848256e-7,"domain_scores_codex":[0.998498,0.00001593435,0.0002684999,0.0004521153,0.0005041563,0.0002613329],"domain_scores_gemma":[0.9986233,0.0004645599,0.00007383468,0.0004017522,0.0002714471,0.0001651107],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.000009812877,0.00006356643,0.0008558196,0.000009130583,0.00000641643,0.00003288397,0.0001269486,0.990972,0.001189786,0.004713312,0.0001587586,0.001861599],"study_design_scores_gemma":[0.0002758276,0.0002295972,0.001616632,0.0004712222,0.000002245986,0.000005877521,6.82479e-7,0.9869934,0.002383809,0.007406517,0.0003099321,0.0003042188],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.02195822,0.000001595937,0.9717268,0.0007571952,0.0004812902,0.0001536761,0.00002486456,0.004717476,0.0001788438],"genre_scores_gemma":[0.6258753,4.448898e-7,0.3728722,0.001148534,0.00005239953,0.00001502163,0.000008944588,0.000017456,0.000009715572],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.6039171,"threshold_uncertainty_score":0.9999945,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2100209048","doi":"","title":"Static And Dynamic Reverse Engineering Techniques for Java Software Systems","year":2019,"lang":"en","type":"book","venue":"Trepo - Institutional Repository of Tampere University","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":81,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"Tekes; University of Victoria","keywords":"Reverse engineering; Computer science; Java; Software engineering; Programming language","retraction":null,"screen_n_in":null,"score":{"opus":0.008589263300448567,"gpt":0.1962618870493534,"spread":0.1876726237489048,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0002142502,0.000301203,0.0004643475,0.0003539673,0.0003000047,0.00006563491,0.0006576127,0.0003084898,7.473998e-7],"category_scores_gemma":[0.0002314489,0.0003401741,0.000159382,0.0001007101,0.0002081789,0.0003962432,0.0003075963,0.0002501694,0.000001957625],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0007151056,"about_ca_system_score_gemma":0.001066081,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001214586,"about_ca_topic_score_gemma":0.000001506489,"domain_scores_codex":[0.998528,0.0000335165,0.0003205405,0.0005755196,0.0003377743,0.0002046236],"domain_scores_gemma":[0.9979793,0.0005290241,0.0003980943,0.0006200466,0.0003778786,0.00009561764],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0004564557,0.0006921808,0.002518645,0.02492475,0.001967033,0.003214582,0.002299593,0.01098852,0.002288051,0.4786044,0.4468595,0.02518624],"study_design_scores_gemma":[0.001293922,0.0009931369,0.0005555958,0.009316411,0.0004959899,0.002233608,0.00003922558,0.1052955,0.0006778148,0.007816388,0.8687854,0.002497026],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.0007538135,0.0005933024,0.9764649,0.00006547235,0.001171719,0.001238386,0.000200113,0.006038635,0.01347372],"genre_scores_gemma":[0.01812653,0.0001098878,0.6713408,0.00005716032,0.0001557185,0.00001499358,0.0001610127,0.00005991006,0.309974],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.470788,"threshold_uncertainty_score":0.999905,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2068493323","doi":"10.1109/tse.2011.56","title":"Size-Constrained Regression Test Case Selection Using Multicriteria Optimization","year":2011,"lang":"en","type":"article","venue":"IEEE Transactions on Software Engineering","topic":"Software Testing and Debugging Techniques","field":"Computer Science","cited_by":79,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Waterloo","funders":"","keywords":"Computer science; Test suite; Test case; Consistency (knowledge bases); Integer programming; Mathematical optimization; Task (project management); Constraint (computer-aided design); Selection (genetic algorithm); Regression testing; Greedy algorithm; Linear programming; Relaxation (psychology); Software; Algorithm; Machine learning; Regression analysis; Artificial intelligence; Software system; Mathematics","retraction":null,"screen_n_in":null,"score":{"opus":0.03160363037280706,"gpt":0.2469561588901617,"spread":0.2153525285173546,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.0001789559,0.0002636428,0.0001839905,0.0002779798,0.0002802958,0.00009091815,0.0002445585,0.0001406984,0.00004291169],"category_scores_gemma":[0.0002923624,0.0002701113,0.00008852467,0.0006140983,0.00002576569,0.0005289317,0.000004691306,0.0002773664,0.000005808294],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001245864,"about_ca_system_score_gemma":0.00004692365,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001632204,"about_ca_topic_score_gemma":0.00000339062,"domain_scores_codex":[0.9987847,0.0000319489,0.000284816,0.0004131713,0.0001719506,0.0003133652],"domain_scores_gemma":[0.9987538,0.0005376507,0.00007838665,0.0003839816,0.000123267,0.0001228913],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00002121871,0.0003879135,0.0003744851,0.0000918373,0.00004832125,0.0006021768,0.001580242,0.9626136,0.004410166,0.00004731314,0.00009243377,0.02973033],"study_design_scores_gemma":[0.0002387569,0.0001250758,0.00004332029,0.0001878856,0.00002243677,0.002246485,0.000005853054,0.9663855,0.03037051,0.00004205878,0.000008099157,0.0003240613],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.008772391,0.00001301761,0.9777734,0.000006441926,0.0006293928,0.00017568,0.000007164491,0.01261108,0.00001143322],"genre_scores_gemma":[0.4769311,0.000001977194,0.5229734,0.00001837047,0.00002083111,0.00001858557,3.684222e-7,0.00002339904,0.00001198384],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.4681587,"threshold_uncertainty_score":0.9999751,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}