{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":200,"total_is_capped":false,"direct_labels_cover":7,"predictions_cover":200,"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":"8a4c8283c409","filters":{"venue":"Evaluation and Program Planning"}},"results":[{"id":"W2017443876","doi":"10.1016/j.evalprogplan.2004.01.001","title":"Making public health programs last: conceptualizing sustainability","year":2004,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Community Health and Development","field":"Health Professions","cited_by":379,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal; McGill University","funders":"Centers for Disease Control and Prevention","keywords":"Sustainability; Process management; Process (computing); Sustainability organizations; Social sustainability; Business; Management science; Knowledge management; Public relations; Environmental resource management; Engineering; Computer science; Political science; Economics; Ecology","retraction":null,"screen_n_in":null,"score":{"opus":0.4589626217430083,"gpt":0.59910954355279,"spread":0.1401469218097817,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.008545569,0.0001771014,0.000274154,0.0001481611,0.002368016,0.00007666917,0.0001266529,0.0001391231,0.0001104959],"category_scores_gemma":[0.00118909,0.0001669815,0.00003436105,0.0004242321,0.0001000423,0.0002196399,0.0001523009,0.0007200803,0.00002441451],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001190863,"about_ca_system_score_gemma":0.004654903,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002370919,"about_ca_topic_score_gemma":0.0003311781,"domain_scores_codex":[0.9959571,0.001516303,0.0006967854,0.0003165721,0.0005529685,0.0009602993],"domain_scores_gemma":[0.9980696,0.0002754112,0.0003338766,0.0002716108,0.0006714704,0.0003780284],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00002331932,0.000145126,0.116803,0.000472153,0.0000125296,0.000001560877,0.06588683,0.00004569526,4.096683e-7,0.005064017,0.0002091779,0.8113362],"study_design_scores_gemma":[0.006021187,0.001032288,0.2701847,0.00192715,0.00002415756,0.00001374023,0.1930638,0.002476081,0.000001025086,0.01919159,0.5054787,0.0005856446],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9219006,0.005265219,0.004658969,0.04312784,0.0007786824,0.0135834,0.000004210036,0.001161988,0.009519127],"genre_scores_gemma":[0.9840126,0.00003172406,0.01024163,0.003883028,0.0001179963,0.001472413,0.0001667781,0.0000198232,0.00005397789],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8107506,"threshold_uncertainty_score":0.9989308,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2033773300","doi":"10.1016/j.evalprogplan.2009.07.005","title":"Challenges and opportunities for evaluating environmental education programs","year":2009,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Environmental Education and Sustainability","field":"Environmental Science","cited_by":245,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Trillium Therapeutics (Canada)","funders":"","keywords":"Environmental education; Environmental quality; Quality (philosophy); Program evaluation; Environmental impact assessment; Bridge (graph theory); Environmental planning; Medical education; Environmental resource management; Engineering; Political science; Psychology; Medicine; Pedagogy; Environmental science; Public administration","retraction":null,"screen_n_in":null,"score":{"opus":0.3024740036663272,"gpt":0.474902373465715,"spread":0.1724283697993879,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.002156493,0.0003741312,0.0005256773,0.00007696925,0.0002788255,0.0001147234,0.0001186273,0.0002129181,0.0002123282],"category_scores_gemma":[0.0001021195,0.0003410526,0.000112327,0.00005140245,0.0002303656,0.0002376187,0.00009941857,0.000187085,0.00001029466],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003900704,"about_ca_system_score_gemma":0.000140417,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000003935722,"about_ca_topic_score_gemma":0.000002580749,"domain_scores_codex":[0.9976606,0.0004079113,0.0004668107,0.0006870872,0.0004737822,0.000303806],"domain_scores_gemma":[0.9991289,0.000111205,0.0003147011,0.0002440995,0.0000110081,0.0001900341],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000004826943,0.0002615368,0.0000802037,0.001161695,0.00001790368,2.935697e-7,0.001073194,0.000002667265,1.621626e-7,0.00005007331,0.0001212749,0.9972262],"study_design_scores_gemma":[0.000226748,0.0004752597,0.001344367,0.001275029,0.0004707756,0.00003207838,0.003735554,0.001010878,9.242191e-8,0.0008011389,0.9902278,0.000400312],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.001115631,0.9895144,0.000008589264,0.0002058141,0.0001128881,0.00518054,0.000005330916,0.00006711279,0.003789703],"genre_scores_gemma":[0.000696412,0.9911318,0.003227708,0.00009971876,0.00007873218,0.003347762,0.0008051704,0.00003439218,0.0005782625],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9968259,"threshold_uncertainty_score":0.9999042,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2022589600","doi":"10.1016/j.evalprogplan.2009.07.001","title":"Peer-led interventions to reduce HIV risk of youth: A review","year":2009,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Adolescent Sexual and Reproductive Health","field":"Health Professions","cited_by":160,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Windsor","funders":"","keywords":"Psychological intervention; Peer education; Condom; Human immunodeficiency virus (HIV); Peer group; Program Design Language; Program evaluation; Public relations; Positive Youth Development; Medicine; Psychology; Environmental health; Medical education; Health education; Political science; Public health; Family medicine; Nursing; Social psychology; Developmental psychology; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.5060799930559863,"gpt":0.6377364766603129,"spread":0.1316564836043266,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.009628613,0.000355545,0.001642347,0.0002707861,0.0005265835,0.00001181368,0.0001832219,0.0002968045,0.000229056],"category_scores_gemma":[0.003365659,0.0002591821,0.0003555339,0.0006290686,0.00004325904,0.00006659175,0.0001070726,0.001154469,0.0001535928],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001467821,"about_ca_system_score_gemma":0.0006802204,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001821187,"about_ca_topic_score_gemma":0.000002618828,"domain_scores_codex":[0.9927177,0.003787921,0.00170969,0.0006391561,0.0007279701,0.0004175667],"domain_scores_gemma":[0.9964201,0.0002729951,0.001607552,0.0004752243,0.0009825834,0.0002415455],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001143937,0.0001591333,0.000108976,0.1068417,0.00005791626,4.985915e-7,0.00168744,9.176818e-7,2.989485e-8,0.00002657049,0.009098678,0.8820067],"study_design_scores_gemma":[0.0002411074,0.0002217606,0.0003846399,0.3037133,0.001823272,0.000003036688,0.0005802501,0.00001240853,2.664595e-8,0.00003198782,0.6927969,0.0001913145],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00003311842,0.9847192,0.0001474997,0.000895564,0.0005549436,0.0116919,0.0001491482,0.0001188341,0.001689769],"genre_scores_gemma":[0.00002391529,0.9924085,0.00124499,0.0002407779,0.000461496,0.001024293,0.001146342,0.00003561784,0.003414123],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8818154,"threshold_uncertainty_score":0.9999861,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2044849666","doi":"10.1016/j.evalprogplan.2010.04.001","title":"Program theory evaluation: Logic analysis","year":2010,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Logic, Reasoning, and Knowledge","field":"Computer Science","cited_by":155,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal; Hôpital Charles-Le Moyne","funders":"Canadian Institutes of Health Research; Canadian Health Services Research Foundation","keywords":"Logic model; Computer science; Engineering; Political science; Public administration","retraction":null,"screen_n_in":null,"score":{"opus":0.06455360024763711,"gpt":0.4119331162399644,"spread":0.3473795159923273,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007113377,0.0001851681,0.0002171975,0.0002649928,0.0003417217,0.0005791797,0.0004001416,0.0001345547,0.0003513141],"category_scores_gemma":[0.0005900353,0.0001500082,0.0001132277,0.001159207,0.00009136916,0.0003945878,0.0001178268,0.0002823331,0.00004160309],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002746759,"about_ca_system_score_gemma":0.000165662,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004848293,"about_ca_topic_score_gemma":0.00001700508,"domain_scores_codex":[0.9975293,0.0005163793,0.0002696043,0.0005414529,0.0008230472,0.0003202336],"domain_scores_gemma":[0.9984279,0.0001976612,0.0001677042,0.0004713621,0.000585636,0.0001497568],"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":[0.000006855738,0.0001487156,0.01055289,0.000006638317,0.0001307557,0.000002198455,0.001716002,0.0002557546,0.00003783486,0.04046949,0.0002117189,0.9464611],"study_design_scores_gemma":[0.0006550761,0.0002620446,0.06985502,0.00001310116,0.0005499329,0.00001602077,0.00012363,0.8815262,0.00005152764,0.03762551,0.009044387,0.0002775151],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5508474,0.009387887,0.2422535,0.001560019,0.002843057,0.008196295,0.000004397256,0.003364141,0.1815432],"genre_scores_gemma":[0.9381653,0.000007106436,0.06063381,0.0001339202,0.0001625565,0.0006642154,0.00006559443,0.000007469597,0.0001600215],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9461836,"threshold_uncertainty_score":0.6117157,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2062658206","doi":"10.1016/j.evalprogplan.2004.10.003","title":"Program sustainability begins with the first events","year":2004,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Community Health and Development","field":"Health Professions","cited_by":146,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Cégep de Rivière-du-Loup; Université de Montréal; McGill University","funders":"","keywords":"Sustainability; Process (computing); Public relations; Promotion (chess); Health promotion; Sociology; Political science; Process management; Business; Computer science; Health care; Law; Politics","retraction":null,"screen_n_in":null,"score":{"opus":0.1202160389815176,"gpt":0.531857968154241,"spread":0.4116419291727235,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.003008221,0.0001302971,0.0001360304,0.0000543809,0.002365393,0.00002636645,0.000120245,0.00008974572,0.00006851872],"category_scores_gemma":[0.0002996185,0.00008085366,0.00001961526,0.000280779,0.00007344459,0.00009724004,0.00008435187,0.000556778,0.00001832266],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003777761,"about_ca_system_score_gemma":0.001342902,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002813911,"about_ca_topic_score_gemma":0.0008335679,"domain_scores_codex":[0.9982103,0.0004272803,0.0002699789,0.0001942513,0.0004393689,0.0004588329],"domain_scores_gemma":[0.9986919,0.0002851882,0.0001410521,0.0002690511,0.0004673993,0.0001454684],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.000219539,0.0003659578,0.333556,0.0004817965,0.00003150518,0.000002878236,0.04240796,0.0004668653,1.850002e-7,0.0009651872,0.001401645,0.6201004],"study_design_scores_gemma":[0.002644977,0.000620089,0.6574028,0.0003710219,0.00003862091,0.000004233581,0.01749638,0.0009394386,9.136135e-7,0.003648694,0.3166612,0.0001716537],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9420856,0.0005383872,0.0002717155,0.04197725,0.0001538096,0.01024654,0.000002019226,0.0003551244,0.004369525],"genre_scores_gemma":[0.990606,0.00001451687,0.00314719,0.001837884,0.00007517654,0.004115413,0.00005867933,0.00001306225,0.0001321416],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6199288,"threshold_uncertainty_score":0.9989334,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2971813575","doi":"10.1016/j.evalprogplan.2019.101713","title":"Governing and managing higher education institutions: The quality audit contributions","year":2019,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Organizational Strategy and Culture","field":"Business, Management and Accounting","cited_by":82,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Accreditation; Audit; Accountability; Higher education; Corporate governance; Quality management; Quality assurance; Business; Quality (philosophy); Quality audit; Quality policy; Quality management system; Public relations; Accounting; Management system; Political science; Operations management; Medicine; Medical education; Engineering; Marketing; Finance","retraction":null,"screen_n_in":null,"score":{"opus":0.0592925696596097,"gpt":0.3886790330649743,"spread":0.3293864634053645,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0007208264,0.00008202052,0.00007045813,0.00003644284,0.0004978278,0.0004118596,0.00005218035,0.00004677246,0.0004225081],"category_scores_gemma":[0.00009845047,0.00006080595,0.00001445484,0.0002426003,0.00003913198,0.0006287072,0.00003889624,0.0001081476,0.00008694795],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002464987,"about_ca_system_score_gemma":0.00006225063,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006697542,"about_ca_topic_score_gemma":0.000008077868,"domain_scores_codex":[0.9993443,0.0000298887,0.0001374079,0.0001572956,0.0002274898,0.0001035542],"domain_scores_gemma":[0.9995301,0.0000347497,0.0001242838,0.00007634854,0.0002250675,0.000009444127],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","study_design_scores_codex":[0.00001997828,0.00007732729,0.1456591,0.0001129198,0.00004023861,2.846449e-7,0.0002318191,0.0006100406,0.000147759,0.7637854,0.003230484,0.08608466],"study_design_scores_gemma":[0.0004214645,0.000008303069,0.6149256,0.00008738106,0.00007423072,0.000002312357,0.0009047966,0.00369187,0.000003243065,0.01008589,0.3696517,0.0001431417],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8694423,0.00462802,0.0002052895,0.01247664,0.0009421514,0.001551447,0.000006988408,0.0002428306,0.1105044],"genre_scores_gemma":[0.9969951,0.00001151528,0.00008194875,0.001451958,0.0004393295,0.00005949334,0.0002269692,0.000005503524,0.0007281709],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7536994,"threshold_uncertainty_score":0.4626167,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2011406769","doi":"10.1016/j.evalprogplan.2010.09.003","title":"Ten steps to making evaluation matter","year":2010,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":81,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Michael's Hospital; University of Toronto","funders":"","keywords":"Program evaluation; Computer science; Program Design Language; Management science; Theory of change; Work (physics); Evaluation methods; Sustainability; Risk analysis (engineering); Process management; Software engineering; Engineering; Political science; Sociology; Business","retraction":null,"screen_n_in":null,"score":{"opus":0.3049186399144428,"gpt":0.5971294410667453,"spread":0.2922108011523025,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.02347861,0.0001635045,0.0001886465,0.0004023692,0.0003430694,0.0007978574,0.0002860999,0.0001007807,0.01146296],"category_scores_gemma":[0.002025584,0.0001269483,0.00004952808,0.0006641402,0.00004330777,0.0005342768,0.00009222368,0.0002593825,0.001265091],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004589132,"about_ca_system_score_gemma":0.000242365,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000005870003,"about_ca_topic_score_gemma":0.00004577776,"domain_scores_codex":[0.9943448,0.0004950737,0.0006077205,0.0005349484,0.003733252,0.0002842521],"domain_scores_gemma":[0.9974783,0.0003805068,0.0002373278,0.0004448476,0.00130455,0.0001544366],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00002182305,0.00003531803,0.1518332,0.000002863232,0.000008958987,3.572057e-7,0.001980236,0.001401486,0.0004974504,0.0001985193,0.005541462,0.8384783],"study_design_scores_gemma":[0.0007397526,0.000132935,0.5727152,0.00004097949,0.00005357278,0.00001179016,0.001089368,0.3467656,0.00009874941,0.004260594,0.07386751,0.0002238931],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9389021,0.0001167196,0.001573088,0.004080735,0.001059392,0.002091978,0.000003948901,0.00008768168,0.05208437],"genre_scores_gemma":[0.9842925,0.000001037381,0.01271059,0.001654076,0.0002095429,0.0004933351,0.00003578522,0.00001292394,0.0005901749],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8382545,"threshold_uncertainty_score":0.9995126,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2048026043","doi":"10.1016/j.evalprogplan.2013.10.004","title":"Early implementation evaluation of a multi-site housing first intervention for homeless people with mental illness: A mixed methods approach","year":2013,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Homelessness and Social Issues","field":"Health Professions","cited_by":80,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; Douglas Mental Health University Institute; Simon Fraser University; Université de Sherbrooke; University of Toronto; University of Winnipeg; Centre for Addiction and Mental Health; Mental Health Commission of Canada; University of Ottawa; Wilfrid Laurier University","funders":"Health Canada; Mental Health Commission","keywords":"Fidelity; Focus group; Mental illness; Housing First; Qualitative research; Scale (ratio); Qualitative property; Intervention (counseling); Psychology; Program evaluation; Process management; Applied psychology; Mental health; Medical education; Medicine; Business; Engineering; Computer science; Marketing; Sociology; Psychiatry; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.1700243146089559,"gpt":0.558821646774928,"spread":0.3887973321659721,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006515278,0.0001612436,0.0003795465,0.0001469457,0.0007800795,0.00007006816,0.00006324053,0.0001271772,0.0001810812],"category_scores_gemma":[0.0001446779,0.0001369626,0.00006138541,0.0002669672,0.00003943942,0.0004011274,0.00003932762,0.0001406531,0.000003467499],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001443797,"about_ca_system_score_gemma":0.0001157579,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002026134,"about_ca_topic_score_gemma":0.002597338,"domain_scores_codex":[0.9967191,0.00156098,0.000546451,0.0002930124,0.0005812211,0.0002992949],"domain_scores_gemma":[0.9978967,0.0002365104,0.0005128731,0.0001210658,0.001161838,0.00007104613],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00006881918,0.0001529931,0.1164468,0.00040579,0.00006980441,2.615416e-8,0.1554031,0.0001583144,0.0001365212,0.00005712143,0.00001478195,0.7270859],"study_design_scores_gemma":[0.006111043,0.0003155573,0.4201884,0.000366396,0.0002932972,1.605333e-7,0.2975237,0.274543,0.0001401604,0.0001940418,0.0001241019,0.0002001584],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9466329,0.0005771787,0.04443663,0.00004877137,0.0002219998,0.00790445,0.00001913274,0.00007070352,0.00008823294],"genre_scores_gemma":[0.9440646,0.00001163382,0.04717651,0.000008354355,0.00007032721,0.007999325,0.0006317032,0.00002671321,0.00001079881],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7268858,"threshold_uncertainty_score":0.599982,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1995587414","doi":"10.1016/j.evalprogplan.2007.01.006","title":"Quasi-experimental evaluation of a national primary school HIV intervention in Kenya","year":2007,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Adolescent Sexual and Reproductive Health","field":"Health Professions","cited_by":79,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Windsor","funders":"Department for International Development","keywords":"Condom; Abstinence; Medicine; Program evaluation; Intervention (counseling); Sexual intercourse; Sexual abstinence; Human immunodeficiency virus (HIV); Human sexuality; Family medicine; Demography; Family planning; Psychology; Population; Environmental health; Nursing; Gender studies; Political science; Psychiatry; Research methodology; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.2501987740899039,"gpt":0.5794862528234834,"spread":0.3292874787335796,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02079852,0.0001094294,0.0001778315,0.0002638394,0.0002287171,0.000006969202,0.00005078744,0.0001245267,0.0003911505],"category_scores_gemma":[0.00079295,0.0000998572,0.00003151933,0.0002646737,0.00004994357,0.0001928949,0.00004221945,0.000344402,0.00002632918],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005734375,"about_ca_system_score_gemma":0.0004774417,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003462664,"about_ca_topic_score_gemma":0.00002002549,"domain_scores_codex":[0.9963483,0.001181814,0.000699425,0.0003012944,0.001199967,0.0002692616],"domain_scores_gemma":[0.998512,0.0001605708,0.0003427273,0.0000985918,0.000795281,0.00009080514],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0005813802,0.003449044,0.8157699,0.0004306163,0.00003112905,9.59578e-7,0.01173868,0.0003677328,0.001308371,0.0005935921,0.0008193182,0.1649092],"study_design_scores_gemma":[0.004165464,0.0005256276,0.9615978,0.000719282,0.00002642011,0.000002424679,0.0109799,0.01971822,0.0001911325,0.0009268885,0.001004841,0.0001419927],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9846598,0.002455155,0.0006309168,0.0002144257,0.0003695197,0.002943386,0.000004071845,0.00005087556,0.008671856],"genre_scores_gemma":[0.997896,0.000005451471,0.001027845,0.000160017,0.000243302,0.000266081,0.0002167282,0.00001083876,0.0001736979],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1647673,"threshold_uncertainty_score":0.7208393,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2016748774","doi":"10.1016/j.evalprogplan.2007.01.002","title":"Key considerations for logic model development in research partnerships: A Canadian case study","year":2007,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":74,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"University of British Columbia","funders":"National Institute on Minority Health and Health Disparities; Canadian Institutes of Health Research; Health Canada","keywords":"General partnership; Accountability; Logic model; Participatory action research; Community-based participatory research; Public relations; Context (archaeology); Knowledge management; Process (computing); Political science; Sociology; Computer science; Public administration","retraction":null,"screen_n_in":null,"score":{"opus":0.9698488794129135,"gpt":0.8027374069994443,"spread":0.1671114724134692,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.04784323,0.00008726201,0.0001269521,0.0006280564,0.002318097,0.00004821274,0.00005343225,0.00008893763,0.0001044004],"category_scores_gemma":[0.005474149,0.00008467078,0.000008353504,0.0004700577,0.00004771811,0.0001382677,0.00002850665,0.0003829421,0.00001762397],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005929462,"about_ca_system_score_gemma":0.005546407,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.01263323,"about_ca_topic_score_gemma":0.5566695,"domain_scores_codex":[0.9961203,0.001416994,0.0007238,0.0002966061,0.0005467269,0.0008955549],"domain_scores_gemma":[0.9947557,0.003807071,0.0001024228,0.0001268682,0.0007950927,0.0004128385],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.00004850536,0.0001425689,0.2502689,0.0001221757,0.00001320472,0.000178114,0.6843768,0.0009569997,0.00001000504,0.003669055,0.005232249,0.0549814],"study_design_scores_gemma":[0.007892724,0.0007972069,0.1036136,0.000303139,0.00003010379,0.0001114901,0.5308795,0.2849087,0.00001417801,0.01011937,0.06078679,0.0005431097],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.983697,0.00003844333,0.001506948,0.003853811,0.00008661446,0.008878284,0.000008481549,0.00004358185,0.001886824],"genre_scores_gemma":[0.966704,7.347332e-7,0.02724065,0.002593886,0.00005080696,0.003288698,0.00001544958,0.000008772723,0.00009708274],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5440363,"threshold_uncertainty_score":0.9989808,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2062053864","doi":"10.1016/j.evalprogplan.2009.02.002","title":"A measure of community members’ perceptions of the impacts of research partnerships in health and social services","year":2009,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":68,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McMaster University; Western University; Brock University; Child and Family Research Institute","funders":"","keywords":"Community-based participatory research; Focus group; Participatory action research; Public relations; Accountability; Participatory evaluation; Construct (python library); Psychology; Knowledge management; Promotion (chess); Applied psychology; Medical education; Business; Sociology; Political science; Medicine; Marketing","retraction":null,"screen_n_in":null,"score":{"opus":0.8581981758020847,"gpt":0.7562091082784566,"spread":0.1019890675236281,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0241248,0.00005307862,0.0001819094,0.0001704846,0.0008590217,0.000006162938,0.0001190613,0.00006746834,0.00003686657],"category_scores_gemma":[0.0008090232,0.00004075834,0.00001545831,0.0005982086,0.0001831416,0.0001026062,0.00005158569,0.0005262321,4.2846e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00007127379,"about_ca_system_score_gemma":0.000843999,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001494145,"about_ca_topic_score_gemma":0.001649593,"domain_scores_codex":[0.9902321,0.008034518,0.0006460313,0.00007971346,0.000714995,0.0002927004],"domain_scores_gemma":[0.9978126,0.001046572,0.0004220315,0.0001359867,0.0004935279,0.00008933009],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"observational","study_design_scores_codex":[0.00005860392,0.00006645908,0.3922592,0.001278389,0.000003823992,2.038608e-8,0.5806416,0.000007893146,0.0003263512,0.0004924409,0.0002839369,0.02458131],"study_design_scores_gemma":[0.0006974827,0.0002558335,0.8424968,0.0006795449,0.000003828842,3.477978e-7,0.1540315,0.0006509477,0.00001415936,0.000686265,0.0004553097,0.00002796494],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9859824,0.0001249045,0.000003321742,0.0119738,0.00002232431,0.001320177,0.00001380218,0.000009058612,0.0005501985],"genre_scores_gemma":[0.9983346,0.00001536785,0.0002707617,0.001290849,0.00001768642,0.00004934527,0.00001084137,0.000002752458,0.000007778634],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4502376,"threshold_uncertainty_score":0.8361221,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2155432674","doi":"10.1016/j.evalprogplan.2013.12.001","title":"Government and voluntary sector differences in organizational capacity to do and use evaluation","year":2013,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":60,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Government (linguistics); Turnover; Business; Work (physics); Organizational effectiveness; Public relations; Capacity building; Public sector; Psychology; Knowledge management; Political science; Management; Engineering; Economics; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.261696516293106,"gpt":0.472350945174198,"spread":0.2106544288810919,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006262791,0.0001416692,0.0001831651,0.0001732062,0.0001872658,0.0008855403,0.00009331982,0.00006432436,0.001895826],"category_scores_gemma":[0.001964564,0.0001099489,0.00001263262,0.0004608835,0.00005765787,0.0008772758,0.00008263477,0.0001068602,0.00003162247],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001045202,"about_ca_system_score_gemma":0.00009511807,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001136215,"about_ca_topic_score_gemma":0.0001371466,"domain_scores_codex":[0.9957517,0.0005028868,0.0004477686,0.0004427293,0.002677528,0.0001774089],"domain_scores_gemma":[0.9983582,0.0005740916,0.0001484087,0.0001547966,0.0006120219,0.0001524736],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0000108929,0.00003269978,0.7862641,0.000004020257,0.000006409363,1.225577e-7,0.0032672,0.0002009152,0.0001129313,0.0001632443,0.0003412322,0.2095962],"study_design_scores_gemma":[0.0005395694,0.0001186119,0.7527942,0.00003670548,0.00001536885,0.000003261494,0.00132718,0.2412901,0.00001752594,0.003231833,0.00051249,0.0001131667],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9960236,0.0002444235,0.0003208615,0.001109518,0.0001277603,0.001640607,0.0000122493,0.00001977623,0.0005011993],"genre_scores_gemma":[0.9953582,0.00001947446,0.003731599,0.0004001597,0.00004510651,0.000321237,0.00002280777,0.000006641792,0.00009474761],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2410892,"threshold_uncertainty_score":0.9990166,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2791939323","doi":"10.1016/j.evalprogplan.2018.02.005","title":"Promoting physical activity, healthy eating and gross motor skills development among preschoolers attending childcare centers: Process evaluation of the Healthy Start-Départ Santé intervention using the RE-AIM framework","year":2018,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Infant Development and Preterm Care","field":"Medicine","cited_by":59,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Saskatchewan Health; Université de Sherbrooke; University of Saskatchewan","funders":"Heart and Stroke Foundation of Canada; Public Health Agency of Canada","keywords":"Gross motor skill; Intervention (counseling); Psychology; Descriptive statistics; Psychological intervention; Fidelity; Program evaluation; Physical activity; Qualitative property; Medical education; Medicine; Gerontology; Physical therapy; Motor skill; Nursing; Engineering; Developmental psychology; Political science; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.07218495248342228,"gpt":0.4488414195479121,"spread":0.3766564670644899,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.003964813,0.0002192799,0.0002720051,0.0001003013,0.0008354248,0.0001283063,0.0001004459,0.0001096233,0.00002177387],"category_scores_gemma":[0.0006605461,0.0001496322,0.00006839846,0.0002838265,0.0001753939,0.0003002048,0.0001052381,0.0003843693,3.074514e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002086061,"about_ca_system_score_gemma":0.0003218787,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002045701,"about_ca_topic_score_gemma":0.00001233812,"domain_scores_codex":[0.9970348,0.0006285783,0.0004227067,0.0003896927,0.001184041,0.0003401433],"domain_scores_gemma":[0.9983867,0.0001378711,0.0005338239,0.0002013436,0.0006402523,0.00009998574],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0003042028,0.00009689748,0.6186455,0.0005387719,0.00009814526,4.270285e-7,0.09948637,0.00004687495,0.0002189481,0.000003404665,0.000009386263,0.280551],"study_design_scores_gemma":[0.001545277,0.0004887239,0.6872237,0.005209138,0.0002557667,0.00001458604,0.004884985,0.2985373,0.001460193,0.0001321827,0.00004554112,0.0002026893],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9950566,0.0003207676,0.0002118222,0.000303764,0.0002674208,0.003652865,0.000006614719,0.00005383111,0.0001263734],"genre_scores_gemma":[0.9974957,0.00000483332,0.001693325,0.0001206733,0.000295489,0.0003014931,0.00006265202,0.00002319835,0.00000268359],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2984904,"threshold_uncertainty_score":0.6425496,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W977025052","doi":"10.1016/j.evalprogplan.2015.07.003","title":"Knowledge brokering in public health: A critical analysis of the results of a qualitative evaluation","year":2015,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":55,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"McGill University; McGill University Health Centre; Université de Montréal","funders":"Canadian Institutes of Health Research","keywords":"Psychological intervention; Interpersonal communication; Qualitative analysis; Qualitative research; Knowledge management; Psychology; Management science; Public relations; Computer science; Social psychology; Medicine; Political science; Sociology; Nursing; Engineering; Social science","retraction":null,"screen_n_in":null,"score":{"opus":0.960302668591302,"gpt":0.8365556028242945,"spread":0.1237470657670074,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.09964266,0.0002228875,0.001546031,0.001296751,0.0003283114,0.00002043448,0.0002679173,0.0002387303,0.00007164481],"category_scores_gemma":[0.03503788,0.0001613779,0.0001749272,0.004965689,0.0002034114,0.0001770121,0.0001525385,0.0006222692,0.000006522322],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008631122,"about_ca_system_score_gemma":0.01141938,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003309455,"about_ca_topic_score_gemma":0.0005862086,"domain_scores_codex":[0.9615321,0.0323819,0.003369272,0.0004722826,0.001650688,0.0005937661],"domain_scores_gemma":[0.9860955,0.008717107,0.002359461,0.000416126,0.002168708,0.0002430878],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000007889943,0.00004713457,0.0001847423,0.005529599,0.00009792251,3.86339e-8,0.1591399,0.00001295085,3.759075e-8,0.000328845,0.0006247687,0.8340262],"study_design_scores_gemma":[0.00150968,0.0001980114,0.002794214,0.0144007,0.001520027,7.4286e-7,0.02405972,0.0156825,5.469568e-8,0.000209298,0.9394128,0.0002122542],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00069695,0.9843714,0.0001125379,0.00243501,0.0004109872,0.009050487,0.000384404,0.00003385744,0.002504342],"genre_scores_gemma":[0.06374652,0.9074879,0.007899274,0.001601286,0.0004414124,0.01617981,0.002248727,0.0001275693,0.0002675048],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.938788,"threshold_uncertainty_score":0.994185,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2000732508","doi":"10.1016/j.evalprogplan.2008.01.003","title":"A framework of operating models for interdisciplinary research programs in clinical service organizations","year":2008,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Interprofessional Education and Collaboration","field":"Health Professions","cited_by":51,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Child and Family Research Institute; Thames Valley Children's Centre; Holland Bloorview Kids Rehabilitation Hospital","funders":"","keywords":"Context (archaeology); Knowledge management; Service (business); Service delivery framework; Service model; Management science; Process management; Computer science; Engineering; Business","retraction":null,"screen_n_in":null,"score":{"opus":0.4872797990941883,"gpt":0.6658440447310565,"spread":0.1785642456368682,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005914301,0.00009410809,0.0002075801,0.0002061381,0.001064903,0.00001661242,0.000110519,0.0002416828,0.00009545886],"category_scores_gemma":[0.001665868,0.00008487645,0.00002396664,0.001260615,0.0000780331,0.0002249506,0.000135225,0.0006596357,0.00001048568],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008537941,"about_ca_system_score_gemma":0.001686061,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004459696,"about_ca_topic_score_gemma":0.0001010093,"domain_scores_codex":[0.9968325,0.001338109,0.0008511988,0.000273434,0.0004195257,0.0002852575],"domain_scores_gemma":[0.9952065,0.001197179,0.000206491,0.0001611364,0.003146561,0.00008216152],"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":[0.0006890938,0.001504079,0.4902714,0.0006770201,0.00005641475,0.000002237531,0.3457855,0.003372292,0.0001315785,0.0252952,0.004517425,0.1276978],"study_design_scores_gemma":[0.002376189,0.0007968043,0.123101,0.00361037,0.0000305043,0.000004357589,0.09849992,0.7365241,0.00001477232,0.03322558,0.001557895,0.000258507],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9820433,0.0002491676,0.007134291,0.00340892,0.0005581919,0.004470423,0.000005910786,0.00006469068,0.002065108],"genre_scores_gemma":[0.9597296,0.00001621838,0.03705801,0.0003546359,0.0002186676,0.002146848,0.0002259134,0.00001915418,0.000231002],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7331518,"threshold_uncertainty_score":0.8190479,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2580523678","doi":"10.1016/j.evalprogplan.2017.01.004","title":"Implementation evaluation of early intensive behavioral intervention programs for children with autism spectrum disorders: A systematic review of studies in the last decade","year":2017,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Behavioral and Psychological Studies","field":"Psychology","cited_by":51,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université du Québec en Outaouais; Université du Québec à Trois-Rivières","funders":"","keywords":"Intervention (counseling); Autism; Psychological intervention; Systematic review; Autism spectrum disorder; Protocol (science); Psychology; Clinical psychology; Medicine; MEDLINE; Developmental psychology; Psychiatry; Alternative medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.6237094503308981,"gpt":0.6133015461684089,"spread":0.01040790416248916,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006932803,0.0003263339,0.001564004,0.0001540827,0.0001345327,0.00004784423,0.0002183505,0.000145271,0.00002747577],"category_scores_gemma":[0.0001776346,0.000181515,0.0002962946,0.0002573523,0.0002056213,0.0000994053,0.00004288034,0.0002258601,0.000001701045],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006607097,"about_ca_system_score_gemma":0.00004768938,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001294372,"about_ca_topic_score_gemma":0.0002459218,"domain_scores_codex":[0.9956856,0.001751614,0.001234173,0.0004082635,0.0006855076,0.0002348536],"domain_scores_gemma":[0.9969544,0.0001720193,0.002111364,0.0003398573,0.000398088,0.00002428652],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.00002097452,0.0002931904,0.00283643,0.1431765,0.0004691701,4.922672e-7,0.003763548,1.004945e-7,6.025942e-9,0.00009327687,0.0000450861,0.8493012],"study_design_scores_gemma":[0.002876838,0.005300769,0.05175592,0.9003553,0.02777315,0.0000647595,0.008879634,0.000008918827,9.740496e-8,0.0007980245,0.00172468,0.0004619313],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.007904261,0.9682391,0.0000172223,0.00008530168,0.0001709282,0.02349697,0.00003002851,0.00001925746,0.00003688494],"genre_scores_gemma":[0.05146291,0.9230843,0.0001424959,0.00002231722,0.00003851996,0.02444263,0.0007702758,0.00002677234,0.000009778504],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8488393,"threshold_uncertainty_score":0.7401966,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2771999481","doi":"10.1016/j.evalprogplan.2017.12.001","title":"Uncovering the mysteries of inclusion: Empirical and methodological possibilities in participatory evaluation in an international context","year":2017,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":45,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Inclusion (mineral); Citizen journalism; Privilege (computing); Context (archaeology); Participatory evaluation; Participatory action research; Empirical evidence; Public relations; Sociology; Empirical research; Power (physics); Engineering ethics; Political science; Psychology; Social science; Engineering; Epistemology; Geography","retraction":null,"screen_n_in":null,"score":{"opus":0.8922807118038334,"gpt":0.7301766260124527,"spread":0.1621040857913807,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06150194,0.0002345279,0.0009157932,0.0004416065,0.0002964109,0.0004517745,0.0005353441,0.0002130379,0.0003130883],"category_scores_gemma":[0.008204847,0.0001384896,0.00009085654,0.0002844521,0.0002750381,0.0007465841,0.0005445777,0.0003413203,0.000003183999],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001533829,"about_ca_system_score_gemma":0.0007090001,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002933971,"about_ca_topic_score_gemma":0.0002657997,"domain_scores_codex":[0.9891104,0.006051161,0.001554442,0.0005223833,0.002554581,0.000207049],"domain_scores_gemma":[0.9949964,0.002942557,0.0009654482,0.0004785121,0.0005500864,0.00006696907],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00003042177,0.00004753995,0.03006868,0.0001551073,0.00001388265,9.290169e-7,0.01093271,0.0002132326,1.791597e-7,0.00006988383,0.000009524905,0.9584579],"study_design_scores_gemma":[0.002250069,0.0007437077,0.1942967,0.008178119,0.0004170715,0.00003306469,0.01486506,0.173913,0.000001856486,0.02117863,0.583485,0.0006377682],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"empirical","genre_scores_codex":[0.2031894,0.7926543,0.00003871861,0.0002769775,0.0004607272,0.002352751,0.00001233493,0.00001553845,0.0009992184],"genre_scores_gemma":[0.5856336,0.4113244,0.001079873,0.0001393139,0.0001568597,0.001426731,0.0001478871,0.00001879178,0.00007252415],"genre_candidate":"review","genre_consensus":null,"teacher_disagreement_score":0.9578201,"threshold_uncertainty_score":0.9822561,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3001351247","doi":"10.1016/j.evalprogplan.2020.101786","title":"Pedagogical considerations for the teaching of evaluation","year":2020,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation of Teaching Practices","field":"Social Sciences","cited_by":44,"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","funders":"","keywords":"Andragogy; Process (computing); Identification (biology); Field (mathematics); Pedagogy; Instructional design; Mathematics education; Teaching method; Computer science; Engineering ethics; Psychology; Engineering; Adult education","retraction":null,"screen_n_in":null,"score":{"opus":0.6344838083863473,"gpt":0.6117541664420982,"spread":0.02272964194424909,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.01617487,0.00007222695,0.0001050891,0.00003406674,0.001278236,0.0001746616,0.00007431147,0.00005912043,0.0003089771],"category_scores_gemma":[0.02855359,0.00005707356,0.00004402567,0.00009296886,0.0001302158,0.0003149755,0.00002001916,0.0001784527,0.000003339728],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00003313177,"about_ca_system_score_gemma":0.0003711667,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001036926,"about_ca_topic_score_gemma":0.00003595082,"domain_scores_codex":[0.9967554,0.001631881,0.0002663004,0.0001875607,0.001023592,0.0001352941],"domain_scores_gemma":[0.9954469,0.003615552,0.0002505463,0.00008277421,0.0005306057,0.00007360082],"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":[0.00005985391,0.00008449308,0.005218524,0.00002809184,0.00007942063,2.154843e-7,0.2427522,0.00730504,0.0001341358,0.09177886,0.004163554,0.6483957],"study_design_scores_gemma":[0.00161189,0.0003916729,0.01102913,0.00006161956,0.000694124,0.000002248711,0.05733529,0.7004948,0.00004144492,0.01478834,0.2133286,0.0002208564],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4891191,0.008027257,0.03083561,0.395994,0.001033271,0.01829105,0.00002549232,0.0006265502,0.05604761],"genre_scores_gemma":[0.9854974,0.00001543454,0.013074,0.000682208,0.0002739875,0.0003794615,0.00001848332,0.000006462474,0.00005255886],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6931897,"threshold_uncertainty_score":0.9831288,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2301037950","doi":"10.1016/j.evalprogplan.2016.03.001","title":"Planning an integrated agriculture and health program and designing its evaluation: Experience from Western Kenya","year":2016,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Child Nutrition and Water Access","field":"Nursing","cited_by":43,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Public Health Ontario; University of Toronto","funders":"PATH; Bill and Melinda Gates Foundation","keywords":"Program evaluation; Agriculture; Occupational safety and health; Environmental planning; Human factors and ergonomics; Engineering; Poison control; Business; Economic growth; Environmental health; Political science; Geography; Medicine; Economics; Public administration","retraction":null,"screen_n_in":null,"score":{"opus":0.1296405227431724,"gpt":0.4453615327165398,"spread":0.3157210099733674,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001544505,0.000271856,0.0002774891,0.0001300705,0.0005841114,0.0007354029,0.0001011963,0.0001373861,0.00004915096],"category_scores_gemma":[0.0001383629,0.0001882592,0.00002403605,0.0002216856,0.0001147192,0.001014245,0.00004758194,0.0001815652,0.000002615666],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008932937,"about_ca_system_score_gemma":0.00005405906,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007596995,"about_ca_topic_score_gemma":0.00002481518,"domain_scores_codex":[0.9974478,0.0005563013,0.0003956616,0.0006407976,0.0006027984,0.0003566451],"domain_scores_gemma":[0.9989136,0.0001253736,0.0002181017,0.0001531618,0.0003091451,0.0002806069],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001468641,0.00012207,0.1448196,0.00003648206,0.00002099251,0.000002457967,0.01440181,0.00002500849,0.0009517054,0.000005800033,0.0002772692,0.8391899],"study_design_scores_gemma":[0.008566248,0.002299514,0.8975104,0.003506133,0.000209219,0.000110152,0.005766731,0.05398994,0.003669366,0.0005518951,0.02285971,0.0009606967],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9843107,0.008377605,0.0006496498,0.003364366,0.0002668333,0.002545894,0.0000212956,0.000392798,0.00007084815],"genre_scores_gemma":[0.9923243,0.00004693877,0.005896619,0.0005781938,0.0002377546,0.000500665,0.0003740982,0.00002388675,0.00001757829],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8382292,"threshold_uncertainty_score":0.7676987,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2791109575","doi":"10.1016/j.evalprogplan.2018.01.008","title":"Factors associated with the implementation of community-based peer-led health promotion programs: A scoping review","year":2018,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":43,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal; Institut Universitaire de Gériatrie de Montréal","funders":"","keywords":"Health promotion; Promotion (chess); Peer review; Occupational safety and health; Environmental health; Psychology; Public relations; Medicine; Nursing; Political science; Public health","retraction":null,"screen_n_in":null,"score":{"opus":0.8768914231027989,"gpt":0.7621768679071069,"spread":0.114714555195692,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.03462872,0.0004167376,0.00145019,0.0002618505,0.00239169,0.00005961619,0.0003460811,0.000202532,0.0002162423],"category_scores_gemma":[0.0018914,0.0002501552,0.0001305547,0.001375393,0.0002212175,0.0001904317,0.00008841536,0.001072027,0.000007786115],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005440431,"about_ca_system_score_gemma":0.007531935,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002094229,"about_ca_topic_score_gemma":0.0008876997,"domain_scores_codex":[0.9755621,0.01968792,0.002108238,0.000339014,0.001572752,0.0007299586],"domain_scores_gemma":[0.9900231,0.003105437,0.004648428,0.0004859685,0.001524021,0.0002130789],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.000003107149,0.0000841828,0.001101045,0.1438656,0.00007488895,7.420933e-8,0.01176069,3.330743e-7,2.379027e-8,0.00001063183,0.002991798,0.8401077],"study_design_scores_gemma":[0.001273974,0.002009182,0.001239969,0.6650711,0.0008088907,9.286105e-7,0.006536102,0.00006564317,3.066444e-7,0.00001076521,0.3226399,0.0003433037],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.001963798,0.9438758,0.0002380374,0.00562518,0.0002221322,0.04755618,0.0001531873,0.0002454963,0.0001201418],"genre_scores_gemma":[0.001665661,0.9579924,0.0007049799,0.006788302,0.0002135807,0.02100438,0.0114251,0.0001366834,0.00006893384],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.8397644,"threshold_uncertainty_score":0.9999951,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1964714699","doi":"10.1016/j.evalprogplan.2009.03.003","title":"Power and perceptions in participatory monitoring and evaluation","year":2009,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":43,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Guelph","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Accountability; Citizen journalism; Participatory evaluation; Bureaucracy; Perception; Power (physics); Participatory action research; Participant observation; Participatory design; Public relations; Politics; Psychology; Business; Political science; Economic growth; Sociology; Engineering; Public administration; Operations management; Economics","retraction":null,"screen_n_in":null,"score":{"opus":0.3716448119431371,"gpt":0.6018709322341447,"spread":0.2302261202910075,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0128765,0.000115235,0.0001543247,0.0002828964,0.0002260918,0.0003581114,0.00006622229,0.00006785036,0.0004304429],"category_scores_gemma":[0.0008250647,0.00009649183,0.00001778114,0.0003432528,0.00005574514,0.000561394,0.00002703985,0.0001335957,0.00001271745],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005056155,"about_ca_system_score_gemma":0.000105917,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004837704,"about_ca_topic_score_gemma":0.000007465026,"domain_scores_codex":[0.9969747,0.0005224984,0.0004350315,0.0003643426,0.001510559,0.0001928355],"domain_scores_gemma":[0.9989929,0.0002550977,0.0001238534,0.000158869,0.0003585484,0.000110684],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00001347234,0.00003624601,0.3704394,0.000001671012,0.000002846629,3.997633e-7,0.005577287,0.0004962815,0.0001606826,0.0001398136,0.00006339543,0.6230686],"study_design_scores_gemma":[0.0007268778,0.0001797033,0.8164224,0.00004786882,0.00002067346,0.000004645942,0.003895731,0.1731154,0.00001900716,0.004780173,0.0006865731,0.000100963],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9929388,0.001722088,0.00008028885,0.001041026,0.0001336096,0.0007384908,9.068048e-7,0.00003343163,0.003311344],"genre_scores_gemma":[0.9982011,0.0000665777,0.001374004,0.000123376,0.00004770341,0.0001293901,0.000007619761,0.000004090185,0.00004612599],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6229676,"threshold_uncertainty_score":0.4713047,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"qualitative","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"}],"label_agreement":"split"},{"id":"W3004981871","doi":"10.1016/j.evalprogplan.2020.101793","title":"Participatory analytic hierarchy process for resource allocation in agricultural development projects","year":2020,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Multi-Criteria Decision Making","field":"Decision Sciences","cited_by":41,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"European Commission; Canadian Patient Safety Institute","keywords":"Analytic hierarchy process; Citizen journalism; Sustainability; Pairwise comparison; Process management; Process (computing); Resource allocation; Sustainable development; Resource (disambiguation); Management science; Business; Computer science; Environmental resource management; Operations research; Engineering; Economics; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.584576250057631,"gpt":0.5587150463818018,"spread":0.02586120367582923,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005950759,0.000140714,0.0002324861,0.0002263631,0.0001851644,0.0003767371,0.0002177312,0.00006688588,0.00004710126],"category_scores_gemma":[0.005847355,0.0001031597,0.00003617337,0.0009082172,0.00003694069,0.0003231012,0.00005628177,0.0001143771,0.00001744078],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005410082,"about_ca_system_score_gemma":0.0001872229,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":9.958491e-7,"about_ca_topic_score_gemma":0.000006716896,"domain_scores_codex":[0.9966959,0.0003381534,0.000741681,0.000560572,0.001397474,0.0002662539],"domain_scores_gemma":[0.9983727,0.0006024317,0.000249816,0.0001353563,0.0004932742,0.0001464586],"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":[0.000216227,0.00008699498,0.02641718,0.00005270495,0.00001485723,0.000003688452,0.04745869,0.01157316,0.0004335194,0.00008520663,0.001012246,0.9126455],"study_design_scores_gemma":[0.001628126,0.0001865793,0.1487344,0.0001502817,0.00002440472,0.000004414093,0.01136827,0.8140553,0.0005742839,0.001006424,0.02198594,0.0002815869],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9912979,0.0003441062,0.004140113,0.001540103,0.00006455893,0.002074497,0.000001854694,0.00007711741,0.0004597669],"genre_scores_gemma":[0.9901942,4.553661e-7,0.008601583,0.00036135,0.000072804,0.0006798532,0.00004150474,0.000008719851,0.00003952643],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9123639,"threshold_uncertainty_score":0.7000253,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2082907843","doi":"10.1016/j.evalprogplan.2012.05.003","title":"Evaluating local food programs: The case of Select Nova Scotia","year":2012,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Organic Food and Agriculture","field":"Agricultural and Biological Sciences","cited_by":40,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Technical University of Nova Scotia; Nova Scotia Department of Agriculture","funders":"","keywords":"Nova scotia; Nova (rocket); Marketing; Local government; Business; Consumption (sociology); Perception; Government (linguistics); Consumer behaviour; Advertising; Psychology; Political science; Geography; Engineering; Sociology; Public administration","retraction":null,"screen_n_in":null,"score":{"opus":0.1144218093796592,"gpt":0.375084479348663,"spread":0.2606626699690038,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001731416,0.0001175742,0.0001171762,0.000007036667,0.0002152212,0.00006661128,0.00008369513,0.00007586576,0.0001779103],"category_scores_gemma":[0.00009649301,0.00003560162,0.00004555397,0.0003737875,0.00006997505,0.0001380947,0.00003969017,0.0001329291,0.000007148713],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00001308428,"about_ca_system_score_gemma":0.00001121235,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001436145,"about_ca_topic_score_gemma":0.0003923702,"domain_scores_codex":[0.9988725,0.0002254974,0.0002005605,0.0001567213,0.0002734465,0.000271304],"domain_scores_gemma":[0.9994679,0.0001404946,0.0001245152,0.00004025119,0.000147288,0.00007952227],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00001038194,0.000108698,0.009742148,0.00001082604,0.00002386771,0.000001450733,0.001202725,0.00001541951,0.005701425,0.00009188586,0.0001983533,0.9828928],"study_design_scores_gemma":[0.003183603,0.01992981,0.7698415,0.000931744,0.001224749,0.005027841,0.08897933,0.04911127,0.02362183,0.001054374,0.03499499,0.002098947],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9954786,0.002653917,0.000006451616,0.0005725901,0.0001087309,0.0007645125,0.000004779094,0.00006251459,0.000347922],"genre_scores_gemma":[0.9991279,0.000002819372,0.0004178903,0.00005734307,0.0002883963,0.00003021914,0.00006312336,0.000001006653,0.0000113532],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9807939,"threshold_uncertainty_score":0.1947993,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2011328276","doi":"10.1016/j.evalprogplan.2014.09.002","title":"Protocol: Realist synthesis of the impact of unemployment insurance policies on poverty and health","year":2014,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Employment and Welfare Studies","field":"Health Professions","cited_by":39,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Wilfrid Laurier University; University of Toronto","funders":"","keywords":"Unemployment; Poverty; Economics; Context (archaeology); Public economics; Actuarial science; Economic growth","retraction":null,"screen_n_in":null,"score":{"opus":0.1826845635748133,"gpt":0.5455077228220494,"spread":0.3628231592472361,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002026206,0.0001027783,0.0002349483,0.00004823311,0.0004257369,0.00000615877,0.00005395669,0.0000448708,0.0000160066],"category_scores_gemma":[0.0003075543,0.00005750737,0.00004006765,0.0001164332,0.00007329533,0.00003531209,0.00005152058,0.0001170528,5.130098e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005110263,"about_ca_system_score_gemma":0.0001087327,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001306716,"about_ca_topic_score_gemma":0.00006043236,"domain_scores_codex":[0.9983891,0.0006600249,0.0003572258,0.0001261647,0.0002838203,0.0001836804],"domain_scores_gemma":[0.9989714,0.0003096189,0.0004015032,0.000146286,0.0001267162,0.00004449957],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0000771239,0.00006120568,0.8907253,0.0002633303,0.00003196575,1.388049e-8,0.003795953,0.00005650123,0.00001564771,0.0006003287,0.00318321,0.1011894],"study_design_scores_gemma":[0.0006278742,0.0005875113,0.993266,0.0009879556,0.00001214082,1.976779e-7,0.0003971399,0.0005861996,0.00004926026,0.0006639207,0.002763997,0.00005782138],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9286068,0.0001793715,0.00001368714,0.007965785,0.00008872261,0.05324915,0.00004227704,0.00006368187,0.00979051],"genre_scores_gemma":[0.984423,0.00001315086,0.0000879629,0.0003762267,0.00004135442,0.01498746,0.000004066337,0.000008460237,0.00005826891],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1025407,"threshold_uncertainty_score":0.3274467,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2737734760","doi":"10.1016/j.evalprogplan.2017.07.006","title":"Finding the sweet spot: Developing, implementing and evaluating a burn out and compassion fatigue intervention for third year medical trainees","year":2017,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Healthcare professionals’ stress and burnout","field":"Health Professions","cited_by":39,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"Carleton University; CARE Canada; Bruyère","funders":"Associated Medical Services","keywords":"Burnout; Compassion fatigue; Intervention (counseling); Medicine; Nursing; Compassion; Emotional exhaustion; Psychology; Medical education; Clinical psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.3951525910122899,"gpt":0.6195339165471256,"spread":0.2243813255348358,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.01219346,0.0001665576,0.0002562099,0.00007473811,0.004950578,0.000198434,0.0001499509,0.0001978816,0.00009893757],"category_scores_gemma":[0.003281252,0.0001120223,0.00004017428,0.00003771757,0.0001019785,0.0002050776,0.0002750713,0.0004746896,0.000001937874],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004899957,"about_ca_system_score_gemma":0.0002687524,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000121817,"about_ca_topic_score_gemma":0.0002172572,"domain_scores_codex":[0.9971694,0.0007776985,0.0006289252,0.0003332973,0.0006273044,0.0004633925],"domain_scores_gemma":[0.9977818,0.000996916,0.0005917669,0.0001770358,0.0002967237,0.0001557296],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00006500779,0.00002271928,0.1560586,0.0005659934,0.0000317042,8.898285e-7,0.01464511,0.000001096376,0.00003021362,0.001166903,0.0006354895,0.8267763],"study_design_scores_gemma":[0.008503877,0.0007176956,0.6431395,0.0240341,0.0001950409,0.00001160917,0.04583163,0.2492274,0.00004186055,0.006887622,0.02086692,0.0005428336],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9756444,0.002280034,0.004741334,0.01273933,0.0006032895,0.00319528,0.00001704504,0.00007705494,0.0007022299],"genre_scores_gemma":[0.9896591,0.0001301883,0.008250846,0.0003663523,0.0003866183,0.0009351574,0.0001076484,0.00001820947,0.0001459315],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8262334,"threshold_uncertainty_score":0.9963449,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2041414609","doi":"10.1016/j.evalprogplan.2009.05.001","title":"A mixed methods evaluation of televised health promotion advertisements targeted at older adults","year":2009,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Behavioral Health and Interventions","field":"Psychology","cited_by":38,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Alberta","funders":"Canadian Institutes of Health Research","keywords":"Credibility; Advertising; Recall; Health promotion; Focus group; Psychology; Population; Consumption (sociology); Qualitative research; Government (linguistics); Source credibility; Promotion (chess); Public health; Medicine; Marketing; Environmental health; Business; Political science; Nursing; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.235724611060932,"gpt":0.5761175877544217,"spread":0.3403929766934897,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.006219268,0.0001446159,0.0002313282,0.0001727096,0.0001987275,0.00001883569,0.00006622863,0.0001151621,0.001942154],"category_scores_gemma":[0.0001446914,0.0001401552,0.00007364809,0.0002726206,0.00002890789,0.0001254494,0.0000152174,0.0001309034,0.00001672168],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001716453,"about_ca_system_score_gemma":0.0001141448,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006861821,"about_ca_topic_score_gemma":0.00001384286,"domain_scores_codex":[0.9960698,0.001989264,0.0006395713,0.00034818,0.0006410409,0.0003121754],"domain_scores_gemma":[0.998677,0.00003791946,0.0004251249,0.0002012309,0.0005032712,0.0001554763],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001414203,0.0007371455,0.004459624,0.00003407164,0.00001968372,2.45651e-7,0.003444746,0.00002283778,0.0001013546,0.00005015848,0.00169839,0.9892903],"study_design_scores_gemma":[0.00450647,0.002560738,0.949688,0.0005639927,0.0001541151,0.000006813729,0.000805181,0.04049858,0.0002368512,0.0002169587,0.0006068945,0.0001554311],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9869623,0.005474268,0.001459871,0.0005938422,0.0006087763,0.003738055,0.00001145355,0.0001049512,0.00104649],"genre_scores_gemma":[0.9868655,0.00001024318,0.01175555,0.0001455006,0.00004180598,0.0005715698,0.000485147,0.00001092803,0.0001137482],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9891349,"threshold_uncertainty_score":0.9989702,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2587275352","doi":"10.1016/j.evalprogplan.2017.02.005","title":"Concept mapping internal validity: A case of misconceived mapping?","year":2017,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":38,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Cégep Marie-Victorin; Université de Montréal","funders":"Canadian Institutes of Health Research","keywords":"Internal validity; Engineering; Psychology; Computer science; Mathematics; Statistics","retraction":null,"screen_n_in":null,"score":{"opus":0.4979744957311376,"gpt":0.5785909440871707,"spread":0.08061644835603315,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.00959699,0.0001351119,0.000269236,0.0002034056,0.0006573299,0.0007386549,0.0003805186,0.00006993498,0.001059931],"category_scores_gemma":[0.002671242,0.0001071661,0.00007508673,0.0001432789,0.0002126008,0.0005965514,0.000134139,0.0001486355,0.00002658757],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002991707,"about_ca_system_score_gemma":0.0001740734,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00007297783,"about_ca_topic_score_gemma":0.00001955565,"domain_scores_codex":[0.9970604,0.0004077161,0.0006672569,0.0003615702,0.001306251,0.0001968324],"domain_scores_gemma":[0.997266,0.0004459979,0.0008813319,0.0005303793,0.0007667334,0.0001095485],"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":[0.00002283497,0.00004290892,0.1210278,0.00001084747,0.00002842706,0.00003735839,0.009306656,0.0002398418,0.0001730199,0.0002514663,0.0009030397,0.8679558],"study_design_scores_gemma":[0.002760631,0.0003797534,0.3224814,0.0003565642,0.00005973629,0.0003762406,0.01463489,0.6146403,0.0006511993,0.006115142,0.03719424,0.0003499398],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9770821,0.0003589245,0.005276906,0.0007840182,0.0005098569,0.0006865867,0.000009311154,0.00003772357,0.0152546],"genre_scores_gemma":[0.9935855,0.000008294409,0.00553329,0.000120684,0.0001138178,0.00006123848,0.0000131098,0.000006384838,0.0005576419],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8676059,"threshold_uncertainty_score":0.9998533,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3006263836","doi":"10.1016/j.evalprogplan.2020.101795","title":"What are the characteristics of the champion that influence the implementation of quality improvement programs?","year":2020,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":38,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal","funders":"Pan American Health Organization","keywords":"Champion; Quality (philosophy); Accountability; Thematic analysis; Health care; Public relations; Proactivity; Psychology; Grounded theory; Quality management; Qualitative research; Business; Sociology; Political science; Marketing; Social psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.5955657876390369,"gpt":0.6774038162348679,"spread":0.081838028595831,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.006657186,0.0001101738,0.0001829518,0.00002494936,0.0008035112,0.0000583448,0.0002808088,0.00005252186,0.00007960962],"category_scores_gemma":[0.0008885231,0.00005610474,0.00004131514,0.000403239,0.0001863892,0.0003164481,0.0001507962,0.0002840211,0.000003662974],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00005364195,"about_ca_system_score_gemma":0.0003918029,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002421346,"about_ca_topic_score_gemma":0.0001047287,"domain_scores_codex":[0.9958352,0.001797492,0.0009052721,0.0002025553,0.0009585306,0.0003009129],"domain_scores_gemma":[0.996747,0.0007737438,0.001611039,0.0002772343,0.0005100477,0.00008093688],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00002631257,0.00001574104,0.5586294,0.0002869109,0.00001474584,4.304633e-8,0.09472767,0.000008382555,0.0006191349,0.0002731633,0.000175855,0.3452227],"study_design_scores_gemma":[0.0005551815,0.000154525,0.8615906,0.0002035199,0.00003046563,2.348037e-7,0.1295761,0.0008442115,0.0003732532,0.0001356706,0.006475243,0.00006098705],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9484285,0.00008333892,0.00008804773,0.0463838,0.0002780702,0.004646896,0.00002822636,0.00002708934,0.00003608879],"genre_scores_gemma":[0.9879419,0.00005197337,0.0001047546,0.01097439,0.00007613613,0.0008117185,0.00002248827,0.000007260994,0.000009354507],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3451617,"threshold_uncertainty_score":0.618004,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1985788625","doi":"10.1016/j.evalprogplan.2009.10.006","title":"Comparing and contrasting poverty reduction performance of social welfare programs across jurisdictions in Canada using Data Envelopment Analysis (DEA): An exploratory study of the era of devolution","year":2009,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Efficiency Analysis Using DEA","field":"Decision Sciences","cited_by":38,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Lakehead University; University of Windsor","funders":"","keywords":"Devolution (biology); Data envelopment analysis; Poverty; Public economics; Welfare; Benchmark (surveying); Social Welfare; Social planner; Population; Economics; Economic growth; Business; Political science; Geography; Sociology; Statistics","retraction":null,"screen_n_in":null,"score":{"opus":0.2189168893901448,"gpt":0.4511346694288697,"spread":0.232217780038725,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005329154,0.0001006931,0.0003604837,0.0001832994,0.0004533714,0.00006379143,0.0002633054,0.00003459681,0.000003298399],"category_scores_gemma":[0.000271792,0.00007565451,0.00003020836,0.001955161,0.00008314398,0.0004364865,0.00009433509,0.0001374724,1.513605e-8],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001498869,"about_ca_system_score_gemma":0.000379261,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.03031481,"about_ca_topic_score_gemma":0.160611,"domain_scores_codex":[0.9968387,0.0005116609,0.0008329029,0.0003489325,0.001308791,0.000159033],"domain_scores_gemma":[0.9984203,0.00007222997,0.0007206585,0.0003209438,0.0004320162,0.00003384248],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00002950528,0.0002110238,0.7764427,0.000005148293,0.0000664897,1.348603e-7,0.01561393,0.06580096,0.0001620473,0.000003960769,0.000001310534,0.1416628],"study_design_scores_gemma":[0.000242228,0.00007213833,0.5204611,0.00002205012,0.0001436548,6.430803e-7,0.01894275,0.4600364,0.00002826221,0.000005605535,0.000004184576,0.00004094421],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9990317,0.0001062357,0.0002212353,0.00007067469,0.00005428471,0.0004889719,0.000008065563,0.000007371018,0.00001148267],"genre_scores_gemma":[0.9992762,0.000001431986,0.0006657612,0.000004999956,0.000008631351,0.000008207317,0.00003158295,0.000002706449,5.31834e-7],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3942354,"threshold_uncertainty_score":0.9761424,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1996512626","doi":"10.1016/j.evalprogplan.2004.04.005","title":"Step down continuing care in the treatment of substance abuse: correlates of participation and outcome effects","year":2004,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Substance Abuse Treatment and Outcomes","field":"Medicine","cited_by":36,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"Center for Substance Abuse Treatment; National Institute on Drug Abuse; Substance Abuse and Mental Health Services Administration; U.S. Department of Health and Human Services","keywords":"Continuing care; Medicine; Continuing education; Substance abuse; Quarter (Canadian coin); Substance use; Ambulatory care; Family medicine; Psychiatry; Health care; Nursing","retraction":null,"screen_n_in":null,"score":{"opus":0.0507464267641588,"gpt":0.4005903750493499,"spread":0.3498439482851911,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00023915,0.00009222853,0.0002231036,0.00006232535,0.00003049488,0.000009371951,0.00001665529,0.00003968841,0.00000332744],"category_scores_gemma":[0.00002628354,0.00005758795,0.00002985703,0.0001291291,0.00004437621,0.00005398695,8.084642e-7,0.00003601056,3.772508e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004915609,"about_ca_system_score_gemma":0.00001157024,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000009934496,"about_ca_topic_score_gemma":0.00009726473,"domain_scores_codex":[0.9992883,0.00006197878,0.0002272105,0.0001129456,0.0002079678,0.000101645],"domain_scores_gemma":[0.9995453,0.0001601969,0.0001066119,0.00009049764,0.00007345279,0.0000239384],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001288233,0.0001281354,0.7597731,0.00005423154,0.00002483465,0.000006006376,0.1382084,0.00009218905,0.00006814531,0.00002924762,3.443288e-7,0.1014866],"study_design_scores_gemma":[0.008417334,0.00153975,0.9773505,0.0005215182,0.0003257896,0.000003904546,0.009624308,0.0006428857,0.001424907,0.00004467608,0.00004221779,0.00006215151],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9846339,0.01391414,0.00001959304,0.00004604927,0.00003868006,0.001263462,6.203618e-7,0.0000167691,0.00006684908],"genre_scores_gemma":[0.9987195,0.0003615764,0.0006995401,0.00001286155,0.00001154362,0.000148445,0.00003519334,0.000005123259,0.000006280235],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2175775,"threshold_uncertainty_score":0.2348368,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2750400467","doi":"10.1016/j.evalprogplan.2017.08.010","title":"On the evaluation of social innovations and social enterprises: Recognizing and integrating two solitudes in the empirical knowledge base","year":2017,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":36,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"University of Ottawa","keywords":"Scholarship; Knowledge management; Empirical research; Field (mathematics); Sociology; Knowledge base; Public relations; Computer science; Political science; Epistemology","retraction":null,"screen_n_in":null,"score":{"opus":0.6846078680975961,"gpt":0.6471283739198773,"spread":0.03747949417771879,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.0390056,0.0001232908,0.0001802763,0.0001963483,0.001593399,0.0008561378,0.000254827,0.00006022994,0.0001443227],"category_scores_gemma":[0.005604547,0.00006998098,0.00003240067,0.0002793555,0.0002420636,0.0003854732,0.00009431857,0.0002406703,0.000003937937],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004328533,"about_ca_system_score_gemma":0.0002509356,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001512121,"about_ca_topic_score_gemma":0.0001495005,"domain_scores_codex":[0.9957359,0.00161689,0.0005498333,0.0002965443,0.001634289,0.0001664762],"domain_scores_gemma":[0.9967299,0.001671499,0.0004872048,0.0002070193,0.000880816,0.00002359867],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00003590344,0.00009879303,0.08210123,0.000009012349,0.00001556848,1.802711e-7,0.06834025,0.00003793235,0.00005167818,0.00621044,0.0007345802,0.8423644],"study_design_scores_gemma":[0.001402505,0.0001307908,0.6369286,0.00008932938,0.00006845092,0.000002816602,0.02666539,0.3137105,0.00002294176,0.01989461,0.0009681939,0.0001158853],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9818749,0.0001768179,0.0001605109,0.007398576,0.0001023979,0.00093383,0.000004495779,0.000009603843,0.009338871],"genre_scores_gemma":[0.9989105,0.00000534311,0.0003463363,0.0003822409,0.00009373339,0.0002202619,0.0000162635,0.000004933532,0.00002034256],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8422486,"threshold_uncertainty_score":0.9997064,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2015075110","doi":"10.1016/j.evalprogplan.2009.12.003","title":"Relationship between perceived improvement and treatment satisfaction among clients of a methadone maintenance program","year":2009,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Substance Abuse Treatment and Outcomes","field":"Medicine","cited_by":35,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université du Québec à Montréal; Douglas Mental Health University Institute; McGill University","funders":"","keywords":"Methadone maintenance; Methadone; Metric (unit); Clinical psychology; Psychology; Scale (ratio); Patient satisfaction; Population; Substance abuse; Customer satisfaction; Medicine; Psychiatry; Nursing; Environmental health; Operations management","retraction":null,"screen_n_in":null,"score":{"opus":0.1097764852465386,"gpt":0.421131667545621,"spread":0.3113551822990824,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0003736423,0.000147977,0.0002596233,0.0001108697,0.0001011958,0.00002574813,0.00001518251,0.00007249381,0.00001387482],"category_scores_gemma":[0.00008996828,0.0001120928,0.00004791165,0.0001452164,0.00005468679,0.0001282621,0.00000496699,0.00007480835,0.000001169623],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001077588,"about_ca_system_score_gemma":0.00003327773,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002607293,"about_ca_topic_score_gemma":0.00001451777,"domain_scores_codex":[0.9989699,0.00006098353,0.0002737808,0.0002309514,0.0002936932,0.0001707441],"domain_scores_gemma":[0.9994302,0.0001060486,0.0001464175,0.0001188243,0.0001032186,0.00009522718],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00006473081,0.00008180988,0.5523172,0.000007378372,0.00003638336,0.000001294054,0.001307355,0.000002434297,0.00006746538,0.00001375631,0.000005343136,0.4460948],"study_design_scores_gemma":[0.004124519,0.002820348,0.9900118,0.0001746001,0.0004739562,0.000004202989,0.000743721,0.001119489,0.0001632745,0.0002136579,0.00005449229,0.00009592147],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9964448,0.0005298884,0.00003508394,0.0001924781,0.00002962072,0.0023237,0.000001491783,0.0001048642,0.0003380902],"genre_scores_gemma":[0.9938683,0.00005283017,0.005546661,0.00003183561,0.00004045792,0.0002120701,0.000116869,0.000008184194,0.0001227997],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4459989,"threshold_uncertainty_score":0.4571013,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2752092069","doi":"10.1016/j.evalprogplan.2017.08.014","title":"Building a community-based culture of evaluation","year":2017,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":34,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Alzheimer Society of Canada; Centre for Community Based Research","funders":"","keywords":"Alliance; Capacity building; Argument (complex analysis); Program evaluation; Sociology; Theory of change; Action research; Conceptual framework; Research program; Public relations; Engineering ethics; Political science; Pedagogy; Engineering; Public administration; Social science; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.4616641335208906,"gpt":0.6253219780203569,"spread":0.1636578444994664,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.03550531,0.000127329,0.0002195801,0.0002014692,0.001336174,0.0006747977,0.0004378841,0.00008870551,0.0004905424],"category_scores_gemma":[0.004687742,0.00009700596,0.00006495624,0.0001945423,0.0001478088,0.000596909,0.0000933128,0.0002502145,0.00001581217],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004699374,"about_ca_system_score_gemma":0.0003234246,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003863371,"about_ca_topic_score_gemma":0.00001928981,"domain_scores_codex":[0.9945277,0.001440086,0.0005525577,0.0002444643,0.003071118,0.0001640566],"domain_scores_gemma":[0.996054,0.0004657337,0.000815312,0.0007926786,0.001792564,0.00007971888],"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":[0.00004179732,0.0001084123,0.0647881,0.00001923242,0.00001801724,2.922887e-7,0.003859647,0.005766472,0.0005247075,0.0005958361,0.000748875,0.9235286],"study_design_scores_gemma":[0.001728658,0.000297654,0.2510177,0.0001452691,0.00008655131,0.000002560253,0.001661795,0.7192286,0.0007979378,0.01389727,0.01098603,0.0001499546],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9703339,0.0004386353,0.003809411,0.0008660518,0.0003418446,0.001339586,0.000007425918,0.00004874543,0.02281435],"genre_scores_gemma":[0.9852898,0.000004751364,0.01418219,0.0001858951,0.00005149415,0.0001535293,0.00004750109,0.000006590521,0.0000782188],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9233786,"threshold_uncertainty_score":0.9999639,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1977960824","doi":"10.1016/j.evalprogplan.2014.04.003","title":"Taking stock of four decades of quantitative research on stakeholder participation and evaluation use: A systematic map","year":2014,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":33,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Ministry of Labour, Employment and Social Solidarity","funders":"","keywords":"Stakeholder; Stock (firearms); Stakeholder engagement; Stakeholder analysis; Environmental planning; Business; Environmental resource management; Geography; Political science; Environmental science; Public relations; Archaeology","retraction":null,"screen_n_in":null,"score":{"opus":0.9237604363882755,"gpt":0.7285006439353664,"spread":0.1952597924529091,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.0759058,0.0003095739,0.001584256,0.001360658,0.0002810878,0.0004212705,0.0002803336,0.0002463586,0.0002159242],"category_scores_gemma":[0.01547435,0.0002150603,0.0001439882,0.001218704,0.0001922987,0.0004648777,0.000090378,0.000385015,0.00002382339],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001319958,"about_ca_system_score_gemma":0.0006768306,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001667602,"about_ca_topic_score_gemma":0.00001818585,"domain_scores_codex":[0.9795019,0.01003712,0.00243372,0.0007033616,0.007007329,0.0003165633],"domain_scores_gemma":[0.9822555,0.01027158,0.003067503,0.0005433288,0.003752592,0.0001094722],"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":[0.00003191809,0.0000818664,0.0007044342,0.02852013,0.0001107636,2.908683e-7,0.002336876,0.0008881291,0.000001155475,0.001386885,0.0001205782,0.965817],"study_design_scores_gemma":[0.001966927,0.003030404,0.01097764,0.2263617,0.003003872,0.0000109884,0.004662837,0.7026354,0.00001146529,0.003387658,0.04319855,0.0007526116],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"empirical","genre_scores_codex":[0.0230713,0.965264,0.000800744,0.00006319011,0.0002683369,0.009801521,0.00002095147,0.00002945721,0.000680513],"genre_scores_gemma":[0.6417393,0.3440747,0.007533412,0.00004231188,0.000150695,0.005765997,0.0003254486,0.00008691047,0.0002812028],"genre_candidate":"review","genre_consensus":null,"teacher_disagreement_score":0.9650643,"threshold_uncertainty_score":0.9928187,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":["metaresearch","bibliometrics"],"domain":"evaluation","study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2074732573","doi":"10.1016/j.evalprogplan.2015.01.004","title":"Application of an organizational evaluation capacity self-assessment instrument to different organizations: Similarities and lessons learned","year":2015,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":32,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"École Nationale d'Administration Publique","funders":"Government of Canada; Australian Government","keywords":"Novelty; Capacity development; Knowledge management; Business; Organization development; Capacity building; Profit (economics); Government (linguistics); Organizational effectiveness; Organizational performance; Marketing; Computer science; Environmental resource management; Psychology; Economics; Economic growth; Social psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.3556781551675651,"gpt":0.5310432671149354,"spread":0.1753651119473704,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01054512,0.0001908545,0.0002701437,0.0003652356,0.0003009219,0.0003708811,0.0002048335,0.00009604044,0.0002735527],"category_scores_gemma":[0.00151796,0.0001596576,0.00002019117,0.001035152,0.00006395359,0.0006254192,0.0001110336,0.000123353,0.00001176944],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002562576,"about_ca_system_score_gemma":0.0007487783,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003044645,"about_ca_topic_score_gemma":0.00005091304,"domain_scores_codex":[0.9942393,0.0008855297,0.0007257551,0.000540494,0.003419181,0.0001897396],"domain_scores_gemma":[0.9948031,0.0002553182,0.0003776959,0.0003585183,0.003919,0.0002863459],"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":[0.00004154332,0.0004674122,0.2528864,0.00002060877,0.0000453937,1.42967e-7,0.01213494,0.02307161,0.0004804356,0.007921414,0.0002454154,0.7026847],"study_design_scores_gemma":[0.001179259,0.0003822554,0.2980562,0.0000247999,0.00008610504,0.000004360952,0.004000443,0.6787487,0.0003245047,0.01567472,0.00134825,0.0001704395],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9774109,0.0001257887,0.01628663,0.003046303,0.0001637558,0.001944923,0.00001912981,0.00009615941,0.0009064699],"genre_scores_gemma":[0.9773716,0.00001644492,0.02147619,0.000218369,0.00006608004,0.0005214966,0.0002927863,0.00001523878,0.00002178447],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7025142,"threshold_uncertainty_score":0.6510648,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4377091894","doi":"10.1016/j.evalprogplan.2023.102318","title":"Mapping the evaluation capacity building landscape: A bibliometric analysis of scholarly communities and themes","year":2023,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":32,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; University of Ottawa","funders":"","keywords":"Scholarship; Bibliometrics; Cover (algebra); Regional science; Sociology; Political science; Library science; Engineering; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.6130985989473278,"gpt":0.5573777674261452,"spread":0.05572083152118257,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","bibliometrics","scholarly_communication"],"consensus_categories":["bibliometrics"],"category_scores_codex":[0.0432756,0.000125935,0.0002820112,0.02843608,0.0006083488,0.001234217,0.000292429,0.00006783288,0.0004574986],"category_scores_gemma":[0.003247499,0.00008109975,0.00007880374,0.09017033,0.0001243323,0.001029177,0.000133635,0.0002004726,0.000008908002],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002677675,"about_ca_system_score_gemma":0.0001201533,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006723902,"about_ca_topic_score_gemma":0.00004733735,"domain_scores_codex":[0.9940853,0.001406105,0.0006152038,0.0002496372,0.003444371,0.0001993941],"domain_scores_gemma":[0.9951001,0.002427261,0.000412694,0.0003706648,0.001627259,0.00006202467],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.000007683697,0.0000163126,0.4084915,0.000007631666,0.0001635411,8.501734e-8,0.0113719,0.006341934,0.00009524641,0.0002165674,0.0002202698,0.5730673],"study_design_scores_gemma":[0.0002512099,0.0000379747,0.500823,0.00002242534,0.0002032012,8.64299e-7,0.01430797,0.481927,0.0000162857,0.001534984,0.0008187661,0.00005633341],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9937729,0.002215658,0.0007172811,0.0007919553,0.000106405,0.0007221942,0.00001034296,0.0000635881,0.001599709],"genre_scores_gemma":[0.9984071,0.0001757785,0.0009892688,0.0001053865,0.00002865961,0.0001683445,0.00007667302,0.00000573126,0.0000430525],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.573011,"threshold_uncertainty_score":0.9998026,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3003697964","doi":"10.1016/j.evalprogplan.2020.101789","title":"Assessing competency-based evaluation course impacts: A mixed methods case study","year":2020,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":31,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University; University of Alberta","funders":"","keywords":"Context (archaeology); Presentation (obstetrics); Process (computing); Knowledge management; Focus group; Medical education; Computer science; Engineering ethics; Engineering; Sociology; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.4928499038396967,"gpt":0.6787705841841661,"spread":0.1859206803444695,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","scholarly_communication","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.04936818,0.0002738693,0.0004115543,0.000249358,0.0006379839,0.001730066,0.0002668687,0.00009758817,0.00150051],"category_scores_gemma":[0.005646829,0.0002162306,0.00009401239,0.001286636,0.00007764383,0.001179296,0.00008169268,0.0002940441,0.00005418792],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00009613822,"about_ca_system_score_gemma":0.001529508,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003022467,"about_ca_topic_score_gemma":0.00004157557,"domain_scores_codex":[0.9873039,0.00605525,0.001119897,0.0008080787,0.00436195,0.0003509213],"domain_scores_gemma":[0.9949549,0.001682679,0.0006327132,0.0004229234,0.001953459,0.0003533471],"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":[0.00003676923,0.0002885021,0.04928688,0.00001049763,0.00003767153,0.00004570014,0.007974488,0.01044823,0.0001257688,0.00002764078,0.0003660563,0.9313518],"study_design_scores_gemma":[0.00263259,0.0007811738,0.06424096,0.00003851803,0.000305328,0.00007323865,0.0576898,0.8723553,0.0000781127,0.0005778493,0.0009797023,0.0002474248],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.959688,0.0008400607,0.03008668,0.003237822,0.0004690084,0.003478701,0.000004500722,0.0001843838,0.002010813],"genre_scores_gemma":[0.9321557,0.000001584999,0.06631465,0.0009062781,0.0001296803,0.0004113376,0.000055681,0.00001782975,0.000007227832],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9311044,"threshold_uncertainty_score":0.9994122,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2965774300","doi":"10.1016/j.evalprogplan.2019.101695","title":"Fostering CHANGE: Lessons from implementing a multi-country, multi-sector nutrition-sensitive agriculture project","year":2019,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Child Nutrition and Water Access","field":"Nursing","cited_by":31,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"Global Affairs Canada","keywords":"Flexibility (engineering); Context (archaeology); Agriculture; Process (computing); Psychological intervention; Popularity; Business; Process management; Sustainability; Computer science; Economic growth; Risk analysis (engineering); Political science; Economics; Medicine; Management","retraction":null,"screen_n_in":null,"score":{"opus":0.1768582222928707,"gpt":0.4447650227706177,"spread":0.267906800477747,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0005422448,0.0002457488,0.0002543514,0.0001570073,0.0003296698,0.0004043691,0.00009633537,0.0001332868,0.00007749769],"category_scores_gemma":[0.00005352305,0.0002208355,0.00006846747,0.0002547646,0.00002609559,0.0004706175,0.0001421617,0.0002441053,0.0000220361],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008101293,"about_ca_system_score_gemma":0.00001529443,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005324806,"about_ca_topic_score_gemma":0.0001874261,"domain_scores_codex":[0.9981354,0.0002270143,0.0003153319,0.0005180564,0.0003960874,0.0004081016],"domain_scores_gemma":[0.9992719,0.0001007484,0.0001851793,0.0001538325,0.0002107803,0.00007759976],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.001169975,0.00257992,0.2111174,0.001060558,0.0003731027,0.00003278783,0.06019867,0.00006078942,0.04551609,0.00004478263,0.003729323,0.6741166],"study_design_scores_gemma":[0.03354156,0.0009034176,0.6308585,0.004533788,0.0005499793,0.00009000685,0.01397365,0.216121,0.02588639,0.0002328961,0.07137194,0.001936871],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.98887,0.001630756,0.0002851399,0.001357226,0.0008889648,0.006007627,0.0002148607,0.0003994512,0.0003459771],"genre_scores_gemma":[0.9879401,0.00002235384,0.008397644,0.0005269932,0.0005761012,0.0006815771,0.001790874,0.00003217097,0.00003216477],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6721797,"threshold_uncertainty_score":0.9005411,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2071135279","doi":"10.1016/j.evalprogplan.2013.06.003","title":"Usability testing, initial implementation, and formative evaluation of an evidence-based intervention: Lessons from a demonstration project to reduce long-term foster care","year":2013,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":30,"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","funders":"National Institutes of Health; University of Wisconsin-Madison","keywords":"Usability; Formative assessment; Psychological intervention; Evidence-based practice; Operationalization; Intervention (counseling); Process management; Psychology; Computer science; Medical education; Applied psychology; Engineering; Nursing; Medicine; Pedagogy; Human–computer interaction","retraction":null,"screen_n_in":null,"score":{"opus":0.7896864516364739,"gpt":0.7360965383819877,"spread":0.05358991325448614,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008035843,0.0001660362,0.0002150682,0.0002372208,0.0006678541,0.0001129122,0.0001113124,0.00009850136,0.000648202],"category_scores_gemma":[0.003013736,0.0001570551,0.00002656603,0.0004704217,0.00009445234,0.001574821,0.00006780621,0.0001944489,0.00001270431],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003039939,"about_ca_system_score_gemma":0.001333262,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001697452,"about_ca_topic_score_gemma":0.001169554,"domain_scores_codex":[0.993853,0.003191004,0.001098083,0.000436303,0.001065155,0.0003564037],"domain_scores_gemma":[0.9945397,0.001383187,0.0006786456,0.0002306446,0.002965962,0.0002019069],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.00004535921,0.00003320993,0.2450245,0.0002045522,0.00000900465,8.018547e-8,0.05125633,0.0000424377,0.0004097591,0.000009476949,0.000124575,0.7028407],"study_design_scores_gemma":[0.002663654,0.0008601608,0.9242556,0.001050597,0.0001098535,0.000001124768,0.04096176,0.02868627,0.0007698127,0.0004355799,0.00003831731,0.000167243],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9854296,0.0001590451,0.00179646,0.001971647,0.0001574851,0.01016689,0.00007473046,0.00006752863,0.0001766445],"genre_scores_gemma":[0.9824431,0.00000211817,0.01118294,0.0009373109,0.0001128251,0.004942334,0.0003656115,0.00001113562,0.000002636488],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7026734,"threshold_uncertainty_score":0.7097357,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2792924106","doi":"10.1016/j.evalprogplan.2018.02.004","title":"Evaluating the implementation of Check & Connect in various school settings: Is intervention fidelity necessarily associated with positive outcomes?","year":2018,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":30,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Fidelity; Intervention (counseling); Dropout (neural networks); Medical education; Psychology; At-risk students; School dropout; Component (thermodynamics); Academic achievement; Mathematics education; Medicine; Computer science; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.0865912338370379,"gpt":0.5070001345973968,"spread":0.4204089007603589,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007713511,0.00009587765,0.0001347872,0.00009003756,0.0004536447,0.0001504625,0.00008935638,0.00005936663,0.0005256128],"category_scores_gemma":[0.001381196,0.00007350239,0.0000310876,0.0004177795,0.0001282429,0.0002291755,0.00002368823,0.0001269718,0.000006930639],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001780147,"about_ca_system_score_gemma":0.0005709041,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002315921,"about_ca_topic_score_gemma":0.004372728,"domain_scores_codex":[0.997703,0.0008227205,0.0003817971,0.0002102365,0.0006812885,0.0002009508],"domain_scores_gemma":[0.9985628,0.0002548563,0.0003872353,0.0000916324,0.0006455138,0.00005798091],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.00003017321,0.00008845013,0.6574418,0.000007099234,0.00007124895,1.088153e-7,0.1558943,0.0000178893,0.00001759984,0.0001817342,0.000341206,0.1859084],"study_design_scores_gemma":[0.001052386,0.0002920987,0.966198,0.000194932,0.0000491685,3.88981e-7,0.03013429,0.0007985187,0.00005170796,0.0008408627,0.0002807548,0.0001068824],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9942203,0.0001076111,0.0001874123,0.002594942,0.0001155798,0.001257226,0.000005583369,0.0000397061,0.00147171],"genre_scores_gemma":[0.9969844,0.000003874459,0.002081224,0.0005884319,0.00004769872,0.0001083437,0.00009568507,0.000006151451,0.00008424642],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3087562,"threshold_uncertainty_score":0.5755092,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3135833119","doi":"10.1016/j.evalprogplan.2021.101929","title":"Ngarraanga Giinganay (‘thinking peacefully’): Co-design and pilot study of a culturally-grounded mindfulness-based stress reduction program with older First Nations Australians","year":2021,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Mindfulness and Compassion Interventions","field":"Psychology","cited_by":28,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Mindfulness; Grounded theory; Psychosocial; Psychological intervention; Focus group; Qualitative research; Anxiety; Psychology; Mental health; Clinical psychology; Medicine; Medical education; Nursing; Psychotherapist; Sociology; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.1166495252526114,"gpt":0.4326570253063607,"spread":0.3160075000537493,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0008758193,0.0002281513,0.0002701562,0.000226169,0.0007791941,0.0002675493,0.0001026704,0.00008032774,0.001205246],"category_scores_gemma":[0.00006832457,0.000194525,0.00004460012,0.0005134293,0.0001040622,0.0001618387,0.00002447386,0.000243317,0.000002929435],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00004054598,"about_ca_system_score_gemma":0.0001007196,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001491857,"about_ca_topic_score_gemma":0.0003942197,"domain_scores_codex":[0.9975991,0.0006772385,0.0004486976,0.0004600743,0.0005548592,0.000260069],"domain_scores_gemma":[0.9986475,0.0001861214,0.0003113769,0.0002536135,0.0004973186,0.0001041278],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.003274953,0.05037924,0.05843783,0.001299687,0.002706267,0.0002092785,0.1139858,0.0215722,0.0002424107,0.01386574,0.008403115,0.7256235],"study_design_scores_gemma":[0.03997723,0.05068416,0.4238876,0.007499168,0.003317468,0.0007496402,0.4128612,0.03682449,0.01204124,0.001728204,0.007568878,0.00286072],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9716275,0.0002467911,0.02326716,0.0003965774,0.0002478401,0.003131467,0.00001435283,0.000173157,0.0008951684],"genre_scores_gemma":[0.9882044,3.514783e-7,0.002959138,0.0000117765,0.0000559654,0.008012811,0.0001837195,0.00002317376,0.0005486922],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7227628,"threshold_uncertainty_score":0.9997078,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1983707297","doi":"10.1016/j.evalprogplan.2012.09.005","title":"Why do women pay more than they should? A mixed methods study of the implementation gap in a policy to subsidize the costs of deliveries in Burkina Faso","year":2012,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Global Maternal and Child Health","field":"Medicine","cited_by":28,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal; Centre Hospitalier de l’Université de Montréal","funders":"Canadian Institutes of Health Research","keywords":"Subsidy; Economics; Business; Public economics; Actuarial science; Environmental health; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.09854096351653664,"gpt":0.5177872288536772,"spread":0.4192462653371406,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002827423,0.00008103668,0.0001765897,0.0001027442,0.00004945019,0.00001198713,0.00006360928,0.00002834391,0.00001645956],"category_scores_gemma":[0.0001380812,0.00004074181,0.00001817953,0.0003172606,0.00002353357,0.00005434224,0.00004944494,0.00009949126,2.216588e-7],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001092758,"about_ca_system_score_gemma":0.00006961139,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003720343,"about_ca_topic_score_gemma":0.0008572323,"domain_scores_codex":[0.9984243,0.0006436487,0.0002855944,0.0000987496,0.0003415765,0.0002061361],"domain_scores_gemma":[0.9995311,0.0000828673,0.0001204909,0.0001226522,0.00008361043,0.00005922131],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001822354,0.0002109122,0.7773063,0.00003833031,0.00001848128,2.896458e-7,0.11091,0.00004079944,0.0002858641,0.00009909549,0.00003140802,0.1108764],"study_design_scores_gemma":[0.001377394,0.0005111017,0.9131128,0.0001383149,0.00003028284,0.000003924036,0.08389423,0.0001340789,0.0002346089,0.0001879202,0.0003348474,0.00004050724],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.995158,0.0002496678,0.00001237868,0.002335706,0.00005284186,0.002074031,0.000004169646,0.000007625581,0.0001055887],"genre_scores_gemma":[0.9985401,0.00001165283,0.0004514401,0.0007077684,0.00003875009,0.0002357563,0.000005772342,0.000005258459,0.00000352024],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1358065,"threshold_uncertainty_score":0.562407,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2026657459","doi":"10.1016/j.evalprogplan.2012.03.006","title":"Improving health equity through theory-informed evaluations: A look at housing first strategies, cross-sectoral health programs, and prostitution policy","year":2012,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Homelessness and Social Issues","field":"Health Professions","cited_by":27,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Michael's Hospital; University of Toronto; McMaster University; Toronto Metropolitan University","funders":"Canadian Institutes of Health Research; Health Canada; Johns Hopkins Bloomberg School of Public Health; Health Promotion Agency; Public Health Agency of Canada","keywords":"Equity (law); Perspective (graphical); Health equity; Inequality; Sociology; Mental health; Equity theory; Mental illness; Variety (cybernetics); Critical theory; Positive economics; Law and economics; Psychology; Public economics; Public relations; Political science; Economics; Health care; Law; Computer science; Psychiatry; Economic Justice","retraction":null,"screen_n_in":null,"score":{"opus":0.254163583276427,"gpt":0.598370269192595,"spread":0.344206685916168,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow","sts"],"consensus_categories":[],"category_scores_codex":[0.009302622,0.0002885007,0.0005186428,0.0001347785,0.004841762,0.00031067,0.00009722989,0.0002423674,0.00008657992],"category_scores_gemma":[0.0007140064,0.0002568318,0.00004948128,0.0003997294,0.0002519662,0.001376325,0.0002294582,0.0004468456,0.00001631545],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001457402,"about_ca_system_score_gemma":0.003813112,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004399706,"about_ca_topic_score_gemma":0.004615206,"domain_scores_codex":[0.9954869,0.001298667,0.0008621934,0.0003853596,0.0007487169,0.001218148],"domain_scores_gemma":[0.9979515,0.0003193891,0.0007786349,0.0002162454,0.0003482716,0.0003859549],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001030934,0.000125771,0.3666534,0.00111748,0.00003027092,2.772319e-7,0.1407991,0.00004279221,0.000004875273,0.0160652,0.00003946053,0.4750183],"study_design_scores_gemma":[0.006777558,0.001695478,0.7267669,0.003335764,0.000140453,0.000007430727,0.2117862,0.01272816,0.00001418498,0.01196825,0.02362139,0.001158264],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9648453,0.01945379,0.002818773,0.001164783,0.0005875786,0.008045687,0.00002581423,0.00052155,0.002536751],"genre_scores_gemma":[0.9936058,0.0003682941,0.003050246,0.0004382938,0.0007483173,0.001207151,0.0004705342,0.00003565017,0.00007571006],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.47386,"threshold_uncertainty_score":0.9999884,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2035078904","doi":"10.1016/j.evalprogplan.2012.03.003","title":"Seeing voices of health disparity: Evaluating arts projects as influence processes","year":2012,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Participatory Visual Research Methods","field":"Social Sciences","cited_by":27,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"St. Michael's Hospital","funders":"","keywords":"Exhibition; Psychological intervention; Narrative; Strengths and weaknesses; Psychology; Observational study; Public relations; Health care; Narrative inquiry; Sociology; Social psychology; Medicine; Political science; Visual arts; Art","retraction":null,"screen_n_in":null,"score":{"opus":0.7688520100965571,"gpt":0.734688925147642,"spread":0.03416308494891518,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03206596,0.0001009245,0.0002002792,0.0001134494,0.0007117568,0.0001232113,0.0001163081,0.00006249417,0.00004664822],"category_scores_gemma":[0.01894999,0.00009509424,0.00002025189,0.0005979534,0.0002317026,0.0007902264,0.00005036209,0.0001500458,0.000006928843],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00008739179,"about_ca_system_score_gemma":0.001488935,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0006543947,"about_ca_topic_score_gemma":0.00008490965,"domain_scores_codex":[0.9947457,0.002629093,0.0003547153,0.0002080791,0.001459084,0.0006033014],"domain_scores_gemma":[0.9979719,0.000812693,0.0003133284,0.0001030312,0.0005485149,0.0002505231],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0000209589,0.0001231878,0.407813,0.0002875396,0.00001701893,2.55542e-7,0.09989427,0.0001404247,0.0001900028,0.0009683344,0.00002498037,0.49052],"study_design_scores_gemma":[0.001843529,0.002307073,0.8264881,0.003269156,0.0001939586,0.000007530516,0.0875677,0.0359552,0.003727834,0.006190164,0.03147892,0.0009708693],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.987488,0.006419761,0.0001960009,0.0004255559,0.00008103764,0.001400136,0.00000105311,0.00009349087,0.003894984],"genre_scores_gemma":[0.9908046,0.00005680289,0.008573182,0.0001018133,0.0001022183,0.0003076218,0.000005786252,0.000008633368,0.00003939349],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4895492,"threshold_uncertainty_score":0.9966918,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2083111387","doi":"10.1016/j.evalprogplan.2011.03.003","title":"Using hybrid models to support the development of organizational evaluation capacity: A case narrative","year":2011,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":26,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"National Research Council Canada","funders":"","keywords":"Work (physics); Narrative; Process management; Program evaluation; Public sector; Internal validity; Knowledge management; Management science; Computer science; Business; Engineering; Political science; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.7330308143647964,"gpt":0.5574659762554713,"spread":0.1755648381093251,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01881071,0.0001447531,0.0001863952,0.0002665419,0.0005804867,0.00014137,0.0001983847,0.0000427903,0.002388567],"category_scores_gemma":[0.0008508524,0.00009832098,0.00003334891,0.0007224936,0.00007100003,0.0005060572,0.00008183911,0.00009939467,0.00002417788],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000105441,"about_ca_system_score_gemma":0.001057443,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002184543,"about_ca_topic_score_gemma":0.00007142494,"domain_scores_codex":[0.9952013,0.0006533752,0.0008209157,0.0003517653,0.0027868,0.0001857706],"domain_scores_gemma":[0.9963752,0.0002275682,0.0003947084,0.000260679,0.002633242,0.000108583],"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":[0.00007015831,0.000166353,0.01011455,0.000009677096,0.00007141515,0.000005276227,0.3669021,0.02788361,0.0002918035,0.001099259,0.0003834312,0.5930024],"study_design_scores_gemma":[0.0006364544,0.0002044773,0.006628038,0.00003822394,0.00008181721,0.0001580963,0.02691053,0.9519537,0.001525133,0.01117237,0.0005079522,0.0001831796],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9571189,0.0000526492,0.03640305,0.0001176911,0.0001828547,0.001601058,0.000007241902,0.0000256233,0.004490922],"genre_scores_gemma":[0.9384346,5.944696e-7,0.06112309,0.0001508219,0.00003294305,0.0001748542,0.0000365796,0.000009047701,0.00003749268],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9240701,"threshold_uncertainty_score":0.9985234,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2058065060","doi":"10.1016/j.evalprogplan.2006.10.005","title":"Beyond the limitations of best practices: How logic analysis helped reinterpret dual diagnosis guidelines","year":2007,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Mental Health and Patient Involvement","field":"Health Professions","cited_by":26,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; Douglas Mental Health University Institute; Université de Montréal; Fonds de Recherche du Québec - Santé; Université du Québec à Montréal","funders":"Canadian Institutes of Health Research","keywords":"Key (lock); Dual (grammatical number); Best practice; Integrated care; Dual diagnosis; Mental health; Service (business); Logic model; Management science; Health care; Computer science; Knowledge management; Process management; Psychology; Sociology; Engineering; Political science; Business; Psychiatry; Computer security","retraction":null,"screen_n_in":null,"score":{"opus":0.7416386789892656,"gpt":0.6116297792655122,"spread":0.1300088997237534,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.007339297,0.0001206965,0.0002070824,0.0001957965,0.0007436726,0.00002505146,0.00008342389,0.000104159,0.00013381],"category_scores_gemma":[0.00190838,0.0000824285,0.00006537004,0.0006011994,0.00005610594,0.0001474723,0.00004805425,0.0002556976,0.00001376196],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006233558,"about_ca_system_score_gemma":0.0001436051,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0003062223,"about_ca_topic_score_gemma":0.0008869768,"domain_scores_codex":[0.9974238,0.0007503831,0.000702959,0.0002207812,0.0005782876,0.0003237982],"domain_scores_gemma":[0.9949307,0.002903495,0.001085986,0.0002055193,0.0007466105,0.0001277087],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002082941,0.000317224,0.6240344,0.0003283156,0.0003535192,0.000002594194,0.02635366,0.0003355633,0.000043468,0.0008676569,0.003396163,0.3437591],"study_design_scores_gemma":[0.005162427,0.004957303,0.3333645,0.002849113,0.00677394,0.000005800371,0.2088846,0.129467,0.0002795411,0.004474948,0.3028351,0.0009456641],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9684432,0.00487444,0.002210381,0.007931796,0.0007521915,0.004366353,0.00006164444,0.00008135543,0.01127864],"genre_scores_gemma":[0.9860849,0.0005303808,0.008092441,0.002905804,0.0001997167,0.0008801055,0.0007728316,0.00001005303,0.0005237052],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3428135,"threshold_uncertainty_score":0.5719804,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1781634349","doi":"10.1016/j.evalprogplan.2015.09.003","title":"Mapping the spatial dimensions of participatory practice: A discussion of context in evaluation","year":2015,"lang":"en","type":"article","venue":"Evaluation and Program Planning","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":26,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Citizen journalism; Participatory GIS; Bridge (graph theory); Context (archaeology); Sociology; Field (mathematics); Relation (database); Set (abstract data type); Politics; Process (computing); Knowledge management; Epistemology; Management science; Computer science; Political science; Engineering; Geography; Data mining","retraction":null,"screen_n_in":null,"score":{"opus":0.571163556037539,"gpt":0.5995651283814842,"spread":0.02840157234394525,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04171402,0.00009722273,0.0002185398,0.0002564744,0.0001002141,0.00006694071,0.0001587071,0.00005806115,0.0001828839],"category_scores_gemma":[0.01047536,0.00004619149,0.00003995435,0.0006686093,0.0001144138,0.0004468706,0.00006910499,0.0001322107,0.00001030072],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00006109064,"about_ca_system_score_gemma":0.0006280105,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009035647,"about_ca_topic_score_gemma":0.00007281646,"domain_scores_codex":[0.9932777,0.002115079,0.0009004275,0.0002451632,0.003309232,0.0001523562],"domain_scores_gemma":[0.9960917,0.001049505,0.0006811364,0.0002932212,0.001805646,0.00007883333],"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":[0.00009411052,0.0001308301,0.09589,0.000005504478,0.00001229479,2.892326e-7,0.02436476,0.009257197,0.0001079237,0.0001581242,0.0004438651,0.8695351],"study_design_scores_gemma":[0.001545605,0.000242586,0.1609367,0.0001270554,0.00005111507,0.000002673816,0.04576763,0.7804164,0.0001126723,0.002662193,0.008055467,0.00007985876],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.985914,0.002262287,0.0014762,0.004375873,0.00038041,0.001976939,0.000003658456,0.00001736807,0.003593253],"genre_scores_gemma":[0.9985191,0.00001184472,0.001041772,0.0001361631,0.00003184359,0.000201033,0.00001679831,0.00000469093,0.0000367331],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8694552,"threshold_uncertainty_score":0.9978598,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4385873038","doi":"10.1016/j.evalprogplan.2023.102354","title":"Characteristics of transformational leadership development programs: A scoping review","year":2023,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":26,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University; University of British Columbia, Okanagan Campus; University of British Columbia","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Checklist; Transformational leadership; Fidelity; Program evaluation; Context (archaeology); Applied psychology; Medical education; Psychology; Transparency (behavior); Baseline (sea); Computer science; Inclusion (mineral); Medicine; Political science; Social psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.9423864761893535,"gpt":0.7666840912630039,"spread":0.1757023849263496,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01248638,0.0002835941,0.001288787,0.0003147335,0.0005439553,0.00002612889,0.000190862,0.0002291041,0.000272914],"category_scores_gemma":[0.001748109,0.0002408274,0.000100377,0.0008566985,0.00007499579,0.0001663727,0.00005604429,0.0005380951,0.0001911384],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002001978,"about_ca_system_score_gemma":0.005945492,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004742465,"about_ca_topic_score_gemma":0.000008813267,"domain_scores_codex":[0.9931451,0.001952972,0.002894702,0.0003474352,0.001052512,0.0006072492],"domain_scores_gemma":[0.9958888,0.001536325,0.001717437,0.0001889564,0.0004749856,0.0001935185],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"systematic_review","study_design_scores_codex":[0.000001075202,0.000007933599,0.0000546653,0.3350689,0.00001907713,2.928695e-7,0.002883412,2.852381e-8,5.830602e-9,0.00001587055,0.0001781205,0.6617706],"study_design_scores_gemma":[0.0001606502,0.0000356666,0.0001370747,0.5252123,0.0001890002,0.000002266195,0.0002682086,0.00003630415,2.718023e-8,0.000003844631,0.4738123,0.0001424018],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.00001752592,0.9821438,0.0001185828,0.0003390156,0.00024748,0.01655458,0.00003685958,0.0001799505,0.000362202],"genre_scores_gemma":[0.000005587378,0.9861132,0.00211257,0.0009342934,0.0001249066,0.009448551,0.001118025,0.00003704854,0.000105825],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.6616282,"threshold_uncertainty_score":0.9996899,"prediction_status":"machine_predicted_unvalidated"},"labels":[{"model":"gemma","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"gpt","categories":[],"domain":null,"study_design":"systematic_review","genre":"review","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"agree"},{"id":"W2904946535","doi":"10.1016/j.evalprogplan.2018.12.003","title":"Towards an assessment for organizational participatory research health partnerships: A systematic mixed studies review with framework synthesis","year":2018,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":26,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Sherbrooke; McGill University Health Centre; McGill University","funders":"","keywords":"General partnership; Participatory action research; Knowledge management; Citizen journalism; Thematic analysis; Sustainability; Action research; Participatory evaluation; Grey literature; Sociology; Psychology; Public relations; MEDLINE; Business; Political science; Qualitative research; Computer science; Social science; Pedagogy","retraction":null,"screen_n_in":null,"score":{"opus":0.9615702679917059,"gpt":0.826293895256178,"spread":0.1352763727355278,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08424394,0.0003941469,0.002223906,0.0003891728,0.0026489,0.00009240596,0.0003381176,0.0002713365,0.0003035871],"category_scores_gemma":[0.03005849,0.0002707437,0.00008452385,0.001457997,0.0002605228,0.000222071,0.0001195075,0.0008697396,0.00005473452],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001140761,"about_ca_system_score_gemma":0.01247058,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001207736,"about_ca_topic_score_gemma":0.00005507705,"domain_scores_codex":[0.967424,0.02609703,0.002426855,0.0008052903,0.002112508,0.001134318],"domain_scores_gemma":[0.9769117,0.01666241,0.001957101,0.0006092329,0.003313819,0.0005457039],"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":[0.000006495594,0.00004139051,0.00008005849,0.739882,0.0001657819,5.00271e-7,0.007849765,3.112029e-7,2.264589e-9,0.001212516,0.009047408,0.2417138],"study_design_scores_gemma":[0.000230195,0.0006421226,0.00004166669,0.5377636,0.0008748829,0.000004776538,0.005880993,0.0002782851,2.996779e-8,0.0004459129,0.4535768,0.0002607021],"study_design_candidate":"systematic_review","study_design_consensus":"systematic_review","genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.000007700074,0.9658883,0.001377578,0.003651064,0.0003180202,0.0283791,0.00008474107,0.0001416651,0.0001518069],"genre_scores_gemma":[0.00005455373,0.8975825,0.03073003,0.00285989,0.0004817777,0.06792769,0.0002412008,0.00007305861,0.00004925872],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.4445294,"threshold_uncertainty_score":0.9999745,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2008748013","doi":"10.1016/j.evalprogplan.2010.11.001","title":"“Show me your impact”: Evaluating transitional justice in contested spaces","year":2010,"lang":"en","type":"review","venue":"Evaluation and Program Planning","topic":"Cambodian History and Society","field":"Social Sciences","cited_by":26,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"International Development Research Centre","funders":"","keywords":"Transitional justice; Economic Justice; Human rights; Politics; Sociology; Political science; Public relations; Public administration; Law","retraction":null,"screen_n_in":null,"score":{"opus":0.2960022377762603,"gpt":0.5569240838286849,"spread":0.2609218460524246,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.007918694,0.0002790096,0.0007017176,0.0001804202,0.0007536099,0.0002895774,0.0001572303,0.0005461039,0.0003746624],"category_scores_gemma":[0.001152955,0.0002603742,0.0002730795,0.000467898,0.000230093,0.0002546964,0.0000165435,0.0008195912,0.00001521859],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003641244,"about_ca_system_score_gemma":0.001638444,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001501627,"about_ca_topic_score_gemma":0.0002533683,"domain_scores_codex":[0.996304,0.001531424,0.0004704011,0.0003994256,0.0009140318,0.0003807397],"domain_scores_gemma":[0.998432,0.0006717281,0.0003383308,0.0001285847,0.0002459867,0.0001834098],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000006581717,0.00004386104,0.000094275,0.0007845965,0.00003112589,0.000003760666,0.0225576,0.00001719296,7.119465e-7,0.000416405,0.0002039189,0.97584],"study_design_scores_gemma":[0.0004897518,0.00009293763,0.0001866713,0.004047948,0.001063514,0.00001074413,0.004969245,0.001809262,2.655672e-8,0.0001516546,0.9867769,0.0004012766],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.001750523,0.9891088,0.0000150093,0.0001007021,0.000545067,0.002232159,0.00003376801,0.0001096474,0.0061043],"genre_scores_gemma":[0.001353166,0.9940362,0.001745585,0.00007750209,0.0009649965,0.0006968097,0.0005898821,0.00004007862,0.0004957682],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.986573,"threshold_uncertainty_score":0.9999849,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}