{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":56,"total_is_capped":false,"direct_labels_cover":0,"predictions_cover":56,"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":"6156bc5eeb8c","filters":{"venue":"Journal of MultiDisciplinary Evaluation"}},"results":[{"id":"W1526225390","doi":"10.56645/jmde.v7i16.314","title":"Empathy Training: Methods, Evaluation Practices, and Validity","year":2011,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Empathy and Medical Education","field":"Medicine","cited_by":141,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Empathy; Psychology; Feeling; Narrative; Social work; Social psychology; Applied psychology; Medical education; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.4604867244391697,"gpt":0.518990695771376,"spread":0.05850397133220631,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02193366,0.0001424714,0.0003393808,0.0002448533,0.0001237625,0.00001535454,0.000079709,0.0001468879,0.0005019744],"category_scores_gemma":[0.01126905,0.000111023,0.000114563,0.0002053519,0.00008700872,0.0005805858,0.0000338309,0.0004027242,0.000006618597],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002117705,"about_ca_system_score_gemma":0.001402692,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001142008,"about_ca_topic_score_gemma":0.000004311882,"domain_scores_codex":[0.9963195,0.001295378,0.0007274853,0.0002113663,0.001274836,0.0001714741],"domain_scores_gemma":[0.9958648,0.0003736347,0.001644388,0.0002057398,0.001611753,0.0002996926],"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":[0.001246358,0.001282665,0.008528979,0.0001217994,0.0001651512,0.00002277171,0.3461845,0.00002094116,0.007435113,0.00004694015,0.0004898257,0.6344549],"study_design_scores_gemma":[0.006387139,0.002906902,0.8042684,0.0004189685,0.003563329,0.001486343,0.1156783,0.04985879,0.005632759,0.009265169,0.0003027741,0.0002311273],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9911806,0.001421194,0.002240027,0.001350752,0.001120258,0.0006918221,0.000001100222,0.00001247101,0.001981804],"genre_scores_gemma":[0.920278,0.0002170273,0.07880034,0.00007185285,0.0005279102,0.00002509563,0.00001944113,0.00001484528,0.00004555407],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7957395,"threshold_uncertainty_score":0.9970595,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1713678554","doi":"10.56645/jmde.v1i1.144","title":"Unpacking the Participatory Process","year":2004,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":116,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Unpacking; Citizen journalism; Process (computing); Participatory action research; Sociology; Computer science; Linguistics; Anthropology; World Wide Web","retraction":null,"screen_n_in":null,"score":{"opus":0.4185578280274691,"gpt":0.5807891987238559,"spread":0.1622313706963868,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02926393,0.0001121422,0.0002086713,0.000244679,0.0003855488,0.0001604463,0.0006156685,0.00005013071,0.0005789404],"category_scores_gemma":[0.002714318,0.000057794,0.0001518386,0.0006487271,0.00009244587,0.001040633,0.00006436052,0.0002526594,0.0001387654],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002293693,"about_ca_system_score_gemma":0.001299178,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000001785504,"about_ca_topic_score_gemma":0.00002441436,"domain_scores_codex":[0.9930677,0.0005869774,0.001180173,0.000174227,0.004807885,0.0001829753],"domain_scores_gemma":[0.9950879,0.0005598975,0.001237641,0.0003490257,0.00266955,0.0000960523],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","study_design_scores_codex":[0.0002061366,0.0003843575,0.0425607,0.00000975351,0.00005739502,0.000009186549,0.03173842,0.8148253,0.002278961,0.001044969,0.0006851172,0.1061997],"study_design_scores_gemma":[0.004535934,0.0007419103,0.5798056,0.0001686388,0.000226649,0.0001467432,0.02624557,0.1564118,0.005251966,0.2249998,0.001193845,0.0002715633],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9832875,0.0002503344,0.004428559,0.009536241,0.0008645875,0.0003006312,0.000001021185,0.00000715259,0.001323989],"genre_scores_gemma":[0.9988037,0.00001722458,0.0006321574,0.0001296857,0.0003247077,0.00001664283,0.00000120677,0.000007606176,0.00006708209],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6584135,"threshold_uncertainty_score":0.999577,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1873333392","doi":"10.56645/jmde.v4i8.30","title":"Culturally Competent Evaluation for Aboriginal Communities: A Review of the Empirical Literature","year":2007,"lang":"en","type":"review","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":56,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"Health Canada","keywords":"Multiculturalism; Context (archaeology); Cultural diversity; Cultural competence; Sociology; Psychology; Social psychology; Public relations; Pedagogy; Political science; Anthropology","retraction":null,"screen_n_in":null,"score":{"opus":0.6594573931484259,"gpt":0.6618411843139824,"spread":0.002383791165556537,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.08787859,0.0004867588,0.002260601,0.0007389896,0.0004117578,0.0001836383,0.001978773,0.0003470955,0.0007540873],"category_scores_gemma":[0.004595303,0.0002360497,0.002144278,0.00201822,0.0001642212,0.0007566405,0.0001861104,0.0009844512,0.00002112597],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009445695,"about_ca_system_score_gemma":0.006304143,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000001782831,"about_ca_topic_score_gemma":0.00002519427,"domain_scores_codex":[0.9774739,0.005859299,0.005416411,0.0003497076,0.01059624,0.0003044811],"domain_scores_gemma":[0.9682752,0.004262854,0.008837668,0.001108295,0.0173806,0.0001353993],"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":[0.00006665403,0.0002392442,0.00009602359,0.01438938,0.0001729143,7.918675e-7,0.003222875,0.0004033358,0.000001256136,0.00006472092,0.009323987,0.9720188],"study_design_scores_gemma":[0.001305111,0.0004767194,0.0008452882,0.1531211,0.003649068,0.0001902408,0.001243416,0.01675828,0.000001846915,0.001270265,0.8208398,0.0002988663],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0003612703,0.9900548,0.0006388404,0.001444583,0.002155505,0.004742222,0.0001063933,0.000005416115,0.0004909953],"genre_scores_gemma":[0.000220923,0.9957064,0.002581603,0.0001957899,0.0005950106,0.0001917464,0.0002752665,0.00002861661,0.0002046842],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.97172,"threshold_uncertainty_score":0.9993292,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1581987850","doi":"10.56645/jmde.v8i17.337","title":"‘Policing Schools’ Strategies: A Review of the Evaluation Evidence","year":2012,"lang":"en","type":"review","venue":"Journal of MultiDisciplinary Evaluation","topic":"Education Discipline and Inequality","field":"Social Sciences","cited_by":44,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"Johns Hopkins University; George Mason University","keywords":"Curriculum; Intervention (counseling); Medical education; Psychology; Scale (ratio); Mathematics education; Medicine; Pedagogy; Psychiatry","retraction":null,"screen_n_in":null,"score":{"opus":0.4899013199200807,"gpt":0.5862737800061536,"spread":0.09637246008607292,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.05452989,0.0003112156,0.001348572,0.0002720801,0.0005256303,0.00007834279,0.001028182,0.0002860788,0.0007516546],"category_scores_gemma":[0.009130977,0.0001965119,0.001042285,0.001283216,0.000190774,0.0017421,0.0001730877,0.0007807203,0.00002784471],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001462122,"about_ca_system_score_gemma":0.01775037,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001473377,"about_ca_topic_score_gemma":0.0001532562,"domain_scores_codex":[0.9835804,0.009011175,0.00265623,0.0002478511,0.004176626,0.0003277347],"domain_scores_gemma":[0.9870535,0.0009984524,0.006715134,0.000638004,0.004415296,0.0001796882],"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":[0.000007782461,0.0002229131,0.0001605054,0.03679248,0.000132638,1.934072e-7,0.008948977,0.00004615753,0.000002844298,0.0004585656,0.001701013,0.9515259],"study_design_scores_gemma":[0.0004726774,0.0001736108,0.002744576,0.6731297,0.01634488,0.00003241017,0.01528239,0.0002012668,0.000003124167,0.002493678,0.2885307,0.0005909664],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0005065457,0.9912782,0.00003661716,0.001806492,0.002008558,0.002750052,0.000008334169,0.000007517787,0.001597643],"genre_scores_gemma":[0.003565629,0.9940354,0.00024136,0.00003507976,0.001877267,0.0001140943,0.00001742715,0.00002454628,0.00008924884],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9509349,"threshold_uncertainty_score":0.9992155,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4384928989","doi":"10.56645/jmde.v19i44.785","title":"Decolonizing Science: Undoing the Colonial and Racist Hegemony of Western Science","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"African cultural and philosophical studies","field":"Social Sciences","cited_by":30,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"Dalhousie University","funders":"Stanford Bio-X; Max-Planck-Gesellschaft; Queen's University; McGill University","keywords":"Undoing; Hegemony; Decolonization; Colonialism; Indigenous; Sociology; Science studies; Epistemology; Postcolonialism (international relations); Social science; Political science; Law; Philosophy; Ecology","retraction":null,"screen_n_in":null,"score":{"opus":0.1054578983877904,"gpt":0.4366287632791224,"spread":0.331170864891332,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["sts"],"consensus_categories":["sts"],"category_scores_codex":[0.012533,0.00006919353,0.0001583494,0.000185682,0.002810765,0.0001007228,0.0004159151,0.00003049218,0.000009319764],"category_scores_gemma":[0.002052449,0.00004089382,0.00004843995,0.00176877,0.003199891,0.0006626009,0.0002576061,0.000141361,0.000004551734],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002213662,"about_ca_system_score_gemma":0.0007991465,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005427994,"about_ca_topic_score_gemma":0.0001854768,"domain_scores_codex":[0.9972481,0.0001932144,0.0003470963,0.0001415664,0.001824871,0.000245219],"domain_scores_gemma":[0.997827,0.0004104946,0.000427042,0.00007745243,0.001153625,0.0001043817],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"observational","study_design_scores_codex":[0.0004525815,0.0002889634,0.271204,0.0000747501,0.00009986405,0.00002981445,0.5484613,0.005432957,0.08343551,0.02033878,0.0004506536,0.06973087],"study_design_scores_gemma":[0.0007779866,0.0003758065,0.9346602,0.0001287282,0.0001041801,0.0000173865,0.02815542,0.008112207,0.0007621924,0.02628308,0.0004707462,0.0001520668],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9838884,0.0003171635,0.00001360401,0.01425801,0.0004058359,0.0002221579,0.000001592693,0.00000957409,0.0008836226],"genre_scores_gemma":[0.999324,0.0001443211,0.0001575026,0.00001526088,0.0003040258,0.000003748782,2.203996e-7,0.000002838307,0.00004807425],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6634563,"threshold_uncertainty_score":0.9995129,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1501592175","doi":"10.56645/jmde.v11i24.422","title":"Utilization Focused Developmental Evaluation: Learning Through Practice","year":2015,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Guelph","funders":"Mastercard Foundation","keywords":"Data collection; Context (archaeology); Curriculum; Intervention (counseling); Sample (material); Citizen journalism; Participatory evaluation; Process (computing); Knowledge management; Computer science; Psychology; Process management; Medical education; Management science; Pedagogy; Sociology; Engineering; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.6300483308941991,"gpt":0.5906089619140568,"spread":0.03943936898014233,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07505964,0.0002055334,0.0003585021,0.0004466075,0.0003693301,0.0002924751,0.0005054672,0.0001214374,0.001525129],"category_scores_gemma":[0.03455161,0.0001541634,0.000148876,0.001161864,0.00006678399,0.004824895,0.0001479755,0.0003848429,0.0003919684],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001259484,"about_ca_system_score_gemma":0.004821962,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000008477923,"about_ca_topic_score_gemma":0.00001992204,"domain_scores_codex":[0.9819041,0.00360681,0.001946343,0.0003551281,0.01194877,0.0002388335],"domain_scores_gemma":[0.9803132,0.001401765,0.002503792,0.000279946,0.01529869,0.0002026479],"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":[0.0009350106,0.0006123038,0.0248473,0.000006944609,0.0001288724,0.00001143702,0.04830743,0.3558161,0.0005688733,0.0004097625,0.009918499,0.5584375],"study_design_scores_gemma":[0.007250878,0.001075007,0.03717446,0.0000788917,0.0003854047,0.0002971171,0.07289588,0.8296646,0.0008106349,0.02095848,0.02908332,0.0003253524],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8673399,0.001570026,0.06135384,0.009704444,0.00325515,0.001420493,0.000002850135,0.0000400949,0.05531318],"genre_scores_gemma":[0.9605368,0.0000604638,0.03832092,0.0001372303,0.0004233122,0.00003074624,0.00002969428,0.00001788668,0.000442998],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5581121,"threshold_uncertainty_score":0.9993876,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2908468726","doi":"10.56645/jmde.v14i31.497","title":"Seeking Culturally Safe Developmental Evaluation: Supporting the Shift in Services for Indigenous Children","year":2018,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"Concordia University","funders":"Health Canada","keywords":"Indigenous; Participatory action research; Mainstream; Psychosocial; Traditional knowledge; Context (archaeology); Intervention (counseling); Citizen journalism; Community-based participatory research; Nursing; Sociology; Psychology; Medical education; Medicine; Political science; Geography; Psychotherapist","retraction":null,"screen_n_in":null,"score":{"opus":0.1359810264428348,"gpt":0.4949917022111761,"spread":0.3590106757683413,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.05631281,0.0001921979,0.0003195554,0.000446208,0.0007906771,0.0003063547,0.0008510069,0.00009414906,0.001016897],"category_scores_gemma":[0.001589511,0.0001127011,0.0001669765,0.0007572734,0.00009061002,0.0014211,0.0001495541,0.0002315027,0.00008264642],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006176612,"about_ca_system_score_gemma":0.001903823,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001985057,"about_ca_topic_score_gemma":0.0009070319,"domain_scores_codex":[0.9915394,0.0009352744,0.002026099,0.0003304815,0.004838228,0.0003304713],"domain_scores_gemma":[0.9930476,0.0008120599,0.002076857,0.0002695645,0.003707135,0.00008680556],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0003829511,0.0002225217,0.6181486,0.00001402192,0.00008863457,0.000001496596,0.09193911,0.01068538,0.001563815,0.00003289183,0.0003164716,0.2766041],"study_design_scores_gemma":[0.00207725,0.0003497708,0.7827318,0.00008175994,0.00008753181,0.00005873963,0.01114584,0.1949206,0.0003367675,0.007876771,0.0001916977,0.0001414892],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9920376,0.0001956463,0.001869359,0.002748803,0.000785093,0.001707217,0.00000703115,0.000006762822,0.0006425186],"genre_scores_gemma":[0.993929,0.00001076485,0.004988817,0.0001567572,0.0007313679,0.00007262408,0.00003665345,0.00001257676,0.00006145283],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2764626,"threshold_uncertainty_score":0.9998963,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1912672976","doi":"10.56645/jmde.v7i16.322","title":"Return on Investment: A Placebo for the Chief Financial Officer… And Other Paradoxes","year":2011,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Intellectual Capital and Performance Analysis","field":"Business, Management and Accounting","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Toronto Metropolitan University; Ministry of Health and Long Term Care","funders":"","keywords":"Return on investment; Officer; Region of interest; Finance; Investment (military); Computer science; Business; Economics; Artificial intelligence; Law; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.07392816003349144,"gpt":0.2925710610319062,"spread":0.2186429009984148,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001691859,0.0001148836,0.0001509068,0.0001966626,0.0002986864,0.00006375655,0.0001400324,0.00004791281,0.0002559519],"category_scores_gemma":[0.0003276754,0.00006665586,0.0001218591,0.0001662593,0.00005203961,0.0005296586,0.00004578142,0.0001246001,0.00002823114],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00002656716,"about_ca_system_score_gemma":0.00003985889,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002767097,"about_ca_topic_score_gemma":0.0000535712,"domain_scores_codex":[0.9990785,0.0000213636,0.000321439,0.0001064279,0.0003472874,0.0001250334],"domain_scores_gemma":[0.9990101,0.0001280642,0.0004441757,0.0001024832,0.0003044296,0.00001076491],"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":[0.07210869,0.005169718,0.1495114,0.001671684,0.002740816,0.00004692707,0.1021566,0.01507807,0.005447957,0.0993208,0.1175186,0.4292287],"study_design_scores_gemma":[0.006325562,0.001239211,0.1711526,0.0003182849,0.001831339,0.00003112224,0.003392514,0.7341368,0.0008953579,0.06090974,0.01921549,0.0005519562],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9930923,0.0004179424,0.0003910697,0.0007167852,0.0003814607,0.000425958,0.000002229218,0.000007327272,0.004564923],"genre_scores_gemma":[0.9974172,0.00002978049,0.000211522,0.0009395789,0.001241277,0.00002467465,0.000003435624,0.00001184259,0.0001207158],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7190587,"threshold_uncertainty_score":0.2802494,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1958376939","doi":"10.56645/jmde.v6i11.207","title":"Debating Professional Designations for Evaluators","year":2009,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Ottawa","funders":"","keywords":"Process (computing); Psychology; Political science; Sociology; Engineering ethics; Pedagogy; Computer science; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.3373365539722523,"gpt":0.5875076878738227,"spread":0.2501711339015704,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03696182,0.0001514877,0.0003143449,0.0006212497,0.0005063645,0.0001390181,0.0004990624,0.00009280973,0.0007632069],"category_scores_gemma":[0.008092914,0.0001029817,0.0002655995,0.0006307475,0.00002872132,0.001268295,0.00003737884,0.0001979082,0.00003998446],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000270023,"about_ca_system_score_gemma":0.001470876,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":3.295748e-7,"about_ca_topic_score_gemma":0.000005111757,"domain_scores_codex":[0.9922615,0.0007935148,0.001694399,0.0002543337,0.004775836,0.0002204122],"domain_scores_gemma":[0.9899909,0.001878916,0.00170744,0.0002814705,0.006003183,0.0001381295],"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":[0.0006416058,0.000892545,0.0109349,0.00001024232,0.00006076342,0.000002253551,0.007500732,0.1208639,0.0187352,0.002744809,0.01302988,0.8245832],"study_design_scores_gemma":[0.002613722,0.001052579,0.1871181,0.00007884251,0.0001166978,0.0000268158,0.002565079,0.6972905,0.001180444,0.1069339,0.0008625199,0.0001608025],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9128692,0.0002436815,0.07229954,0.01047641,0.001832655,0.001146805,0.000007337244,0.00001454526,0.001109783],"genre_scores_gemma":[0.9689445,0.000009200168,0.02987401,0.0001485577,0.0004514765,0.00003439159,0.00001330876,0.000008480579,0.0005159985],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8244224,"threshold_uncertainty_score":0.9916505,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2165257846","doi":"10.56645/jmde.v11i25.433","title":"The \"Usability\" of Evaluation Reports: A Precursor to Evaluation Use in Government Organizations","year":2015,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"École Nationale d'Administration Publique","funders":"Social Sciences and Humanities Research Council; Social Sciences and Humanities Research Council of Canada; Australian Government","keywords":"Summative assessment; Formative assessment; Government (linguistics); Public relations; Usability; Business; Treasury; Data collection; Thematic analysis; Program evaluation; Qualitative research; Psychology; Political science; Public administration; Computer science; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.2916528544849001,"gpt":0.5123901997387377,"spread":0.2207373452538375,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1419134,0.0001956673,0.0004005693,0.000436185,0.00020908,0.0001981594,0.0004622726,0.0001029957,0.0006723349],"category_scores_gemma":[0.09524897,0.0001277272,0.0001335977,0.002311657,0.00008715726,0.001721319,0.0001857103,0.0002259195,0.00004052207],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002074073,"about_ca_system_score_gemma":0.00390064,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003317238,"about_ca_topic_score_gemma":0.0008898177,"domain_scores_codex":[0.9743057,0.004122294,0.003681018,0.0004513661,0.01717171,0.0002678975],"domain_scores_gemma":[0.9732037,0.001821174,0.003840705,0.0009020069,0.02000094,0.0002314259],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","study_design_scores_codex":[0.000824029,0.0009453921,0.3233956,0.000008523831,0.00007034252,0.000004044265,0.02066414,0.4845032,0.00186633,0.000240195,0.003777992,0.1637002],"study_design_scores_gemma":[0.001995044,0.0006626174,0.533599,0.00008255702,0.0002197084,0.00003910429,0.008102494,0.4344434,0.001028272,0.01920165,0.0004841303,0.000142047],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9889758,0.0003132236,0.002855014,0.002423551,0.001532802,0.002504751,0.000008020618,0.00000591409,0.001380877],"genre_scores_gemma":[0.9961736,0.00002338928,0.003318578,0.00003563051,0.0001245781,0.0001376908,0.00001406041,0.00001564332,0.0001568416],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2102033,"threshold_uncertainty_score":0.9123721,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2994567589","doi":"10.56645/jmde.v13i28.462","title":"Translating Project Achievements into Strategic Plans: A Case Study in Utilization-Focused Evaluation","year":2017,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Guelph","funders":"","keywords":"Process management; Program evaluation; Strategic planning; Management science; Engineering management; Business; Political science; Engineering; Public administration; Marketing","retraction":null,"screen_n_in":null,"score":{"opus":0.6383590666042548,"gpt":0.6091607377580164,"spread":0.02919832884623841,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.06284669,0.000247183,0.0004732632,0.001065582,0.001064518,0.0007153057,0.0008495991,0.0001107411,0.0004268962],"category_scores_gemma":[0.003278277,0.000184391,0.0001540883,0.0005755197,0.0000728055,0.002836868,0.0001257332,0.0003680126,0.00002270018],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005173789,"about_ca_system_score_gemma":0.002239915,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002852422,"about_ca_topic_score_gemma":0.009234397,"domain_scores_codex":[0.9872218,0.002664636,0.002483182,0.0004787017,0.006885875,0.0002657479],"domain_scores_gemma":[0.9923338,0.0005140776,0.003023906,0.0008478293,0.003177431,0.0001029982],"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":[0.0004062175,0.001514586,0.5862045,0.00001730269,0.00008119552,0.0002621842,0.05161577,0.0469554,0.0005534938,0.00003244593,0.00004136764,0.3123155],"study_design_scores_gemma":[0.00665824,0.0008977189,0.3139246,0.0000879432,0.0001620864,0.0001654617,0.06109272,0.6107671,0.00003843828,0.006033073,0.000008946207,0.0001637189],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9935336,0.0001159186,0.0009634929,0.0008810454,0.0007878607,0.002518345,0.000006130751,0.000008322869,0.001185297],"genre_scores_gemma":[0.9977452,0.00001399927,0.001876256,0.000012481,0.0001813725,0.000104042,0.00001411817,0.00001658269,0.00003590645],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5638117,"threshold_uncertainty_score":0.9649966,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3094330052","doi":"10.56645/jmde.v16i36.623","title":"Evaluation Policy and Organizational Evaluation Capacity Building: Application of an Ecological Framework across Cultural Contexts","year":2020,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"University of Ottawa; Saudi Arabian Cultural Bureau; National Science Foundation","keywords":"Context (archaeology); Focus group; Conceptual framework; Capacity building; Situated; Government (linguistics); Public relations; Political science; Sociology; Psychology; Environmental resource management; Geography; Business; Social science; Marketing; Economics","retraction":null,"screen_n_in":null,"score":{"opus":0.3125964842906351,"gpt":0.5461170546069485,"spread":0.2335205703163134,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04701893,0.0002522032,0.0005321215,0.0003127015,0.0004143298,0.0002201581,0.0005761835,0.0002551547,0.0007623515],"category_scores_gemma":[0.03256073,0.0001875976,0.0001500656,0.001444777,0.0002169879,0.00210263,0.0001751958,0.0003816976,0.00002480731],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006468546,"about_ca_system_score_gemma":0.001752015,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001188529,"about_ca_topic_score_gemma":0.00003619131,"domain_scores_codex":[0.9842613,0.003063057,0.002098303,0.000553377,0.009757684,0.0002663171],"domain_scores_gemma":[0.9788213,0.001175665,0.002711902,0.0003784494,0.0165883,0.0003244281],"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":[0.0007162151,0.001353619,0.06529303,0.00004580909,0.0001953207,0.000001492493,0.03558611,0.2523138,0.04215214,0.008398825,0.0002263442,0.5937173],"study_design_scores_gemma":[0.001914263,0.0006469739,0.3186539,0.0000269037,0.0002047106,0.0000234316,0.002548242,0.6175789,0.001345822,0.0568529,0.00006047452,0.0001434989],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9705718,0.0002451364,0.02149724,0.005788176,0.0003241833,0.001355097,0.00002491267,0.0000155491,0.0001779271],"genre_scores_gemma":[0.9877003,0.00003241248,0.01109997,0.0002336132,0.0007975702,0.00005377877,0.00006103752,0.0000162391,0.000005018666],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5935737,"threshold_uncertainty_score":0.9812946,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4384932226","doi":"10.56645/jmde.v19i44.809","title":"Steps Toward Evaluation as Decluttering: Learnings from Hawaiian Epistemology","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Indigenous; CLARITY; Negotiation; Epistemology; Identity (music); Field (mathematics); Meaning (existential); Sociology; Social science; Aesthetics; Philosophy; Ecology","retraction":null,"screen_n_in":null,"score":{"opus":0.3574561697276343,"gpt":0.5449069672074174,"spread":0.187450797479783,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.05022797,0.0002430802,0.0005283382,0.001160541,0.0003717786,0.0002566874,0.0008014199,0.0001839803,0.005878124],"category_scores_gemma":[0.00818814,0.0001863408,0.0002938974,0.00139684,0.0000959228,0.001334519,0.0002509413,0.000417603,0.00255157],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005057035,"about_ca_system_score_gemma":0.001506878,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003991497,"about_ca_topic_score_gemma":0.00006203009,"domain_scores_codex":[0.985773,0.002696099,0.002140015,0.0005199294,0.008506575,0.0003643326],"domain_scores_gemma":[0.990879,0.001926594,0.002153918,0.0005576487,0.004262101,0.000220736],"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":[0.0006810221,0.0002687024,0.1015086,0.00001663029,0.0001861791,0.00004756739,0.02249938,0.1822451,0.01362281,0.000207185,0.01131448,0.6674023],"study_design_scores_gemma":[0.002270831,0.0005174312,0.4593824,0.00006568663,0.0001685512,0.00005141639,0.009611815,0.4895169,0.0005067382,0.03566944,0.002052014,0.0001867838],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9813511,0.0002371662,0.002211588,0.0104521,0.0025117,0.0006487091,0.000008714718,0.00004124151,0.002537724],"genre_scores_gemma":[0.9966856,0.00007483596,0.001602147,0.0001323266,0.000580223,0.00004451096,0.00008238774,0.00002425017,0.0007736761],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6672155,"threshold_uncertainty_score":0.998225,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2553254625","doi":"10.56645/jmde.v12i27.454","title":"Debate on the Appropriate Methods for Conducting Impact Evaluation of Programs within the Development Context","year":2016,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Impact evaluation; Context (archaeology); Program evaluation; Computer science; Management science; Intervention (counseling); Impact assessment; Process management; Evaluation methods; Risk analysis (engineering); Psychology; Political science; Business; Engineering; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.6229679264758866,"gpt":0.6137793347535752,"spread":0.009188591722311457,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1990921,0.0001963609,0.0003690485,0.0002604434,0.0004943751,0.0001265749,0.0007112193,0.00007742583,0.0004060477],"category_scores_gemma":[0.01551316,0.0000673592,0.0002935991,0.0004838099,0.0001433777,0.0006335622,0.00008984889,0.0001834682,0.00002290901],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000498205,"about_ca_system_score_gemma":0.002289622,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000002580683,"about_ca_topic_score_gemma":0.00003263156,"domain_scores_codex":[0.9867742,0.00504862,0.002274995,0.0002908922,0.005380057,0.0002312406],"domain_scores_gemma":[0.9812877,0.006353096,0.003743076,0.0005129456,0.008023367,0.00007985104],"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":[0.0003899821,0.0001541862,0.002679575,0.000004838014,0.0001245403,7.330209e-8,0.008254159,0.01041958,0.01144586,0.0003549402,0.0001315567,0.9660407],"study_design_scores_gemma":[0.006761027,0.001955881,0.07345316,0.0004863645,0.0005341875,0.00003306768,0.02276497,0.7646508,0.05936598,0.06918363,0.0005078907,0.0003030727],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9423179,0.0001775909,0.04826767,0.005905836,0.0009008266,0.002194718,0.000003083992,0.000005618062,0.000226741],"genre_scores_gemma":[0.981334,0.000006091768,0.01814369,0.00004163829,0.0001611066,0.0001958684,0.000003687833,0.00001380872,0.0001000537],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9657376,"threshold_uncertainty_score":0.9927796,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2311665238","doi":"10.56645/jmde.v12i26.437","title":"Book Review: Evaluating Communication for Development: A Framework for Social Change","year":2016,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Guelph","funders":"","keywords":"Social change; Sociology; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.5475983103789693,"gpt":0.6087747331216172,"spread":0.0611764227426479,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04425866,0.0001710157,0.0004457327,0.000301519,0.0006991327,0.00008067839,0.0006788276,0.0001192339,0.001220863],"category_scores_gemma":[0.01153881,0.0001025169,0.0002984672,0.0003307385,0.000057832,0.001472148,0.0001104419,0.0001298722,0.00004531691],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003885004,"about_ca_system_score_gemma":0.0008374989,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":2.845505e-7,"about_ca_topic_score_gemma":0.000004964449,"domain_scores_codex":[0.9936478,0.0008373325,0.00193188,0.0002738821,0.003061647,0.0002474873],"domain_scores_gemma":[0.9868597,0.004340711,0.002648437,0.0003854488,0.00567688,0.00008886378],"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":[0.0004022414,0.0001530748,0.000582914,0.0001234686,0.00005592851,1.257594e-7,0.005470809,0.00004656855,0.0006862254,0.0007032807,0.07345148,0.9183239],"study_design_scores_gemma":[0.01290883,0.002293864,0.09692661,0.008351177,0.0008196495,0.00003137333,0.001691704,0.09599756,0.001841305,0.1475078,0.6307495,0.0008805376],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.06497585,0.1292817,0.6400506,0.1540274,0.002048559,0.009282116,0.00005983384,0.00004027705,0.0002336197],"genre_scores_gemma":[0.3089563,0.01563419,0.6481081,0.01361445,0.004150811,0.005398609,0.0001408523,0.0001292783,0.003867397],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.9174433,"threshold_uncertainty_score":0.9996921,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3128096052","doi":"10.56645/jmde.v17i38.673","title":"Evaluation in Our New Normal Environment: Navigating the Challenges with Data Collection","year":2021,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Data collection; Credibility; Best practice; Data quality; Computer science; Quality (philosophy); Public relations; Data science; Risk analysis (engineering); Process management; Business; Political science; Marketing; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.368895379755028,"gpt":0.5129477895526718,"spread":0.1440524097976439,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.05688851,0.0001619282,0.00028346,0.0002100286,0.0003174497,0.0001856446,0.0007352044,0.00007972737,0.0006514844],"category_scores_gemma":[0.003378215,0.00009697505,0.0000729303,0.0007906125,0.00003173723,0.00187518,0.0002749134,0.0005115703,0.00005369027],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004788222,"about_ca_system_score_gemma":0.003363334,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001608769,"about_ca_topic_score_gemma":0.0007146276,"domain_scores_codex":[0.9872029,0.002733161,0.001297559,0.0004312875,0.008126257,0.0002088352],"domain_scores_gemma":[0.994957,0.0006572852,0.001319603,0.0008958564,0.002069267,0.0001009875],"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":[0.0002504083,0.0003263514,0.01924724,0.000007095569,0.00007899556,0.0000144742,0.01033135,0.2357539,0.0009699123,0.0000644438,0.001364916,0.7315909],"study_design_scores_gemma":[0.002972607,0.000306229,0.2702905,0.0001526953,0.0002053019,0.0001595059,0.02223692,0.6982861,0.0003935857,0.00440225,0.0004629114,0.0001313555],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9375741,0.005121533,0.004823931,0.04899565,0.0008673664,0.0008753542,0.00000526074,0.000007447773,0.00172937],"genre_scores_gemma":[0.9935388,0.0006560086,0.005102351,0.00005124969,0.0003981951,0.00001912379,0.00003513123,0.00001237869,0.0001867671],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7314596,"threshold_uncertainty_score":0.9711317,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1750660279","doi":"10.56645/jmde.v8i18.349","title":"An In-Depth International Comparison of Major Donor Agencies: How Do They Systematically Conduct Country Program Evaluation?","year":2012,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"International Development and Aid","field":"Social Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"Japan International Cooperation Agency; World Bank Group; Asian Development Bank; United States Agency for International Development","keywords":"Viewpoints; Christian ministry; Ministry of Foreign Affairs; Program evaluation; Political science; Evaluation methods; International development; Public administration; Economic growth; Public relations; Economics; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.1399939193363626,"gpt":0.4766054579176641,"spread":0.3366115385813015,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01573825,0.0001477466,0.0003509971,0.0003785865,0.0001960672,0.0001346689,0.0004979334,0.0001253623,0.0003250004],"category_scores_gemma":[0.001106356,0.0001218708,0.0001092954,0.0002549715,0.0001222347,0.001814591,0.00004960494,0.0002163378,0.0000104042],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009961278,"about_ca_system_score_gemma":0.001026538,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00005080508,"about_ca_topic_score_gemma":0.0005736573,"domain_scores_codex":[0.9940456,0.001162438,0.001006953,0.0001510714,0.003362624,0.0002713223],"domain_scores_gemma":[0.9956139,0.0002790522,0.001251875,0.0001404022,0.002570519,0.0001441896],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004477153,0.003975624,0.7321331,0.0001884237,0.0004124207,0.000004712704,0.1887736,0.004316584,0.007462785,0.01946788,0.0009290145,0.04188813],"study_design_scores_gemma":[0.00525483,0.0008561339,0.6831034,0.001250657,0.0005466921,0.00002860538,0.1691296,0.1271628,0.001111518,0.009720099,0.001230144,0.0006054542],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9846064,0.0004127884,0.0001754854,0.001655121,0.002170036,0.001556325,0.000009139876,0.00001500475,0.009399649],"genre_scores_gemma":[0.9932857,0.00003055364,0.005773093,0.00001026044,0.000690638,0.00009350322,0.00005357242,0.00001201548,0.00005064803],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1228462,"threshold_uncertainty_score":0.5454596,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4386812841","doi":"10.56645/jmde.v19i45.743","title":"The Evaluation of Equity-Focused Community Coalitions: A Review of the Empirical Literature","year":2023,"lang":"en","type":"review","venue":"Journal of MultiDisciplinary Evaluation","topic":"Community Health and Development","field":"Health Professions","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Victoria","funders":"","keywords":"CLARITY; Framing (construction); Equity (law); Sociology; Scope (computer science); Empirical research; Public relations; Engineering ethics; Management science; Political science; Epistemology; Computer science; Economics; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.7978476605117001,"gpt":0.6764633025978322,"spread":0.1213843579138679,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts","research_integrity"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1164502,0.0003684278,0.001966516,0.0003540794,0.002676689,0.00001738187,0.001527424,0.0005077192,0.000126948],"category_scores_gemma":[0.02250002,0.0001866222,0.000968446,0.001826787,0.0001901709,0.0001833748,0.001408102,0.004740481,0.00002730934],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00173321,"about_ca_system_score_gemma":0.01705914,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003326614,"about_ca_topic_score_gemma":0.0002523731,"domain_scores_codex":[0.9348994,0.05486866,0.005471089,0.0001708387,0.004167553,0.0004223989],"domain_scores_gemma":[0.968234,0.01239948,0.00915282,0.001522044,0.008529922,0.0001617094],"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":[0.00007065956,0.0003087626,0.00006955928,0.202573,0.0002707292,6.543034e-7,0.01009242,0.00001529046,4.992094e-7,0.00008044061,0.01442518,0.7720928],"study_design_scores_gemma":[0.001710647,0.0002872077,0.005029739,0.7580544,0.005019798,0.00003039616,0.00345072,0.0005770497,3.993796e-7,0.006883256,0.2187134,0.0002430365],"study_design_candidate":"systematic_review","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.0005825476,0.9857224,0.00001136223,0.003717171,0.001779816,0.007457183,0.0001211968,0.00001280518,0.0005955286],"genre_scores_gemma":[0.0001564024,0.9982248,0.0001930927,0.0001720501,0.0002755198,0.0005773241,0.0002233757,0.00004387873,0.0001335887],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.7718498,"threshold_uncertainty_score":0.9986217,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2576058593","doi":"10.56645/jmde.v3i4.77","title":"Taking Evaluation Contexts Seriously: A Cross-Cultural Evaluation in Extreme Unpredictability","year":2006,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Psychology; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.3547556918813846,"gpt":0.5431977453023824,"spread":0.1884420534209978,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1176223,0.0003473032,0.0006445102,0.001033514,0.000416268,0.0006114856,0.0007184701,0.0002308734,0.00491274],"category_scores_gemma":[0.01322588,0.0002602323,0.0003437359,0.001518659,0.0002079836,0.004420624,0.0001460237,0.0004993116,0.00009817223],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002314597,"about_ca_system_score_gemma":0.00240866,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001136224,"about_ca_topic_score_gemma":0.00134104,"domain_scores_codex":[0.974858,0.004671799,0.004009826,0.0007250709,0.015264,0.0004712826],"domain_scores_gemma":[0.9777062,0.001159106,0.003914775,0.0006744469,0.01640102,0.0001444683],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0004076407,0.0004851742,0.4260064,0.00001132583,0.00002451409,0.00000426306,0.002561962,0.3383353,0.004636534,0.0001164328,0.0004725136,0.2269379],"study_design_scores_gemma":[0.004011939,0.000177009,0.5174589,0.00007240122,0.0001149259,0.00002918229,0.001077575,0.4588423,0.0002669222,0.01770517,0.00008835511,0.0001552902],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9875534,0.001057614,0.00102566,0.00151826,0.001735023,0.001976166,0.0000110293,0.00002000757,0.005102837],"genre_scores_gemma":[0.9976248,0.00002031656,0.00131242,0.00003936384,0.0006339797,0.0001375618,0.00008782414,0.00002084855,0.00012293],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2267827,"threshold_uncertainty_score":0.999985,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2998328657","doi":"10.56645/jmde.v15i33.539","title":"Using Action Research to Build Evaluation Capacity in Public Health Organizations","year":2019,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"École Nationale d'Administration Publique","funders":"","keywords":"Capacity building; Public health; Action research; Qualitative research; Data collection; Action (physics); Program evaluation; Public relations; Medical education; Psychology; Political science; Sociology; Medicine; Nursing; Public administration; Pedagogy","retraction":null,"screen_n_in":null,"score":{"opus":0.8641173841470653,"gpt":0.654224582662043,"spread":0.2098928014850223,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1507688,0.0001555965,0.0004026162,0.002888237,0.0004248057,0.0003114785,0.0005842885,0.000114285,0.00263456],"category_scores_gemma":[0.01187545,0.0001262624,0.00009136929,0.0054733,0.00004708282,0.002251493,0.0001716484,0.0005575411,0.0004028584],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003213295,"about_ca_system_score_gemma":0.005832805,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00006654676,"about_ca_topic_score_gemma":0.0004910512,"domain_scores_codex":[0.9784869,0.006539054,0.002163814,0.0004874468,0.01183186,0.0004909032],"domain_scores_gemma":[0.9824738,0.000997468,0.001230888,0.0005626535,0.01444561,0.0002895573],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0001396358,0.0006770688,0.2780677,0.00002236143,0.00003021142,9.623441e-7,0.01788604,0.5484121,0.01844817,0.0005066087,0.001343399,0.1344657],"study_design_scores_gemma":[0.001405676,0.0004435903,0.3869733,0.00008482278,0.00001376599,0.000020698,0.007402115,0.5974493,0.0002675965,0.005192472,0.0006340464,0.000112611],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9752389,0.00008772983,0.008164819,0.0127372,0.0013491,0.001709353,0.000004199184,0.000007879612,0.0007008599],"genre_scores_gemma":[0.9921501,0.00002427776,0.007297126,0.0001120133,0.0002421334,0.00002305033,0.00001346829,0.0000173419,0.0001204642],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1388934,"threshold_uncertainty_score":0.9998032,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2945434485","doi":"10.56645/jmde.v15i32.513","title":"Communication for Social Change","year":2019,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Guelph","funders":"","keywords":"Deci-; Capacity building; Public relations; Political science; Knowledge management; Business; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.4924261702304836,"gpt":0.5826940578875175,"spread":0.09026788765703392,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.02288135,0.00009262635,0.000256065,0.0003225436,0.0002391537,0.0001031426,0.0005155214,0.00007341976,0.00120287],"category_scores_gemma":[0.0008582454,0.00006663739,0.0002038042,0.0003280558,0.00002998554,0.001249472,0.00008142013,0.0001376987,0.0001852765],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001688726,"about_ca_system_score_gemma":0.0002919145,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000001834987,"about_ca_topic_score_gemma":0.00001518236,"domain_scores_codex":[0.9956186,0.0005535985,0.0009776225,0.0001531147,0.002565712,0.0001312984],"domain_scores_gemma":[0.9944062,0.0008718073,0.001343109,0.0003106573,0.003017764,0.00005050527],"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":[0.001097595,0.0006816748,0.1120972,0.00004325504,0.000102707,6.485616e-7,0.03861801,0.005888313,0.005906874,0.004386713,0.01127065,0.8199064],"study_design_scores_gemma":[0.003499868,0.0005815742,0.6264889,0.0000466118,0.00008027095,0.00001144348,0.004120118,0.3266606,0.0002390953,0.02952314,0.008601354,0.0001469602],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9856999,0.0002684208,0.002229513,0.008198777,0.0009391094,0.001016665,0.000006775503,0.000006264741,0.001634558],"genre_scores_gemma":[0.9948933,0.00002890338,0.004010875,0.0001083507,0.0003952775,0.00005055002,0.00001751706,0.000008720059,0.0004864889],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8197594,"threshold_uncertainty_score":0.9997101,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2798177048","doi":"10.56645/jmde.v14i30.477","title":"Evaluating Community Development Projects Using the OECD Evaluation Criteria","year":2018,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Community Development and Social Impact","field":"Economics, Econometrics and Finance","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":false},"ca_institutions":"","funders":"International Fund for Agricultural Development; Université Laval; UNICEF","keywords":"Program evaluation; Community development; Relevance (law); Psychological intervention; Sustainability; Strengths and weaknesses; Environmental planning; Intervention (counseling); Business; Environmental resource management; Evaluation methods; Economic growth; Political science; Engineering; Geography; Psychology; Public administration; Environmental science","retraction":null,"screen_n_in":null,"score":{"opus":0.4724031975862495,"gpt":0.4691999126206124,"spread":0.003203284965637099,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.03745093,0.0001776123,0.0003495909,0.0003334831,0.001801316,0.0001569298,0.0004984448,0.0001019133,0.0007296298],"category_scores_gemma":[0.001693804,0.0001510731,0.0001261707,0.0003057992,0.0001436142,0.0007011501,0.0002309402,0.0005104346,0.00008371556],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001072736,"about_ca_system_score_gemma":0.0007961672,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009576427,"about_ca_topic_score_gemma":0.0001281796,"domain_scores_codex":[0.9964399,0.001332989,0.001311738,0.000142242,0.0005157976,0.0002573166],"domain_scores_gemma":[0.9963185,0.0002847472,0.00161057,0.0003826263,0.001330555,0.00007299817],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0007721469,0.001920128,0.08738582,0.0001479464,0.0009826343,0.000003756134,0.5513911,0.005395835,0.01351378,0.001280427,0.001677292,0.3355292],"study_design_scores_gemma":[0.002681748,0.0006704498,0.4635874,0.0001617761,0.0001570654,0.00003994255,0.01060495,0.4967449,0.001282692,0.0233244,0.000351487,0.0003931739],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9932564,0.0005268171,0.001814474,0.0005039283,0.001092268,0.0007271442,0.000004155103,0.00001090665,0.002063854],"genre_scores_gemma":[0.9865964,0.00001251446,0.01279003,0.0000493298,0.0004514519,0.00001898452,0.00001820704,0.00002077497,0.00004228996],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5407861,"threshold_uncertainty_score":0.9994982,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2992830025","doi":"10.56645/jmde.v11i25.406","title":"The Promise of Participatory Evaluation in Family-Centered Rehabilitation Settings: A Qualitative Study","year":2015,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Family and Disability Support Research","field":"Psychology","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Children's Hospital of Eastern Ontario; University of Ottawa","funders":"","keywords":"Empowerment; Thematic analysis; Citizen journalism; Focus group; Participatory action research; Medical education; Qualitative research; Rehabilitation; Logic model; Psychological intervention; Service delivery framework; Psychology; Program evaluation; Nursing; Service (business); Medicine; Sociology; Political science; Social science; Business","retraction":null,"screen_n_in":null,"score":{"opus":0.3205534779548873,"gpt":0.5440565552918495,"spread":0.2235030773369622,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.07007644,0.0001499576,0.0003431702,0.0003048445,0.0001178321,0.0000267326,0.0002806407,0.0000936021,0.00007694009],"category_scores_gemma":[0.008443997,0.0001066832,0.0001317762,0.0004849744,0.0002383241,0.0004049554,0.00007992543,0.0003367334,0.00002110078],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006217187,"about_ca_system_score_gemma":0.0009414421,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009933499,"about_ca_topic_score_gemma":0.0002181393,"domain_scores_codex":[0.9841533,0.01098115,0.001719855,0.0002519366,0.002593987,0.000299729],"domain_scores_gemma":[0.993081,0.002302879,0.000988288,0.0004330659,0.003056583,0.0001382176],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.004280122,0.004524321,0.1629157,0.00003588423,0.0001305207,0.00000539567,0.803929,0.001675237,0.001099987,0.0000720688,0.001665914,0.01966577],"study_design_scores_gemma":[0.005329242,0.00261295,0.3850217,0.00005879575,0.000100857,0.000003266221,0.5949266,0.00792911,0.00003150243,0.003902767,0.00001341965,0.00006972049],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9940596,0.0007666646,0.0000387231,0.00119788,0.0006481562,0.00269041,0.000005501872,0.000006097417,0.0005869243],"genre_scores_gemma":[0.9992683,0.000002867821,0.000247275,0.000008668593,0.00007806356,0.0003355932,0.00001104653,0.00001389541,0.00003429705],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.222106,"threshold_uncertainty_score":0.9999083,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4384929312","doi":"10.56645/jmde.v19i44.773","title":"Decolonizing Evaluation of Indigenous Land-Based Programs: A Settler Perspective on What We Can Learn from the LANDBACK Movement","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Aurora College","funders":"","keywords":"Indigenous; Sovereignty; Perspective (graphical); Land rights; Sociology; Environmental ethics; Political science; Ecology; Law; Ethnology; Computer science; Politics; Artificial intelligence","retraction":null,"screen_n_in":null,"score":{"opus":0.3042104605577486,"gpt":0.5120849863126739,"spread":0.2078745257549252,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03766251,0.0001908737,0.0003557373,0.0005043393,0.00037849,0.0002971232,0.0005608976,0.00009801366,0.0005376642],"category_scores_gemma":[0.002367873,0.000113001,0.0002312158,0.001182141,0.00008085235,0.0007125887,0.00009091457,0.0003128007,0.0001210315],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009722608,"about_ca_system_score_gemma":0.00208512,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00009190908,"about_ca_topic_score_gemma":0.00113473,"domain_scores_codex":[0.9876252,0.002236961,0.001219133,0.0003391587,0.008324793,0.0002547126],"domain_scores_gemma":[0.990465,0.001966243,0.001671911,0.0004852957,0.005317609,0.0000939223],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0005670809,0.0005838952,0.1147579,0.00000895101,0.0001737145,0.000004548927,0.05583861,0.4551748,0.0008007975,0.00004571433,0.00118578,0.3708582],"study_design_scores_gemma":[0.003324469,0.000966818,0.424026,0.0003198223,0.0002533458,0.000001985306,0.05391983,0.4963037,0.0007609843,0.0198304,0.0001551808,0.0001374145],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9741374,0.0006845727,0.0001731863,0.02252723,0.0009213213,0.001351645,0.00001933164,0.00001531527,0.0001699559],"genre_scores_gemma":[0.9986628,0.000214579,0.000436949,0.0001829572,0.0002718756,0.00008191392,0.00006457369,0.000016266,0.00006809078],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3707207,"threshold_uncertainty_score":0.9909289,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1771824113","doi":"10.56645/jmde.v4i7.8","title":"Evaluation, Sustainable Development, and the Environment in the South Pacific","year":2007,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Island Studies and Pacific Affairs","field":"Social Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Nipissing University","funders":"","keywords":"Argument (complex analysis); Sustainable development; Vulnerability (computing); Orthodoxy; Function (biology); Perspective (graphical); Value (mathematics); Empirical research; Regional science; Sociology; Environmental resource management; Environmental planning; Political science; Epistemology; Geography; Computer science; Economics; Law","retraction":null,"screen_n_in":null,"score":{"opus":0.04057001783778182,"gpt":0.3327616416763509,"spread":0.2921916238385691,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.07867575,0.00008018753,0.0001341148,0.00009885913,0.001124451,0.00005767375,0.0001710099,0.00004957567,0.00005385493],"category_scores_gemma":[0.0007115423,0.0000413679,0.00004900445,0.0002027489,0.0002626184,0.0001787634,0.00004256617,0.000169474,0.000004364873],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000334401,"about_ca_system_score_gemma":0.000387388,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003335867,"about_ca_topic_score_gemma":0.0002984438,"domain_scores_codex":[0.9960206,0.001330483,0.0004612346,0.0001011602,0.001836811,0.0002497439],"domain_scores_gemma":[0.9985948,0.0005171713,0.0003481697,0.00009905782,0.0004039132,0.00003685516],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.0002046854,0.00008760146,0.01029119,0.000006746988,0.00004288694,0.00001141864,0.9322935,0.0006954022,0.000003970699,0.003709985,0.000302105,0.05235047],"study_design_scores_gemma":[0.002279957,0.00004950138,0.107618,0.00001620402,0.0001093772,0.00000493556,0.874562,0.001590047,0.000005385938,0.007185429,0.006505874,0.00007328492],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9482102,0.002339573,0.0009025006,0.006369513,0.0002487554,0.001213869,4.24949e-7,0.000003270663,0.04071188],"genre_scores_gemma":[0.9989669,0.0001749355,0.0002936495,0.00001298033,0.0001765615,0.00001912394,0.000001347142,0.000003677907,0.0003508152],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.09732683,"threshold_uncertainty_score":0.9486972,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4362503601","doi":"10.56645/jmde.v18i42.721","title":"Competitive champions versus cooperative advocates","year":2022,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Centre for Advancing Health Outcomes","funders":"","keywords":"Interview; Nonprobability sampling; Enthusiasm; Public relations; Extant taxon; Data collection; Sociology; Psychology; Knowledge management; Social psychology; Political science; Social science; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.2869313256836769,"gpt":0.5289861200793462,"spread":0.2420547943956693,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01995559,0.0001350416,0.0002999738,0.0005233706,0.0008899349,0.0001035482,0.0005844239,0.0000312115,0.009895207],"category_scores_gemma":[0.002173053,0.0001013755,0.0001826887,0.0008998308,0.00007742024,0.000782571,0.0003049395,0.0003978111,0.0001272857],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005111353,"about_ca_system_score_gemma":0.001295025,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000027977,"about_ca_topic_score_gemma":0.00003007126,"domain_scores_codex":[0.9912657,0.001854908,0.001141252,0.000247102,0.005313955,0.0001771312],"domain_scores_gemma":[0.9936605,0.001776162,0.00123521,0.0002953465,0.002920011,0.0001127449],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.004217175,0.001630991,0.01338238,0.000006884534,0.0002843155,0.0000360195,0.04051197,0.7991644,0.006069127,0.0168547,0.01633394,0.1015081],"study_design_scores_gemma":[0.01427381,0.006804482,0.1786221,0.00004458109,0.0002770003,0.0001798008,0.1644834,0.5780171,0.001802838,0.01394556,0.04101679,0.0005326016],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9710362,0.0004134076,0.001252145,0.008196087,0.003905603,0.0005584175,0.00003467546,0.00001261296,0.01459088],"genre_scores_gemma":[0.9979879,0.00002833196,0.001004655,0.00007594457,0.0002270668,0.00005102197,0.00002597764,0.000009452207,0.0005896095],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2211474,"threshold_uncertainty_score":0.9910099,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4362504947","doi":"10.56645/jmde.v3i4.83","title":"American Evaluation Association","year":2006,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Association (psychology); Psychology; Psychotherapist","retraction":null,"screen_n_in":null,"score":{"opus":0.1641309330495719,"gpt":0.5246990005505626,"spread":0.3605680675009907,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.06676973,0.0001657123,0.000396558,0.0007131931,0.0002579569,0.000235857,0.0004474178,0.00008348229,0.001732419],"category_scores_gemma":[0.005915922,0.0001227184,0.000252252,0.00118955,0.00005464983,0.001357868,0.00006503518,0.0002491735,0.0002553151],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001532801,"about_ca_system_score_gemma":0.001145194,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003373453,"about_ca_topic_score_gemma":0.0001676237,"domain_scores_codex":[0.9835074,0.002090052,0.001978865,0.0002929886,0.01187753,0.0002531557],"domain_scores_gemma":[0.9851593,0.001279559,0.004204542,0.0003681362,0.008896633,0.00009180347],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0002062987,0.0005100397,0.4075599,0.000004058996,0.00006624311,0.000002683699,0.001181928,0.1650438,0.005027703,0.000351764,0.02391721,0.3961284],"study_design_scores_gemma":[0.001565642,0.0002873642,0.6307778,0.00001715977,0.0001694239,0.00001317009,0.00109818,0.3424588,0.0005057204,0.02098011,0.002002029,0.0001245541],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9813612,0.0002286333,0.003196411,0.005101102,0.001349181,0.0005237914,0.000004928995,0.00001310646,0.008221703],"genre_scores_gemma":[0.9952995,0.00001932453,0.002866098,0.00005681981,0.0008128215,0.00002462093,0.00002496587,0.00001224191,0.0008836084],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3960038,"threshold_uncertainty_score":0.9991801,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4384928951","doi":"10.56645/jmde.v19i44.779","title":"Decolonizing Evaluation of Indigenous Guidance and Counseling Approaches: A Review of Selected Evaluated Programs","year":2023,"lang":"en","type":"review","venue":"Journal of MultiDisciplinary Evaluation","topic":"Educational and Psychological Assessments","field":"Psychology","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Indigenous; Colonialism; Curriculum; Social connectedness; Medical education; Culturally sensitive; Sociology; Medicine; Psychology; Psychotherapist; Political science; Pedagogy; Social psychology; Ecology","retraction":null,"screen_n_in":null,"score":{"opus":0.4869309888252101,"gpt":0.5408138259197823,"spread":0.0538828370945722,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":[],"category_scores_codex":[0.02989973,0.000443943,0.00216942,0.0005684939,0.0001218088,0.00002039899,0.000465331,0.0004351716,0.0002737427],"category_scores_gemma":[0.001719546,0.0003348546,0.0004724915,0.001827573,0.0001288511,0.0001727719,0.00007975294,0.0006293301,0.00002118957],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0004574411,"about_ca_system_score_gemma":0.002535603,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001011136,"about_ca_topic_score_gemma":0.000006565921,"domain_scores_codex":[0.9881525,0.004697409,0.003515705,0.0005202701,0.002782662,0.0003314386],"domain_scores_gemma":[0.9862227,0.0009258789,0.006828967,0.000431366,0.005464833,0.0001262234],"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":[0.00006601748,0.001087564,0.000359874,0.0395953,0.001069293,0.000003732091,0.003947685,0.00005988104,0.000004489815,0.00001722945,0.0002246847,0.9535642],"study_design_scores_gemma":[0.009043726,0.006330176,0.04593934,0.8379067,0.05910387,0.001386689,0.005348267,0.009280958,0.000005732091,0.005239731,0.01827989,0.002134979],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.02195891,0.9721584,0.00005605439,0.00007130672,0.0009150898,0.004551355,0.00003190659,0.00001738827,0.0002395398],"genre_scores_gemma":[0.002474452,0.9947248,0.001366933,0.000009594484,0.0002516643,0.0006289628,0.0004212605,0.00006235747,0.00006005281],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9514292,"threshold_uncertainty_score":0.9999104,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1906589293","doi":"10.56645/jmde.v8i17.329","title":"Thinking Beyond Measurement, Description and Judgement: Fourth Generation Evaluation in Family-Centered Pediatric Healthcare Organizations","year":2012,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Child and Adolescent Health","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Health care; General partnership; Context (archaeology); Judgement; Ideal (ethics); Engineering ethics; Dialectic; Family centered care; Psychology; Sociology; Management science; Knowledge management; Computer science; Epistemology; Political science; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.2150583790358597,"gpt":0.4294796476801095,"spread":0.2144212686442499,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0208822,0.0002056133,0.0003234858,0.0006306352,0.0009586291,0.00003346418,0.0001314496,0.0001921483,0.00009523784],"category_scores_gemma":[0.001128885,0.0001860204,0.00005609508,0.0006623684,0.00001944837,0.001306812,0.0001001018,0.0006617868,0.0000188688],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001930868,"about_ca_system_score_gemma":0.001319793,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00003842461,"about_ca_topic_score_gemma":0.0003715798,"domain_scores_codex":[0.9927538,0.002690372,0.001513979,0.0002382951,0.00231548,0.0004880639],"domain_scores_gemma":[0.9949704,0.0001078358,0.001353775,0.000194976,0.00313034,0.0002427063],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0001273252,0.0006635919,0.9469299,0.0003095241,0.00002638991,8.630527e-7,0.03219673,0.001298402,0.005370037,0.0002812739,0.001621334,0.01117461],"study_design_scores_gemma":[0.004917541,0.0002007986,0.939411,0.0004490526,0.0002973237,0.00001049303,0.006432308,0.04688677,0.0000291667,0.00110854,0.00006621228,0.000190805],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9830345,0.005523323,0.00335027,0.003050704,0.002755766,0.002125495,0.000008277731,0.00002034783,0.0001313368],"genre_scores_gemma":[0.9939239,0.001094621,0.001620446,0.0003014275,0.002881368,0.00002081074,0.0001136839,0.00003214709,0.00001157109],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.04558837,"threshold_uncertainty_score":0.7585693,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4377141999","doi":"10.56645/jmde.v19i43.825","title":"The Program Evaluation Standards in Evaluation Scholarship and Practice","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Scholarship; Medical education; Systematic review; Resource (disambiguation); Professional association; Political science; Psychology; Medicine; Sociology; Public relations; MEDLINE; Computer science","retraction":null,"screen_n_in":null,"score":{"opus":0.4232894218728951,"gpt":0.6372998949468092,"spread":0.214010473073914,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3790371,0.0001892196,0.0003038074,0.0009145065,0.0006558075,0.0006930978,0.0005136005,0.0001246966,0.0003026288],"category_scores_gemma":[0.09272421,0.0001184993,0.0001200524,0.002301801,0.0001309019,0.003038795,0.0001673053,0.0004701435,0.00007268391],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001071092,"about_ca_system_score_gemma":0.003431386,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004357917,"about_ca_topic_score_gemma":0.0001828681,"domain_scores_codex":[0.955358,0.0125864,0.002989699,0.0006084908,0.02794649,0.0005108871],"domain_scores_gemma":[0.975273,0.006448465,0.001822502,0.0004998627,0.01582194,0.0001342089],"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":[0.0004974358,0.0001567432,0.0136492,0.000005358968,0.00003124994,0.000003138141,0.003079645,0.03653042,0.0003511346,0.0001264793,0.001263711,0.9443055],"study_design_scores_gemma":[0.00246046,0.0004113656,0.4005967,0.0000632298,0.0001604921,0.00004148816,0.007113128,0.5621546,0.00004350179,0.02296677,0.003887026,0.000101243],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9734507,0.001763823,0.0002184273,0.01709392,0.001352734,0.003475526,0.000006582515,0.00002611597,0.002612134],"genre_scores_gemma":[0.9969889,0.000597044,0.001614096,0.00004814541,0.0002328345,0.0003938515,0.00002138741,0.00001541432,0.00008831292],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9442043,"threshold_uncertainty_score":0.9149182,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1756996990","doi":"10.56645/jmde.v6i11.221","title":"Nonmarket Economic Valuation and \"The Economist’s Fallacy\"","year":2009,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Financial Reporting and Valuation Research","field":"Business, Management and Accounting","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Memorial University of Newfoundland","funders":"","keywords":"Nonmarket forces; Fallacy; Economics; Valuation (finance); Positive economics; Neoclassical economics; Public economics; Classical economics; Microeconomics; Factor market; Epistemology; Philosophy","retraction":null,"screen_n_in":null,"score":{"opus":0.04982295590130277,"gpt":0.3348057041719891,"spread":0.2849827482706864,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01278143,0.0001173668,0.0002361471,0.0002652841,0.0002832942,0.0002927418,0.0001682354,0.0000578188,0.0001982502],"category_scores_gemma":[0.001470335,0.00008243191,0.000120276,0.0001228033,0.00007517316,0.001240674,0.00005782425,0.0001789486,0.00008416976],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0001478119,"about_ca_system_score_gemma":0.0002051063,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00008880088,"about_ca_topic_score_gemma":0.0000341387,"domain_scores_codex":[0.9982033,0.0001155279,0.0007928163,0.0001569882,0.0005722521,0.0001590691],"domain_scores_gemma":[0.9976963,0.0001937069,0.00118537,0.0001579752,0.000747387,0.00001922066],"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":[0.002625181,0.0004720805,0.06400419,0.0001110329,0.000161596,0.00001370495,0.00114604,0.04101168,0.001254462,0.02112298,0.01682646,0.8512506],"study_design_scores_gemma":[0.003119979,0.00005038253,0.4361172,0.00003376232,0.00011368,0.00001878349,0.0002028293,0.4970731,0.00001538943,0.0624087,0.0007482909,0.00009789217],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9787222,0.0002221074,0.0003991842,0.01502449,0.0004847032,0.0005641627,3.64181e-7,0.00001149846,0.004571317],"genre_scores_gemma":[0.9980024,0.00003703647,0.0002412381,0.0002599214,0.001278659,0.0000137801,0.000007269086,0.000008634268,0.0001510381],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8511527,"threshold_uncertainty_score":0.4429814,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2611404306","doi":"10.56645/jmde.v13i28.456","title":"The World of Evaluation: Challenges Faced by Student Evaluators","year":2017,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Mathematics education; Psychology; Pedagogy; Sociology","retraction":null,"screen_n_in":null,"score":{"opus":0.3297996531442704,"gpt":0.5830362193061646,"spread":0.2532365661618942,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.09773402,0.0002310224,0.0005094382,0.0003984426,0.001513306,0.000496556,0.001848588,0.00008882129,0.001127014],"category_scores_gemma":[0.008137129,0.0001334005,0.000321104,0.000294611,0.0002503475,0.00147977,0.0003471869,0.0003190308,0.00006917878],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003663333,"about_ca_system_score_gemma":0.00124227,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000004257908,"about_ca_topic_score_gemma":0.0003391746,"domain_scores_codex":[0.9805449,0.002409217,0.002452587,0.0003806772,0.01392201,0.0002906215],"domain_scores_gemma":[0.9825478,0.00208679,0.005642987,0.001482818,0.008081329,0.0001583097],"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":[0.0005314693,0.0007191881,0.04032724,0.00001762804,0.0003071069,0.000002886709,0.008373955,0.01928286,0.003475819,0.0009490901,0.01626585,0.9097469],"study_design_scores_gemma":[0.004553164,0.0007230154,0.7414655,0.0001390468,0.0004082386,0.0000181685,0.01019524,0.2157513,0.002251477,0.01721614,0.007032822,0.0002458856],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9595702,0.006640609,0.0002120935,0.02023276,0.002707904,0.001154166,0.000009270266,0.000006547432,0.009466401],"genre_scores_gemma":[0.9972107,0.0008620375,0.0004192652,0.0000192436,0.0003016235,0.00005439047,0.0000046726,0.00001478894,0.001113269],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.909501,"threshold_uncertainty_score":0.9997866,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4362504882","doi":"10.56645/jmde.v3i5.54","title":"Collaborative Evaluation","year":2006,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.1473091922505518,"gpt":0.5253125317131393,"spread":0.3780033394625875,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.05034616,0.0001677379,0.0003466404,0.0006963146,0.0002847832,0.0002228836,0.0004594266,0.000095017,0.003099008],"category_scores_gemma":[0.003255861,0.0001170806,0.0001845056,0.001309838,0.00007330431,0.001558432,0.00006950708,0.0002044089,0.0002023297],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0005028403,"about_ca_system_score_gemma":0.001957787,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000008795309,"about_ca_topic_score_gemma":0.00009838622,"domain_scores_codex":[0.9861593,0.001916948,0.001908709,0.0002860645,0.009528342,0.0002006361],"domain_scores_gemma":[0.9838074,0.0008641679,0.002148795,0.0003825707,0.01270671,0.00009034848],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0005958613,0.0009724667,0.09187198,0.000009473781,0.00009194232,0.00000995454,0.004453016,0.4817135,0.01149915,0.002191609,0.04291185,0.3636792],"study_design_scores_gemma":[0.002817465,0.000371452,0.4343958,0.00003401565,0.0001784224,0.00003284871,0.003101407,0.4888171,0.00144346,0.06587128,0.002774206,0.0001625174],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9748165,0.0007867049,0.004213102,0.003441431,0.001517223,0.0007939289,0.000009881122,0.00001058169,0.01441065],"genre_scores_gemma":[0.9947186,0.00002051232,0.004053476,0.00003751695,0.0006436118,0.00003147589,0.00003076677,0.00001084593,0.0004531913],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3635167,"threshold_uncertainty_score":0.9978123,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4362503613","doi":"10.56645/jmde.v18i42.711","title":"Empowerment Evaluation of Programs Involving Youth","year":2022,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa; University of Winnipeg","funders":"","keywords":"Empowerment; Participatory evaluation; Program evaluation; Youth empowerment; Citizen journalism; Psychology; Participatory action research; Intervention (counseling); Positive Youth Development; Set (abstract data type); Medical education; Applied psychology; Sociology; Computer science; Political science; Medicine; Social science; Developmental psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.3965048437440594,"gpt":0.526744258074393,"spread":0.1302394143303336,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.1013192,0.0001582994,0.0003915739,0.0007876208,0.0004493439,0.00008838611,0.000715299,0.00004724129,0.003887039],"category_scores_gemma":[0.002197916,0.0001223398,0.0002765446,0.001126126,0.00006567212,0.0008668402,0.0003327267,0.0003268559,0.000017944],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0008745576,"about_ca_system_score_gemma":0.002081149,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.000007280636,"about_ca_topic_score_gemma":0.00001666979,"domain_scores_codex":[0.9751464,0.003704246,0.002368391,0.0003218361,0.01823323,0.000225932],"domain_scores_gemma":[0.9894966,0.0003880274,0.003130117,0.0004783766,0.006398164,0.0001087361],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0005280332,0.001571954,0.09687027,0.00001542009,0.0001221901,0.000002854831,0.04659998,0.4402607,0.004861497,0.0002098708,0.000869175,0.4080881],"study_design_scores_gemma":[0.004299934,0.001835294,0.1652665,0.00006335345,0.0005017038,0.00004003374,0.07099652,0.7423244,0.0006566264,0.01347529,0.0003391278,0.0002012566],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9937835,0.000773647,0.0005276538,0.0008415558,0.001573268,0.0009886493,0.00001191813,0.00000757622,0.001492208],"genre_scores_gemma":[0.9982895,0.00002085254,0.001275581,0.00002720774,0.0001494126,0.0000740598,0.00004154753,0.00001270326,0.0001091751],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4078868,"threshold_uncertainty_score":0.9970235,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2997715932","doi":"10.56645/jmde.v15i33.575","title":"Retrospective Pretest and Counterfactual Self-Report: Different or Same?","year":2019,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Resilience and Mental Health","field":"Psychology","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Counterfactual thinking; Retrospective cohort study; Self-efficacy; Psychology; Research design; Randomized controlled trial; Clinical psychology; Medicine; Social psychology; Statistics; Mathematics; Internal medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.03922127755039199,"gpt":0.4213653749666471,"spread":0.382144097416255,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.001456141,0.0001367923,0.000294485,0.0001262783,0.00009347365,0.00002391148,0.0001063229,0.00008913585,0.000874839],"category_scores_gemma":[0.00007831782,0.00008963793,0.00006289074,0.00009023193,0.00003744771,0.00029016,0.00004882887,0.0002419184,0.00006941667],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0003918505,"about_ca_system_score_gemma":0.000176779,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001527263,"about_ca_topic_score_gemma":0.00002642008,"domain_scores_codex":[0.998072,0.000225408,0.0005774722,0.0002290688,0.0006955269,0.0002004892],"domain_scores_gemma":[0.9986153,0.0001042573,0.0006687692,0.0002079422,0.0002809814,0.0001227704],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.003129766,0.001037278,0.9335409,0.00008998202,0.0002088758,0.000164885,0.03117415,0.00006359115,0.004116251,0.0001140328,0.0007166055,0.02564363],"study_design_scores_gemma":[0.0024208,0.00344515,0.9840168,0.000125714,0.0001128327,0.001531933,0.005686008,0.001826268,0.0002597031,0.0003964134,0.00006259079,0.0001157741],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9957346,0.0004557903,0.00003808448,0.0003311241,0.001593822,0.0006416333,0.000004276913,0.00001186846,0.001188851],"genre_scores_gemma":[0.9981678,0.00007934677,0.0002242984,0.00002853007,0.0002086872,0.00001161106,0.000004882329,0.00001323825,0.001261628],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.05047587,"threshold_uncertainty_score":0.9578874,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1488273786","doi":"10.56645/jmde.v1i1.148","title":"The State of Evaluation in Canada","year":2004,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"Health Canada; Government of Canada; Industry Canada; Australian Government; Transport Canada","keywords":"State (computer science); Political science; Evaluation methods; Public administration; Library science; Regional science; Sociology; Computer science; Engineering; Reliability engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.1579830133115361,"gpt":0.4877877855008924,"spread":0.3298047721893563,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.04439157,0.0001026927,0.00025047,0.0002967471,0.0001582427,0.00005591445,0.0004185003,0.00002578457,0.0001952862],"category_scores_gemma":[0.003561124,0.0000598346,0.00008971518,0.0007500778,0.0000511886,0.0006075352,0.00006206142,0.0001979217,0.00001073646],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001913846,"about_ca_system_score_gemma":0.01388898,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.02857226,"about_ca_topic_score_gemma":0.6946353,"domain_scores_codex":[0.9887437,0.001031384,0.001958453,0.0001640252,0.007918932,0.0001834979],"domain_scores_gemma":[0.9928364,0.001075521,0.001741349,0.0003298453,0.0039459,0.00007096136],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","study_design_scores_codex":[0.0001368439,0.00005556459,0.02352932,0.000002519406,0.00001433397,0.000002524423,0.001808614,0.7336122,0.0006457884,0.00005179408,0.0001707677,0.2399697],"study_design_scores_gemma":[0.002462261,0.0001727616,0.6508724,0.00005782896,0.00003587068,0.00001590383,0.003239116,0.2886246,0.001229324,0.05310217,0.0001148273,0.00007289255],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9934697,0.0005740071,0.00059981,0.003477334,0.0009200249,0.0004627712,0.000003535057,0.000001113402,0.0004917664],"genre_scores_gemma":[0.9992451,0.00009568643,0.0005127703,0.00002486402,0.00005640499,0.00001351868,0.00000264321,0.000005335753,0.00004366056],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.666063,"threshold_uncertainty_score":0.9917014,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4389686508","doi":"10.56645/jmde.v19i46.893","title":"In Plain Sight or Just Plain Obscured?: A Review of Professional Evaluation Associations’ Frameworks for Evaluation Practice Supporting Equity, Diversity and Inclusion (EDI)","year":2023,"lang":"en","type":"review","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Carleton University","funders":"","keywords":"Inclusion (mineral); Equity (law); Professional development; Diversity (politics); Presentation (obstetrics); Psychology; Medical education; Pedagogy; Political science; Medicine","retraction":null,"screen_n_in":null,"score":{"opus":0.5524989813377544,"gpt":0.6559057552488365,"spread":0.1034067739110821,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow","sts","open_science"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.3169106,0.0005686677,0.002625308,0.001908965,0.002522973,0.0001240705,0.001235999,0.0009848885,0.0008842965],"category_scores_gemma":[0.1473193,0.0003912927,0.0007384921,0.002677873,0.0001045675,0.002784206,0.008485941,0.001519112,0.0000191022],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002936787,"about_ca_system_score_gemma":0.009255413,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001481388,"about_ca_topic_score_gemma":0.0001513507,"domain_scores_codex":[0.9509225,0.01442443,0.008913349,0.00112889,0.02393271,0.0006780382],"domain_scores_gemma":[0.9451863,0.0207816,0.01764957,0.0006767996,0.01548409,0.0002216628],"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":[0.0004392272,0.0007479399,0.0002758983,0.01530537,0.0002969877,0.000008203097,0.011349,0.001419414,0.000004191086,0.00007849206,0.005893645,0.9641817],"study_design_scores_gemma":[0.01476706,0.002310761,0.006568525,0.2561881,0.02229696,0.0002106983,0.00989009,0.5765778,0.00000981687,0.03924856,0.07027178,0.001659813],"study_design_candidate":"design_other","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.002355805,0.9734864,0.0009270121,0.005387165,0.004461323,0.01294739,0.000213204,0.00001829028,0.0002034608],"genre_scores_gemma":[0.003915619,0.9905691,0.00282873,0.0001807087,0.0008953361,0.0006753307,0.0006814814,0.00005831666,0.0001953726],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.9625218,"threshold_uncertainty_score":0.9998539,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4386812791","doi":"10.56645/jmde.v19i45.701","title":"Excessive Evaluation Anxiety (XEA): The Last Two Decades","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Problem Solving Skills Development","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Ottawa","funders":"","keywords":"Stakeholder; Scopus; Anxiety; Psychology; Thematic analysis; Intervention (counseling); Inclusion (mineral); Resistance (ecology); Applied psychology; Public relations; Political science; Social psychology; Sociology; MEDLINE; Social science; Qualitative research","retraction":null,"screen_n_in":null,"score":{"opus":0.1194933989239602,"gpt":0.4969720774571387,"spread":0.3774786785331785,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.02463518,0.000216781,0.0003191416,0.0003963199,0.001477754,0.00003104787,0.0004717688,0.0001666247,0.0005267449],"category_scores_gemma":[0.003095615,0.0001412281,0.0001582756,0.0009960826,0.00007867821,0.00056216,0.0003054632,0.000889979,0.0005880278],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001154523,"about_ca_system_score_gemma":0.002336603,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001636096,"about_ca_topic_score_gemma":0.0001987865,"domain_scores_codex":[0.9919956,0.00315332,0.001425207,0.0002741312,0.002650121,0.0005016513],"domain_scores_gemma":[0.9932817,0.001470807,0.001721762,0.0003840661,0.003005368,0.0001363415],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"observational","study_design_gemma":"observational","study_design_scores_codex":[0.0005148977,0.0003889235,0.5318592,0.0002820805,0.0004695313,0.00002546501,0.09974188,0.07872742,0.01318001,0.0003692785,0.07833081,0.1961105],"study_design_scores_gemma":[0.00437182,0.0001502309,0.8505015,0.0007621192,0.0003522522,0.00003074415,0.009574588,0.1212749,0.0002566132,0.01158074,0.0009153279,0.0002291045],"study_design_candidate":"observational","study_design_consensus":"observational","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9847904,0.0006394651,0.0005777911,0.00761742,0.002968833,0.002080013,0.000007327497,0.00006922233,0.001249513],"genre_scores_gemma":[0.9961677,0.0003089611,0.001678338,0.0001512499,0.0009399792,0.0003173664,0.0000581273,0.00003848127,0.0003398429],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3186423,"threshold_uncertainty_score":0.9998222,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4389672983","doi":"10.56645/jmde.v19i46.971","title":"Building Spaces for Dialogues to Rethink Evaluator Competencies: Lessons from the Webinars Organized by the Evaluation Centre for Complex Health Interventions","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Toronto","funders":"","keywords":"Psychological intervention; General partnership; Sustainability; Political science; Public relations; Capacity building; Diversity (politics); Sociology; Engineering ethics; Psychology; Engineering","retraction":null,"screen_n_in":null,"score":{"opus":0.4878210702307514,"gpt":0.5796790978903889,"spread":0.09185802765963741,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08628207,0.0002617324,0.0005913465,0.0003780634,0.001835139,0.0005200599,0.001274733,0.00009390596,0.0005963356],"category_scores_gemma":[0.01733043,0.0001451176,0.000570278,0.001404734,0.0001154299,0.0006531226,0.0002419302,0.0002642117,0.00008261446],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0006926871,"about_ca_system_score_gemma":0.001696581,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00004315579,"about_ca_topic_score_gemma":0.0006291792,"domain_scores_codex":[0.9880425,0.003295754,0.002231157,0.0005143845,0.005463848,0.0004523673],"domain_scores_gemma":[0.9805071,0.01023758,0.002357464,0.0006663407,0.006035125,0.0001964357],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0008063603,0.0003799325,0.003163186,0.00005562135,0.0003204088,3.781619e-7,0.04256512,0.06134799,0.01423991,0.001134653,0.6319903,0.2439962],"study_design_scores_gemma":[0.004473466,0.0009183857,0.07942727,0.0003612,0.0004513563,0.000005364103,0.03314406,0.8088829,0.0002997275,0.05463023,0.01717115,0.0002349179],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5937175,0.0009798642,0.06203971,0.3342324,0.00239981,0.005964953,0.0005957481,0.00004128589,0.00002874977],"genre_scores_gemma":[0.9708983,0.0001259333,0.02588683,0.000873149,0.0007850747,0.0004724551,0.000566364,0.00004302751,0.0003488221],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7475349,"threshold_uncertainty_score":0.9994643,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4362503582","doi":"10.56645/jmde.v18i42.703","title":"How did conservation agriculture go to scale?","year":2022,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Agriculture and Rural Development Research","field":"Agricultural and Biological Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"Business Development Bank of Canada; University of Guelph","funders":"Global Affairs Canada","keywords":"Food security; Conservation agriculture; Livelihood; Agriculture; Sustainable agriculture; Business; Environmental resource management; Promotion (chess); Economic growth; Agricultural economics; Geography; Economics; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.04144172568117667,"gpt":0.2953618935618358,"spread":0.2539201678806591,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00190384,0.0001297042,0.0001899995,0.00004496331,0.000672041,0.00009890234,0.0003433605,0.00005337069,0.0005534075],"category_scores_gemma":[0.0001216192,0.00004337198,0.0001346763,0.0007677621,0.0000179011,0.000361378,0.00024442,0.0003081923,0.00002206818],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002507721,"about_ca_system_score_gemma":0.00005287206,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001624122,"about_ca_topic_score_gemma":0.0001248968,"domain_scores_codex":[0.9972951,0.0003774556,0.0003449436,0.0001891672,0.001574032,0.0002193079],"domain_scores_gemma":[0.9986288,0.0001230108,0.0003270668,0.00004172255,0.0007271616,0.0001522237],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"bench_or_experimental","study_design_gemma":"observational","study_design_scores_codex":[0.000276275,0.0003572962,0.01876238,0.000007859289,0.0000426683,0.00001726892,0.001485488,0.0009625733,0.7739935,0.00005116525,0.1089621,0.09508147],"study_design_scores_gemma":[0.0004513664,0.001226555,0.9393172,0.00002640366,0.00004524794,0.0001441177,0.005487008,0.0005279924,0.003628102,0.0007926307,0.0481335,0.0002198851],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.910901,0.0001935609,0.00002393108,0.08780132,0.0003335602,0.0004787962,0.00001635163,0.00001312072,0.0002383681],"genre_scores_gemma":[0.9962367,0.00002541592,0.0004806354,0.0001934037,0.000489116,0.00005402194,0.0001279145,0.000001054046,0.002391672],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9205548,"threshold_uncertainty_score":0.6059424,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4399921253","doi":"10.56645/jmde.v20i47.999","title":"Canadian Evaluation Society Tribute to Dr. Michael Scriven","year":2024,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Canadian Evaluation Society","funders":"","keywords":"Tribute; Political science; Sociology; Art; Media studies; Law","retraction":null,"screen_n_in":null,"score":{"opus":0.230736999126543,"gpt":0.5334483235436518,"spread":0.3027113244171088,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.05427643,0.0002088202,0.0003443547,0.001093864,0.0003798104,0.0005440928,0.0006238548,0.000135995,0.002784644],"category_scores_gemma":[0.002488826,0.0001539356,0.0004074048,0.00162613,0.00003851164,0.001521578,0.0001001588,0.0003594007,0.0008924501],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002290449,"about_ca_system_score_gemma":0.006930554,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0005520665,"about_ca_topic_score_gemma":0.00963334,"domain_scores_codex":[0.9895504,0.0009548165,0.001537792,0.0004396861,0.007159501,0.0003578263],"domain_scores_gemma":[0.9926752,0.0006729377,0.0004839338,0.0004313539,0.005241096,0.0004954989],"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":[0.000148018,0.0001554688,0.00734065,0.00002897721,0.0002241865,0.00001993143,0.02135684,0.1264891,0.01063368,0.0002729735,0.2782781,0.5550521],"study_design_scores_gemma":[0.0009935123,0.0003858947,0.06821255,0.0001731236,0.0002623599,0.00005694952,0.003541925,0.8697288,0.0004615483,0.01169654,0.04426735,0.0002194188],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8708482,0.002422662,0.01210067,0.1052349,0.005488574,0.001576849,0.00005875041,0.00003106783,0.002238351],"genre_scores_gemma":[0.9932936,0.00006377081,0.003452295,0.0003512015,0.0007438724,0.00005062739,0.00003998998,0.00002133439,0.001983282],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7432398,"threshold_uncertainty_score":0.9998855,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4410799871","doi":"10.56645/jmde.v21i50.1175","title":"From Studies to Systems: Ray Rist’s Influence on Evaluation Systems: Insights from International Research Group for Policy and Program Evaluation (INTEVAL)","year":2025,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université Laval","funders":"","keywords":"Group (periodic table); Psychology; Management science; Political science; Engineering; Chemistry","retraction":null,"screen_n_in":null,"score":{"opus":0.4299539096790813,"gpt":0.6383168529536529,"spread":0.2083629432745717,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","metaepi_narrow"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.06507224,0.0003365275,0.0007191424,0.002806047,0.0006954283,0.0009496718,0.0009502497,0.0002107124,0.00006762541],"category_scores_gemma":[0.04000937,0.0002476193,0.0001795834,0.001649889,0.0001783546,0.001583142,0.0003406169,0.000444273,0.00006182295],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003528444,"about_ca_system_score_gemma":0.00258687,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0002814736,"about_ca_topic_score_gemma":0.0002216032,"domain_scores_codex":[0.9769911,0.00572097,0.00279209,0.0009479204,0.01316647,0.0003815073],"domain_scores_gemma":[0.9627817,0.007645054,0.001584503,0.0007311038,0.02704172,0.0002158873],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00204978,0.0007901948,0.003324804,0.00006589515,0.0007190056,0.000002244634,0.01378081,0.5444557,0.003725868,0.001867459,0.01070039,0.4185179],"study_design_scores_gemma":[0.004607259,0.001269025,0.08456887,0.001414207,0.0003880952,0.000003467392,0.01985047,0.8471175,0.0001530962,0.03522355,0.005176064,0.0002283418],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9689114,0.00546234,0.002932201,0.008068181,0.004967612,0.008732039,0.00007973056,0.0000269398,0.0008195342],"genre_scores_gemma":[0.9924945,0.0002292953,0.002702291,0.0001038793,0.001534197,0.002496963,0.0001384329,0.00002120511,0.0002792605],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.4182895,"threshold_uncertainty_score":0.9999976,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4389686746","doi":"10.56645/jmde.v19i46.881","title":"Glocal Evaluation Competencies for Learning As We Go: Zooming in and zooming out to connect system-level solutions to local beneficiaries","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Royal Roads University; Canadian Evaluation Society","funders":"","keywords":"Glocalization; Perspective (graphical); Beneficiary; Zoom; Knowledge management; Adaptation (eye); Computer science; Psychology; Business; Artificial intelligence; Political science; Engineering; Globalization","retraction":null,"screen_n_in":null,"score":{"opus":0.4029661667108118,"gpt":0.5222759465913903,"spread":0.1193097798805785,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04928083,0.0002256375,0.0005306872,0.001611302,0.0007836455,0.0002892465,0.0003649021,0.0001197905,0.0001024379],"category_scores_gemma":[0.009994521,0.0001865846,0.0001498599,0.00130779,0.00005560159,0.0008626751,0.0002705702,0.0002731018,0.0002349158],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0009541228,"about_ca_system_score_gemma":0.001122034,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002731569,"about_ca_topic_score_gemma":0.001311149,"domain_scores_codex":[0.9920462,0.001135966,0.001779995,0.0004694407,0.00412281,0.0004456161],"domain_scores_gemma":[0.9926208,0.002883974,0.0007890998,0.0002540204,0.003210449,0.0002416547],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0004436224,0.00006876039,0.007457181,0.00004209575,0.00003393807,0.00000386292,0.02358278,0.7740499,0.00532163,0.0006949222,0.000732809,0.1875685],"study_design_scores_gemma":[0.001805765,0.0007556062,0.0825835,0.0004149305,0.00008434526,0.00002110088,0.07144132,0.8381158,0.0001598542,0.003976553,0.0004373041,0.0002039793],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.883814,0.0002803104,0.1038275,0.009085377,0.001113691,0.00162764,0.00001496216,0.00002976219,0.0002066888],"genre_scores_gemma":[0.9898022,0.00003085094,0.009435663,0.00006489865,0.0002265777,0.0001834787,0.00002034195,0.00001996101,0.00021601],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1873645,"threshold_uncertainty_score":0.9983447,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2564563180","doi":"10.56645/jmde.v2i3.107","title":"Evaluation in Canada","year":2005,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"World Bank Group","keywords":"Political science; Business","retraction":null,"screen_n_in":null,"score":{"opus":0.2399987003130568,"gpt":0.5137436117701629,"spread":0.2737449114571061,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0610721,0.0001282495,0.00029265,0.0005663062,0.0001036597,0.00007235225,0.0004630363,0.00004956297,0.003761616],"category_scores_gemma":[0.004276656,0.00009407204,0.00009366534,0.0007518908,0.00002085806,0.001291626,0.00006143937,0.0002444543,0.00008532499],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.003866249,"about_ca_system_score_gemma":0.0134458,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.01573176,"about_ca_topic_score_gemma":0.7698318,"domain_scores_codex":[0.9842972,0.001637574,0.001895086,0.000237549,0.01171859,0.0002140122],"domain_scores_gemma":[0.9929475,0.0007047929,0.001211158,0.0003226838,0.004698169,0.0001156502],"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":[0.00005620988,0.00007294895,0.07212818,0.00000129564,0.00001193035,0.000002267732,0.001019454,0.3630908,0.0003514582,0.00002402104,0.002093354,0.5611481],"study_design_scores_gemma":[0.00149229,0.00006583078,0.4166149,0.00002330288,0.00004536847,0.00001727347,0.001817181,0.5760356,0.0002320701,0.00259726,0.0009834322,0.00007546236],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9874687,0.0005236888,0.0004506533,0.006564054,0.0009107664,0.0004331495,0.000002513935,0.000002436905,0.003644051],"genre_scores_gemma":[0.9976022,0.00003037431,0.001796605,0.0001069756,0.0003287549,0.00001765416,0.000004915612,0.000007198002,0.0001053523],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7541,"threshold_uncertainty_score":0.9999577,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2996945182","doi":"10.56645/jmde.v15i33.609","title":"Big Shoes to Fill: An Evaluation Journey in the Footsteps of Daniel L. Stufflebeam","year":2019,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Athabasca University; Laurentian University","funders":"","keywords":"Context (archaeology); Relevance (law); Foundation (evidence); Intervention (counseling); Program evaluation; Evaluation methods; Sociology; Engineering ethics; Psychology; Engineering; History; Political science; Archaeology; Law","retraction":null,"screen_n_in":null,"score":{"opus":0.3461999936431824,"gpt":0.5442462073245252,"spread":0.1980462136813428,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.08369327,0.0001906544,0.0004594211,0.001063136,0.0001396053,0.0001719831,0.001141122,0.0001024207,0.001276964],"category_scores_gemma":[0.005191211,0.0001122192,0.0001755394,0.001651208,0.00004320336,0.001322057,0.0001218168,0.000333397,0.0001137059],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000289168,"about_ca_system_score_gemma":0.001612876,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00001291002,"about_ca_topic_score_gemma":0.0002162184,"domain_scores_codex":[0.9843395,0.003384002,0.002264303,0.0003593171,0.00940001,0.0002528602],"domain_scores_gemma":[0.9910341,0.001041823,0.001728225,0.0007794129,0.005255705,0.0001607622],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","study_design_scores_codex":[0.0008812541,0.001389428,0.1954548,0.00002980267,0.00006514984,0.000004010657,0.05311293,0.3879632,0.01738889,0.0002937601,0.007051991,0.3363648],"study_design_scores_gemma":[0.002722431,0.001813355,0.6082678,0.0001422284,0.0001068615,0.00003266642,0.02546236,0.3534339,0.0007939024,0.006065798,0.0009931654,0.0001654782],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.987825,0.0001471477,0.00087589,0.005641827,0.001403438,0.001475259,0.000007336686,0.000004071964,0.002620046],"genre_scores_gemma":[0.997468,0.00002209572,0.001705771,0.0002255411,0.0002815925,0.00003487878,0.00001440382,0.00001259593,0.000235154],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.412813,"threshold_uncertainty_score":0.999636,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2527468053","doi":"10.56645/jmde.v2i2.131","title":"An Update on Evaluation in Canada","year":2005,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Political science; Psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.1992817783507851,"gpt":0.5200236305055952,"spread":0.32074185215481,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.03897716,0.0001547093,0.0003000125,0.0006601571,0.000136134,0.000102857,0.0005614682,0.00005729958,0.003248174],"category_scores_gemma":[0.001654454,0.0001127034,0.00008202303,0.0006610126,0.00002128577,0.001651335,0.00003648707,0.0002879539,0.0001294754],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002849146,"about_ca_system_score_gemma":0.007386341,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.006966823,"about_ca_topic_score_gemma":0.5666797,"domain_scores_codex":[0.9874651,0.001742262,0.001677066,0.0002983282,0.008595182,0.0002220054],"domain_scores_gemma":[0.9950003,0.0005228419,0.001167149,0.00046876,0.002679375,0.0001615649],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.0001524915,0.0001462348,0.03861579,0.000001009988,0.000006978877,0.000002487079,0.000752893,0.5586497,0.0004079195,0.00005270478,0.001234282,0.3999775],"study_design_scores_gemma":[0.001528921,0.000217044,0.3936639,0.00002551797,0.00002638599,0.0000100731,0.001437778,0.5999073,0.0003746537,0.002061993,0.000656168,0.0000902962],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9900646,0.0001438624,0.0003364722,0.006594589,0.0008240904,0.0004533252,0.000004663811,0.000003536953,0.001574833],"genre_scores_gemma":[0.9976199,0.00002804698,0.001601378,0.0002949967,0.0003710942,0.00001854734,0.00001615808,0.000009661601,0.00004019902],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5597128,"threshold_uncertainty_score":0.9996459,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4394973106","doi":"10.56645/jmde.v19i45.947","title":"In Memoriam - Chris Coryn","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Legal case studies and regulations","field":"Social Sciences","cited_by":0,"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":"McKnight Foundation; Eli Lilly and Company; Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung; National Science Foundation","keywords":"Sociology; Theology; Psychology; Philosophy","retraction":null,"screen_n_in":null,"score":{"opus":0.07754206848215654,"gpt":0.4367946990376323,"spread":0.3592526305554757,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.005099433,0.00005608864,0.0001364914,0.0002731924,0.0004080674,0.00002403005,0.0001083646,0.0000571907,0.0001825758],"category_scores_gemma":[0.0005915412,0.00004605508,0.00008490763,0.0007667235,0.00008865942,0.0004474036,0.00004395127,0.0001258823,0.00005028136],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002466193,"about_ca_system_score_gemma":0.0003496116,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0001268952,"about_ca_topic_score_gemma":0.0008915971,"domain_scores_codex":[0.9982042,0.0003294649,0.0003572155,0.0000810548,0.0008587423,0.0001693189],"domain_scores_gemma":[0.9990414,0.000183903,0.0002360955,0.00006970592,0.0004070723,0.00006180604],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"observational","study_design_scores_codex":[0.0005361068,0.0009147971,0.06250708,0.00007725529,0.0002394622,0.0005003427,0.5125057,0.08101654,0.002920415,0.01330233,0.2724781,0.05300189],"study_design_scores_gemma":[0.003784999,0.000371201,0.7770542,0.0001755101,0.0001629013,0.00002326866,0.07454868,0.03720091,0.00008267902,0.0256639,0.0806011,0.0003306226],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9696665,0.0003467751,0.00003400172,0.01773501,0.004347224,0.0002938142,0.000002283372,0.00001818968,0.007556183],"genre_scores_gemma":[0.9965906,0.0003294167,0.0002112631,0.00001625568,0.001783685,0.000007376172,0.000002874845,0.000005907833,0.001052671],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7145472,"threshold_uncertainty_score":0.3138565,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1544634347","doi":"10.56645/jmde.v3i4.95","title":"Review of Research Evaluation, Volumes 13(3), 14(1), and 14(2)","year":2006,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Psychology","retraction":null,"screen_n_in":null,"score":{"opus":0.365610475907359,"gpt":0.5976328797105598,"spread":0.2320224038032008,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.1401359,0.000163858,0.0005315281,0.0009528412,0.0003183579,0.0001145004,0.0005345612,0.0001011873,0.002121838],"category_scores_gemma":[0.01099615,0.0001156701,0.0001875697,0.001437846,0.0002340323,0.001130983,0.000182311,0.0003697567,0.00006601198],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.0002758199,"about_ca_system_score_gemma":0.001560359,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.0000236253,"about_ca_topic_score_gemma":0.0000894363,"domain_scores_codex":[0.9809058,0.003595121,0.002648067,0.0003668861,0.01221111,0.0002729598],"domain_scores_gemma":[0.9750879,0.00175574,0.001882059,0.0005359554,0.02061916,0.0001192301],"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":[0.0004422955,0.001093293,0.1404426,0.001383985,0.0001176799,0.000006996122,0.00297252,0.01709108,0.004698203,0.0006506827,0.181977,0.6491237],"study_design_scores_gemma":[0.003752837,0.0009793263,0.6121509,0.004977843,0.0003664153,0.0001047852,0.00193567,0.2788985,0.0009194429,0.08759868,0.008037434,0.0002781913],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9085251,0.0732238,0.0007892241,0.008825188,0.0009153711,0.001488719,0.00001025228,0.00000898341,0.006213333],"genre_scores_gemma":[0.9885162,0.007312647,0.002882691,0.00008622125,0.0004283751,0.00004112217,0.00002108718,0.00001437337,0.0006972747],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.6488455,"threshold_uncertainty_score":0.9987904,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4389686269","doi":"10.56645/jmde.v19i46.969","title":"Rethinking Evaluator Competencies in an Age of Discontinuity – Implications of Inequities, Sustainability and the Pandemic for Training Evaluators: An Introduction to a Special Volume","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Toronto","funders":"","keywords":"Pandemic; Sustainability; Coronavirus disease 2019 (COVID-19); Sociology; Psychology; Political science; Engineering ethics; Medicine; Engineering; Biology; Ecology","retraction":null,"screen_n_in":null,"score":{"opus":0.2642530573592116,"gpt":0.5106165978136467,"spread":0.246363540454435,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08981225,0.0001582579,0.0006476411,0.0008272323,0.0002692367,0.00009795136,0.0005534426,0.00008795942,0.0001193417],"category_scores_gemma":[0.0132547,0.0001056192,0.0001587714,0.001269496,0.000392522,0.001340584,0.0001629185,0.000226892,0.000001112112],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.000389264,"about_ca_system_score_gemma":0.001505302,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00002425869,"about_ca_topic_score_gemma":0.001304685,"domain_scores_codex":[0.9921766,0.002341077,0.002280471,0.0003908828,0.002559356,0.0002515939],"domain_scores_gemma":[0.991218,0.001786584,0.001615805,0.0005604773,0.004703206,0.0001159633],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"observational","study_design_scores_codex":[0.003147109,0.0004273737,0.2799111,0.0001037758,0.00006027241,8.894615e-7,0.3393214,0.1099198,0.0119981,0.005804996,0.0004692042,0.2488359],"study_design_scores_gemma":[0.002420173,0.0007929274,0.5547088,0.00003908398,0.00008877806,0.000006553846,0.05065106,0.2192802,0.00005515017,0.1717734,0.00009238653,0.0000914781],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9851533,0.0000502769,0.003282748,0.00918938,0.0004988657,0.00176508,0.00002802904,0.000009902108,0.00002245699],"genre_scores_gemma":[0.9956896,0.00001899465,0.003059539,0.00002009767,0.0009980089,0.0001132053,0.00004177975,0.00001140465,0.00004733119],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.2886704,"threshold_uncertainty_score":0.9950571,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4377142049","doi":"10.56645/jmde.v19i43.837","title":"The Integration of the Program Evaluation Standards into an Evaluation Toolkit for a Transformative Model of Care for Mental Health Service Delivery","year":2023,"lang":"en","type":"article","venue":"Journal of MultiDisciplinary Evaluation","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":0,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Centre for Advancing Health Outcomes; University of British Columbia","funders":"","keywords":"Transformative learning; Mental health; Process management; Process (computing); Health care; Logic model; Service delivery framework; Computer science; Service (business); Knowledge management; Psychology; Business; Sociology; Pedagogy; Political science","retraction":null,"screen_n_in":null,"score":{"opus":0.5826331380281115,"gpt":0.6905983737903973,"spread":0.1079652357622859,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"codex-gemma-dda1882f352a","candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.05977577,0.0001774696,0.0003914659,0.0002986143,0.002300326,0.000022602,0.0004385331,0.0001231439,0.00002011808],"category_scores_gemma":[0.003660395,0.0001116567,0.0001987368,0.0008951135,0.0001009223,0.001017149,0.00006867678,0.0002827493,0.000001171243],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002661077,"about_ca_system_score_gemma":0.01338348,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00008886348,"about_ca_topic_score_gemma":0.004339368,"domain_scores_codex":[0.9888519,0.003991684,0.002568502,0.0002340179,0.003912329,0.0004415601],"domain_scores_gemma":[0.9762647,0.001649095,0.00328739,0.000339314,0.01832924,0.0001303025],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"simulation_or_modeling","study_design_scores_codex":[0.00160954,0.0001264576,0.0003579667,0.001149311,0.0000562486,9.808001e-9,0.4733142,0.06926936,0.006028236,0.0002359238,0.001919235,0.4459335],"study_design_scores_gemma":[0.004729926,0.001562636,0.00687673,0.0004885338,0.0001869026,6.464368e-7,0.1252444,0.8551877,0.0009209247,0.004304712,0.0004101928,0.00008670457],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9328139,0.0001857558,0.01574043,0.02379157,0.001000433,0.02556633,0.0008395668,0.00002181803,0.00004017299],"genre_scores_gemma":[0.9861566,0.0001081171,0.009111408,0.0002716679,0.0001721217,0.003802707,0.0003387779,0.00002575025,0.00001282367],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.7859183,"threshold_uncertainty_score":0.9989985,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}