{"meta":{"query_hash":"586ff0674e7a","filters":{"venue":"Evaluation"},"cohort_total":46,"direct_labels_cover":2,"predictions_cover":46,"exported":46,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/586ff0674e7a","api":"https://metacan.xera.ac/api/v1/cohort?venue=Evaluation"},"results":[{"id":"W1972177361","doi":"10.1177/1356389008095487","title":"A Transaction Cost-Based Approach to Partnership Performance Evaluation","year":2008,"lang":"en","type":"article","venue":"Evaluation","topic":"Banking stability, regulation, efficiency","field":"Economics, Econometrics and Finance","cited_by":58,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Safety Canada","funders":"","keywords":"General partnership; Transaction cost; Corporate governance; Key (lock); Process management; Business; Database transaction; Limited partnership; Audit; Industrial organization; Risk analysis (engineering); Computer science; Accounting; Computer security; Finance","score_opus":0.12777864542751377,"score_gpt":0.29956379061594945,"score_spread":0.17178514518843568,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1972177361","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8676813,0.00027261625,0.10836761,0.0002913707,0.00028120072,0.0019755827,0.000009267884,0.00007238308,0.021048648],"genre_scores_gemma":[0.9966395,0.0000075242842,0.0021685925,0.00012960237,0.00010107781,0.0006704689,0.00013894244,0.000021675873,0.00012261528],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99810874,0.00014178359,0.0005738424,0.0005453437,0.0003716303,0.0002586755],"domain_scores_gemma":[0.99886274,0.00005145515,0.00024809007,0.00043023442,0.00033523675,0.00007222453],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0057524946,0.00015075546,0.00021396443,0.00032948016,0.0002838127,0.00003792688,0.00014807444,0.00010813716,0.00053273264],"category_scores_gemma":[0.00045071635,0.00018546048,0.0000797864,0.00068504876,0.00005313155,0.00036689243,0.0000068395166,0.00010443004,0.0004262632],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009352007,0.00061206147,0.08576047,0.00005526858,0.000023342798,9.2672494e-8,0.004250935,0.83713883,0.0000845227,0.0069225193,0.0005376242,0.0645208],"study_design_scores_gemma":[0.000720458,0.00005966373,0.34952644,0.0000063823986,0.00001466842,0.0000016299521,0.00001760095,0.64723617,0.00019916796,0.0011914985,0.0008699911,0.000156321],"about_ca_topic_score_codex":0.00003819433,"about_ca_topic_score_gemma":0.000009654508,"teacher_disagreement_score":0.26376596,"about_ca_system_score_codex":0.0007171135,"about_ca_system_score_gemma":0.00017827305,"threshold_uncertainty_score":0.75628585},"labels":[],"label_agreement":null},{"id":"W2014366393","doi":"10.1177/1356389003094006","title":"Evaluability Assessment: A Tool for Incorporating Evaluation in Social Change Programmes","year":2003,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":61,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Calgary","funders":"","keywords":"Behaviour change; Social change; Theory of change; Process management; Political science; Engineering ethics; Management science; Sociology; Public relations; Psychology; Engineering; Psychological intervention","score_opus":0.4539931329436821,"score_gpt":0.5881362452183663,"score_spread":0.13414311227468417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2014366393","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9666018,0.00014199883,0.008847391,0.0023480614,0.0009464109,0.009916979,0.000018141756,0.000057110727,0.011122104],"genre_scores_gemma":[0.9837788,0.0000022154602,0.008671657,0.00028212252,0.00028671673,0.0067166216,0.00014313804,0.000017798486,0.00010098233],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9859606,0.0043205954,0.0015095995,0.00087435526,0.006894502,0.0004403371],"domain_scores_gemma":[0.9941632,0.0010792287,0.0007343861,0.0004998351,0.0034611674,0.00006217799],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.11808656,0.00022942794,0.0003616562,0.00047265718,0.00048393814,0.0003664751,0.0003139536,0.00016008824,0.0024369794],"category_scores_gemma":[0.013360029,0.00019872378,0.00016141747,0.0014786106,0.00007076483,0.0013693722,0.00004091121,0.00018983333,0.000081349084],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005073347,0.00034721973,0.14789762,0.000021944626,0.000016173582,1.5099911e-7,0.0032805898,0.00228102,0.00026125592,0.020239754,0.00075291377,0.8248506],"study_design_scores_gemma":[0.0026764646,0.000147888,0.36203307,0.000015061472,0.000064256164,8.96538e-7,0.00092909427,0.4717394,0.00010413649,0.15933119,0.00274651,0.00021201151],"about_ca_topic_score_codex":0.000022984923,"about_ca_topic_score_gemma":0.00032365287,"teacher_disagreement_score":0.8246386,"about_ca_system_score_codex":0.0009796759,"about_ca_system_score_gemma":0.0018175447,"threshold_uncertainty_score":0.99847496},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"design_other","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"},{"model":"opus","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"low"}],"label_agreement":"split"},{"id":"W2026507234","doi":"10.1177/1356389009360471","title":"Peace and Conflict Impact Assessment (PCIA) in Community Development: A Case Study from Mozambique","year":2010,"lang":"en","type":"article","venue":"Evaluation","topic":"Environmental and Social Impact Assessments","field":"Environmental Science","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McGill University","funders":"","keywords":"Situational ethics; Government (linguistics); Work (physics); Quality (philosophy); Public relations; Political science; Public administration; Sociology; Business; Engineering; Law","score_opus":0.04812356254503167,"score_gpt":0.4044966418320075,"score_spread":0.3563730792869758,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2026507234","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9967102,0.000011588945,0.000042330947,0.00004036263,0.000082071674,0.0009058603,0.000005028745,0.00001650614,0.0021860776],"genre_scores_gemma":[0.9986909,0.0000054809575,0.001049251,0.000058671427,0.000020953368,0.000085378495,0.000033700166,0.0000109958055,0.00004469132],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99800646,0.00084488647,0.00022285031,0.00019551639,0.0005315369,0.00019871845],"domain_scores_gemma":[0.99942726,0.00012211176,0.00008571443,0.00023084755,0.000007853375,0.0001261959],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.0027481962,0.00013769756,0.00013912935,0.000036027533,0.0004245664,0.00007280808,0.00011497051,0.000080411795,0.0018758907],"category_scores_gemma":[0.00007668365,0.0001332421,0.000020664114,0.00010993112,0.00010129615,0.0003665232,0.00019003294,0.0004949886,0.00003992237],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000008219318,0.00069359393,0.9360495,0.0000010595998,0.000016467886,0.000040803152,0.033815444,0.000088417,0.005771766,0.0000016173027,0.00002255956,0.02349054],"study_design_scores_gemma":[0.0010709569,0.00010210094,0.9849606,0.0000047531826,0.000022360808,0.000016251352,0.011492932,0.0018274206,0.00013740144,0.00012016785,0.00009767673,0.00014739108],"about_ca_topic_score_codex":0.054682385,"about_ca_topic_score_gemma":0.09561539,"teacher_disagreement_score":0.048911072,"about_ca_system_score_codex":0.0007275356,"about_ca_system_score_gemma":0.00007283349,"threshold_uncertainty_score":0.99903655},"labels":[],"label_agreement":null},{"id":"W2036342338","doi":"10.1177/1356389003009002006","title":"Reducing Anxiety and Resistance in Policy and Programme Evaluations","year":2003,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Carleton University; Simon Fraser University","funders":"","keywords":"Resistance (ecology); Anxiety; Process (computing); Psychology; Political science; Public relations; Social psychology; Computer science; Ecology; Biology","score_opus":0.22745225598283272,"score_gpt":0.5341188055677645,"score_spread":0.3066665495849318,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2036342338","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.927227,0.0010863282,0.0010385077,0.004147313,0.00014918143,0.0009285228,0.0000025934112,0.00002116411,0.06539937],"genre_scores_gemma":[0.99166626,0.00011829271,0.006782691,0.00018928727,0.0000475874,0.00012254261,0.0000064512683,0.000006336754,0.0010605699],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9959349,0.000977838,0.00053481443,0.00043378127,0.0018923517,0.00022629551],"domain_scores_gemma":[0.9984055,0.0004563653,0.00017686235,0.00031201093,0.0005606918,0.00008853835],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.023249589,0.00010334754,0.00014634828,0.00039065175,0.00018826147,0.0002326783,0.00009531732,0.00004665985,0.00046502936],"category_scores_gemma":[0.011726447,0.00008681909,0.000021624217,0.0010704515,0.000068462,0.0005294187,0.000026549782,0.00009917357,0.00003563167],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000030684703,0.00010939215,0.18504576,0.000014617118,0.0000105489435,3.8667824e-7,0.007324254,0.0013976722,0.000811402,0.11043094,0.0011023094,0.693722],"study_design_scores_gemma":[0.0017183833,0.000046485035,0.6733787,0.00004436907,0.000026264246,0.000007946299,0.0016144798,0.06652263,0.00025334777,0.23145357,0.024710108,0.00022369546],"about_ca_topic_score_codex":0.00005519082,"about_ca_topic_score_gemma":0.0011273858,"teacher_disagreement_score":0.6934983,"about_ca_system_score_codex":0.00018341784,"about_ca_system_score_gemma":0.0006529009,"threshold_uncertainty_score":0.9965982},"labels":[],"label_agreement":null},{"id":"W2052280380","doi":"10.1177/1356389012461192","title":"The ethical sensitivity of evaluators: A qualitative study using a vignette design","year":2012,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":21,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université Laval","funders":"Government of Canada","keywords":"Vignette; Context (archaeology); Psychology; Qualitative research; Quality (philosophy); Social psychology; Engineering ethics; Sociology; Epistemology; Social science","score_opus":0.6117213930755443,"score_gpt":0.6475090385779075,"score_spread":0.03578764550236324,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2052280380","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9256657,0.00016237266,0.07076379,0.0009906868,0.00050821685,0.0014389176,0.0000026990028,0.000013440878,0.00045423085],"genre_scores_gemma":[0.99742955,0.0000028832744,0.0021185281,0.00016852886,0.00013904557,0.000068692025,0.000002031403,0.000008013604,0.00006275527],"study_design_codex":"qualitative","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9546191,0.035233486,0.0011530494,0.0003522681,0.00827651,0.00036556702],"domain_scores_gemma":[0.9850491,0.011283144,0.0008137304,0.0005737742,0.0021894267,0.000090818816],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.264315,0.00013162196,0.00024042251,0.00015138397,0.00043682638,0.00011775502,0.0002201448,0.0001013989,0.00053209614],"category_scores_gemma":[0.020995993,0.00007855737,0.00008456729,0.0007916494,0.00013649817,0.000511099,0.00007839686,0.0002643565,0.00016195008],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00054340134,0.0013875549,0.034663957,0.000008863947,0.00025465735,8.335158e-7,0.6717294,0.033953384,0.014572494,0.004389659,0.0052730865,0.2332227],"study_design_scores_gemma":[0.0009668425,0.00029974763,0.07285091,0.000014271665,0.00016026449,0.0000035077499,0.10266712,0.81044865,0.0012905382,0.010842168,0.00029554914,0.00016043612],"about_ca_topic_score_codex":0.000051071587,"about_ca_topic_score_gemma":0.000060513295,"teacher_disagreement_score":0.7764953,"about_ca_system_score_codex":0.00018336471,"about_ca_system_score_gemma":0.00060078927,"threshold_uncertainty_score":0.98725057},"labels":[],"label_agreement":null},{"id":"W2081322617","doi":"10.1177/1356389005058475","title":"Polity, Politics and Policy Evaluation in Belgium","year":2005,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":50,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Polity; Parliament; Politics; Institutionalisation; Political science; Public administration; Corporate governance; Government (linguistics); Public policy; Phenomenon; Political economy; Comparative politics; Sociology; Economics; Law","score_opus":0.2900194241285808,"score_gpt":0.5822176039709627,"score_spread":0.2921981798423819,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2081322617","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.93253773,0.00023006277,0.0003182405,0.025774797,0.00023384675,0.00081376993,0.000006123882,0.000022901566,0.040062524],"genre_scores_gemma":[0.9966677,0.000028637007,0.00070711726,0.00120021,0.00047842346,0.000089654204,0.000029642202,0.000008666857,0.0007899657],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9928079,0.0010785771,0.0007931291,0.00043253944,0.0045818794,0.00030595172],"domain_scores_gemma":[0.9975584,0.00041874152,0.00022975243,0.0004467783,0.0012302184,0.00011608843],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.02595729,0.0001358167,0.00018663077,0.00084719766,0.00015463437,0.00021234101,0.00023368275,0.0000913504,0.0032186336],"category_scores_gemma":[0.005280739,0.00011607637,0.000042125273,0.0010144262,0.000063985666,0.00089945574,0.000082722785,0.00013036659,0.0006668479],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000014485392,0.00012323749,0.06395793,0.0000040253535,0.0000076192473,1.6569616e-7,0.0028277996,0.01213827,0.00038513597,0.031798914,0.002217567,0.88652486],"study_design_scores_gemma":[0.0010006714,0.00003545527,0.44723678,0.00000777492,0.000017563163,0.0000028971408,0.0003449154,0.45643333,0.00015790593,0.08832588,0.0063394806,0.00009735498],"about_ca_topic_score_codex":0.00017726302,"about_ca_topic_score_gemma":0.0007824812,"teacher_disagreement_score":0.8864275,"about_ca_system_score_codex":0.000746948,"about_ca_system_score_gemma":0.0012906517,"threshold_uncertainty_score":0.9976926},"labels":[],"label_agreement":null},{"id":"W2093062308","doi":"10.1177/1356389012442445","title":"A socio-political framework for evaluability assessment of participatory evaluations of partnerships: Making sense of the power differentials in programs that involve the state and civil society","year":2012,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Ottawa","funders":"","keywords":"General partnership; Citizen journalism; Participatory evaluation; Politics; Public relations; Civil society; Participatory action research; Power (physics); State (computer science); Political science; Public administration; Sociology; Computer science","score_opus":0.5679951806189625,"score_gpt":0.5888862518149395,"score_spread":0.020891071195976996,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2093062308","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.985726,0.00019095182,0.009451217,0.0017517401,0.00030689852,0.002307969,0.000027476279,0.0000050911863,0.00023265649],"genre_scores_gemma":[0.996661,0.0000067858236,0.0026688022,0.00014159708,0.000036528752,0.00045059464,0.0000073354904,0.000009496822,0.000017823168],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99113566,0.0030960995,0.0012758508,0.00030565483,0.003780151,0.00040659108],"domain_scores_gemma":[0.9924256,0.0043313988,0.0009870699,0.0007152928,0.001469915,0.00007070656],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.047197003,0.00015518868,0.0003871417,0.00007841457,0.00016912995,0.00004924174,0.0002737185,0.00011591106,0.0004916904],"category_scores_gemma":[0.005120374,0.000087166256,0.00025866248,0.0004930697,0.00052802195,0.00033359128,0.00014065468,0.00019942789,0.0000017277173],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000039265975,0.0004748185,0.9385109,0.000071084694,0.00006529143,7.200268e-9,0.019447044,0.0006740426,0.0004842955,0.03284596,0.00009421614,0.0072930665],"study_design_scores_gemma":[0.0006107047,0.00010268879,0.7568241,0.0001066052,0.00015584391,2.9394604e-7,0.008235553,0.10450143,0.0009003061,0.12844557,0.000035672394,0.000081244514],"about_ca_topic_score_codex":0.0000116285655,"about_ca_topic_score_gemma":0.00011421087,"teacher_disagreement_score":0.18168682,"about_ca_system_score_codex":0.0002152437,"about_ca_system_score_gemma":0.0007585249,"threshold_uncertainty_score":0.98111117},"labels":[],"label_agreement":null},{"id":"W2111087525","doi":"10.1177/1356389009360478","title":"Conceptual and Methodological Challenges in Producing Research Syntheses for Decision-and Policy-Making: An Illustrative Case in Primary Healthcare","year":2010,"lang":"en","type":"article","venue":"Evaluation","topic":"Primary Care and Health Outcomes","field":"Health Professions","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal; Université de Sherbrooke","funders":"Canadian Institutes of Health Research; Canada Research Chairs; Ministère de la Santé et des Services sociaux; Canadian Health Services Research Foundation","keywords":"Rigour; Operationalization; Viewpoints; Management science; Health care; Engineering ethics; Foundation (evidence); Process (computing); Deliberation; Knowledge management; Computer science; Process management; Political science; Engineering; Epistemology; Politics","score_opus":0.7415406961974293,"score_gpt":0.6603945259424397,"score_spread":0.0811461702549896,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2111087525","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.964657,0.002166956,0.00006289258,0.027871966,0.0002511682,0.0034030974,0.000021082264,0.00002522038,0.0015406164],"genre_scores_gemma":[0.9873283,0.000546186,0.009213011,0.0014876333,0.00043707775,0.0009383432,0.000011526133,0.000016452048,0.000021481968],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99226505,0.0055377022,0.0005980605,0.000578089,0.00044379296,0.00057729083],"domain_scores_gemma":[0.9754418,0.023406064,0.00013814743,0.0002952589,0.0005699374,0.00014877379],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.029985141,0.00013282328,0.00038302064,0.0004798286,0.0005963564,0.0000134480715,0.000092558286,0.000353697,0.00007371111],"category_scores_gemma":[0.038897146,0.00010969538,0.000017955608,0.00025991106,0.00020131047,0.00027299253,0.00012795914,0.0012029347,0.000003421886],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007931867,0.00013132926,0.035290286,0.0007643816,0.000008784073,0.000051749168,0.11886171,0.000009012369,0.0006129691,0.052314155,0.00041432923,0.7907481],"study_design_scores_gemma":[0.002004216,0.0009883186,0.84978604,0.00080942997,0.000018446182,0.00011198581,0.07729019,0.0014342659,0.000029124649,0.06383831,0.003362771,0.0003269044],"about_ca_topic_score_codex":0.000929682,"about_ca_topic_score_gemma":0.030046659,"teacher_disagreement_score":0.81449574,"about_ca_system_score_codex":0.0006084354,"about_ca_system_score_gemma":0.0026924126,"threshold_uncertainty_score":0.99883443},"labels":[],"label_agreement":null},{"id":"W2111340440","doi":"10.1177/1356389009105883","title":"How Legitimate and Justified are Judgments in Program Evaluation?","year":2009,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":43,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Value (mathematics); Function (biology); Psychology; Order (exchange); Social psychology; Computer science; Economics","score_opus":0.3428563984356187,"score_gpt":0.5551089461910942,"score_spread":0.21225254775547547,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2111340440","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96743304,0.00031164574,0.00037259585,0.016790414,0.0004258271,0.003843633,0.000005646445,0.000062554,0.010754623],"genre_scores_gemma":[0.9968715,0.000017643717,0.0013399652,0.0004808336,0.00008137139,0.00048497992,0.000049939594,0.000006013042,0.00066777045],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99333256,0.000764783,0.0005463072,0.0005169652,0.004592764,0.00024662632],"domain_scores_gemma":[0.9976911,0.00019499088,0.000366344,0.00042019758,0.0012411351,0.00008622392],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.018362312,0.0001523123,0.00022036525,0.00041134784,0.00015336122,0.0007103499,0.00022714165,0.000081333754,0.00084383297],"category_scores_gemma":[0.0026617073,0.0001226223,0.00004232702,0.00080525485,0.00003987708,0.0011248793,0.000031699383,0.00012487313,0.00011482002],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000025109974,0.00015732454,0.013122111,0.000003075247,0.000004716643,6.298372e-7,0.00057983556,0.001941315,0.0003168528,0.0003987791,0.0018308804,0.98161936],"study_design_scores_gemma":[0.0013299375,0.00012363918,0.501936,0.000023513894,0.00003151119,0.00000167968,0.00080030423,0.4754462,0.00010094876,0.017744372,0.002347804,0.00011405372],"about_ca_topic_score_codex":0.000004404109,"about_ca_topic_score_gemma":0.000101090496,"teacher_disagreement_score":0.98150533,"about_ca_system_score_codex":0.00020674337,"about_ca_system_score_gemma":0.00027846888,"threshold_uncertainty_score":0.9239379},"labels":[],"label_agreement":null},{"id":"W2116343005","doi":"10.1177/1356389012453289","title":"Towards an evidence base of theory-driven evaluations: Some questions for proponents of theory-driven evaluation","year":2012,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":24,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"York University; University of Toronto; St. Michael's Hospital","funders":"","keywords":"Management science; Causality (physics); Psychological intervention; Promotion (chess); Development theory; Theory of change; Computer science; Psychology; Engineering ethics; Sociology; Political science; Economics; Economic growth; Engineering","score_opus":0.37736500390818084,"score_gpt":0.5722308654359345,"score_spread":0.19486586152775365,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2116343005","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86966616,0.002029552,0.1168348,0.0011058884,0.0016350107,0.007055422,0.00014446375,0.00006180904,0.001466905],"genre_scores_gemma":[0.98666024,0.000067669265,0.01094041,0.00013643593,0.00034448164,0.0013627841,0.00022686628,0.000029100704,0.00023200815],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.98019284,0.007960978,0.00206492,0.00069640414,0.00860393,0.0004809391],"domain_scores_gemma":[0.9846474,0.003484014,0.0016312294,0.0013083214,0.008709921,0.00021913562],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.104484,0.0002903029,0.0005167416,0.0007531174,0.00033891943,0.00010645154,0.00079106534,0.0001824574,0.0052380725],"category_scores_gemma":[0.035492517,0.00023197771,0.00025576373,0.0010116531,0.00024257696,0.0043167444,0.00010690806,0.00015412766,0.00014985878],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0007164042,0.0014692893,0.025019761,0.00010589181,0.00017096882,6.8586544e-8,0.011802788,0.06610368,0.027999293,0.20627294,0.0009374373,0.6594015],"study_design_scores_gemma":[0.0018157351,0.00061341777,0.12930736,0.00021265754,0.0007162943,0.000001950991,0.0017059443,0.5733773,0.009243588,0.28253624,0.00019668233,0.0002728408],"about_ca_topic_score_codex":0.00002832712,"about_ca_topic_score_gemma":0.00004052329,"teacher_disagreement_score":0.6591286,"about_ca_system_score_codex":0.0004366177,"about_ca_system_score_gemma":0.0021589294,"threshold_uncertainty_score":0.9956713},"labels":[],"label_agreement":null},{"id":"W2128827052","doi":"10.1177/1356389009105884","title":"Critical Connections between Participatory Evaluation, Organizational Learning and Intentional Change in Pluralistic Organizations","year":2009,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":84,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"","keywords":"Praxis; Citizen journalism; Sociology; Resistance (ecology); Process (computing); Organizational learning; Participatory action research; Organization development; Epistemology; Psychology; Knowledge management; Political science; Computer science","score_opus":0.34842097867393707,"score_gpt":0.5590078097974384,"score_spread":0.2105868311235013,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2128827052","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9656277,0.00032888804,0.008712304,0.01967823,0.0005410725,0.0012393533,0.000019854593,0.00008691916,0.0037656908],"genre_scores_gemma":[0.99817777,0.000014788179,0.00042147023,0.00051410834,0.0003066624,0.00013198525,0.00029346632,0.000011254289,0.00012852467],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9938929,0.0013719725,0.0008459202,0.00042300383,0.0032189058,0.00024727744],"domain_scores_gemma":[0.99278617,0.0013774202,0.00020330882,0.00020592935,0.005293069,0.00013409088],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.01267365,0.00015303114,0.00022024271,0.0005594899,0.0004610747,0.00027558868,0.00016487714,0.00010366692,0.011743152],"category_scores_gemma":[0.034166325,0.00014300161,0.000032200063,0.00208726,0.00009882198,0.0009961962,0.000043768592,0.0002207755,0.00029297],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000012361186,0.00016695779,0.8767616,0.000004692659,0.000015407191,5.0444913e-7,0.0047862697,0.0034267898,0.00033790158,0.052694105,0.00042273954,0.06137066],"study_design_scores_gemma":[0.00062780915,0.0000793889,0.6779665,0.00001935413,0.000072163966,0.0000030833667,0.0007567983,0.25420234,0.000039562117,0.06569722,0.00040540364,0.00013037668],"about_ca_topic_score_codex":0.000018889175,"about_ca_topic_score_gemma":0.00011294318,"teacher_disagreement_score":0.25077555,"about_ca_system_score_codex":0.0003001598,"about_ca_system_score_gemma":0.00060758844,"threshold_uncertainty_score":0.98916024},"labels":[],"label_agreement":null},{"id":"W2130276747","doi":"10.1177/1356389015593357","title":"Towards a comprehensive framework for the evaluation of small and medium enterprise policy","year":2015,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of British Columbia","funders":"","keywords":"Additionality; Relevance (law); Public policy; Policy analysis; Process management; Business; Management science; Knowledge management; Computer science; Political science; Economics; Public economics; Public administration; Economic growth","score_opus":0.4988657924257035,"score_gpt":0.574468999590146,"score_spread":0.07560320716444247,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2130276747","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.724194,0.0022599397,0.23669784,0.026420504,0.0017335059,0.004328517,0.0000337664,0.000029612189,0.004302332],"genre_scores_gemma":[0.99111056,0.000037485523,0.0075392663,0.0005634573,0.00031489236,0.00032858108,0.000019026409,0.000008251424,0.00007846703],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9932095,0.00096221705,0.0006148007,0.00029674167,0.0047601885,0.00015654344],"domain_scores_gemma":[0.9899485,0.0020753134,0.00045589934,0.00049777917,0.0069250916,0.00009742288],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.030734038,0.00011669674,0.00020239675,0.00024030142,0.00013270754,0.00012152822,0.0003244335,0.00008305423,0.00036750198],"category_scores_gemma":[0.020636955,0.00007208049,0.00007153903,0.00055184006,0.000103980725,0.00024812034,0.00008198497,0.00009073166,0.000036001995],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00018457943,0.00007992711,0.005062198,0.000011574569,0.000060676106,4.188981e-8,0.011760584,0.019257698,0.00019132164,0.010928967,0.003929641,0.9485328],"study_design_scores_gemma":[0.0015211235,0.00016395721,0.114888825,0.000021411874,0.00016129362,0.0000015527995,0.0034680725,0.58593744,0.000351086,0.28925675,0.004149003,0.00007950145],"about_ca_topic_score_codex":0.00007209998,"about_ca_topic_score_gemma":0.000075696335,"teacher_disagreement_score":0.9484533,"about_ca_system_score_codex":0.00018982218,"about_ca_system_score_gemma":0.0021575463,"threshold_uncertainty_score":0.99806327},"labels":[],"label_agreement":null},{"id":"W2133487638","doi":"10.1177/1356389012452052","title":"Applications of contribution analysis to outcome planning and impact evaluation","year":2012,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":44,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Accountability; General partnership; Participatory evaluation; Context (archaeology); Participatory planning; Citizen journalism; Theory of change; Process management; Process (computing); Value (mathematics); Management science; Appeal; Conceptual framework; Knowledge management; Public relations; Sociology; Computer science; Political science; Business; Public administration; Environmental planning; Engineering","score_opus":0.3089908516284891,"score_gpt":0.6239675731195479,"score_spread":0.31497672149105876,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2133487638","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8170985,0.00045789767,0.17897859,0.0003859971,0.00012398083,0.0012746659,0.000020354686,0.0000145714575,0.0016454589],"genre_scores_gemma":[0.99837637,0.000002308337,0.0009015559,0.00010003592,0.000099110715,0.00031838156,0.00013409088,0.0000047654157,0.00006336592],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.994406,0.0006871073,0.0008399382,0.00027206357,0.0035808806,0.00021403242],"domain_scores_gemma":[0.9959242,0.00056153705,0.00046913346,0.00043904711,0.0024425576,0.00016352547],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.034517214,0.000113389215,0.00028397792,0.0009128731,0.00015045394,0.00008520959,0.00016359605,0.00006327224,0.003159418],"category_scores_gemma":[0.0033353234,0.000086491134,0.0001213477,0.002497255,0.000027858054,0.00068384095,0.00004763156,0.000064810854,0.00015929806],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000019178764,0.000050127663,0.82389617,0.0000014911046,0.00010025096,6.4614176e-9,0.00092222355,0.036873713,0.00055159716,0.00065994373,0.00017906491,0.1367462],"study_design_scores_gemma":[0.00037762488,0.000044588593,0.753021,0.0000028525865,0.0006131851,5.1769604e-7,0.0003534239,0.24259792,0.00029459273,0.0018472763,0.0007664554,0.000080584825],"about_ca_topic_score_codex":0.00002591621,"about_ca_topic_score_gemma":0.00001997886,"teacher_disagreement_score":0.20572421,"about_ca_system_score_codex":0.0002516858,"about_ca_system_score_gemma":0.00017361452,"threshold_uncertainty_score":0.99775183},"labels":[],"label_agreement":null},{"id":"W2133783179","doi":"10.1177/1356389008101967","title":"Using Logic Analysis to Evaluate Knowledge Transfer Initiatives","year":2009,"lang":"en","type":"article","venue":"Evaluation","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal","funders":"Canadian Institutes of Health Research; U.S. Public Health Service; Canadian Health Services Research Foundation","keywords":"Popularity; Knowledge transfer; Knowledge management; Process (computing); Politics; Order (exchange); Technology transfer; Logic model; Institutional logic; Management science; Computer science; Political science; Sociology; Engineering; Business; Social science; Public administration","score_opus":0.8591549218648961,"score_gpt":0.7677565598744323,"score_spread":0.0913983619904638,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2133783179","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.83785594,0.000037305104,0.12636456,0.010127704,0.00033651764,0.002332649,0.000026499258,0.0000868971,0.022831907],"genre_scores_gemma":[0.97753507,0.0000049297355,0.004085818,0.017814845,0.00020964349,0.00015143237,0.000021787126,0.000007785758,0.00016870574],"study_design_codex":"qualitative","study_design_gemma":"observational","domain_scores_codex":[0.9938597,0.0036000877,0.0007848495,0.00035919066,0.0008499806,0.00054617523],"domain_scores_gemma":[0.9975807,0.0007281092,0.00012249847,0.0002922317,0.0010280499,0.00024842078],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.0105098095,0.00013149982,0.0002594521,0.00068557734,0.0007638956,0.00001553295,0.00016194163,0.00008726326,0.0040259114],"category_scores_gemma":[0.0022253962,0.00011971657,0.00007408754,0.002792341,0.000023800363,0.00036781884,0.000020616439,0.00019791834,0.0010986906],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00032215752,0.0004751079,0.04848676,0.00021627375,0.00065458304,0.0000031077232,0.61253035,0.07969255,0.05069364,0.031226622,0.011980694,0.16371816],"study_design_scores_gemma":[0.0015887185,0.00022763868,0.80790955,0.000052696127,0.000771156,2.6914563e-7,0.0029735903,0.1775627,0.00023956629,0.0047330996,0.0036555866,0.000285442],"about_ca_topic_score_codex":0.00006552365,"about_ca_topic_score_gemma":0.00026386784,"teacher_disagreement_score":0.7594228,"about_ca_system_score_codex":0.0006522835,"about_ca_system_score_gemma":0.0012652145,"threshold_uncertainty_score":0.9996791},"labels":[],"label_agreement":null},{"id":"W2143790740","doi":"10.1177/1356389008090857","title":"Health Technology Assessment in the Canadian Health Policy Arena","year":2008,"lang":"en","type":"article","venue":"Evaluation","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Université de Montréal","funders":"Canadian Institutes of Health Research; Canada Research Chairs","keywords":"Autonomy; Agency (philosophy); Accountability; Health technology; Public relations; Politics; Health care; Perspective (graphical); Public administration; Political science; Sociology; Social science","score_opus":0.45063542929334766,"score_gpt":0.5225140240842393,"score_spread":0.07187859479089165,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2143790740","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0643134,0.0032913252,0.0007528627,0.921379,0.00035645757,0.0020470074,0.00012047159,0.000042781394,0.0076966495],"genre_scores_gemma":[0.9389544,0.00023241596,0.0014390869,0.058536958,0.00028014154,0.00029501872,0.00013953011,0.000018037836,0.00010437297],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","domain_scores_codex":[0.994418,0.0012318563,0.0029646717,0.0004678575,0.0002567694,0.0006608341],"domain_scores_gemma":[0.99732083,0.00025232826,0.0015289846,0.0005961271,0.00008990949,0.000211802],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.04650024,0.00014802748,0.00063712645,0.0014834984,0.00076606584,0.00004715114,0.00033844233,0.00013289605,0.00020505767],"category_scores_gemma":[0.0024969492,0.00016333847,0.000055079756,0.00092153996,0.00008227393,0.00024516726,0.000019683855,0.00033273143,0.0007873935],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.0000029906269,0.00015244607,0.1446466,0.000104390296,0.000027984863,0.0000017351011,0.01389114,0.0012283666,2.1346506e-7,0.72628397,0.106879584,0.006780594],"study_design_scores_gemma":[0.0011196592,0.00019952416,0.84287906,0.00006223865,0.0000017046744,0.000038607162,0.0024847041,0.01621,3.0472415e-7,0.058350723,0.0783843,0.00026914952],"about_ca_topic_score_codex":0.50614643,"about_ca_topic_score_gemma":0.5867924,"teacher_disagreement_score":0.87464106,"about_ca_system_score_codex":0.0076675736,"about_ca_system_score_gemma":0.009894305,"threshold_uncertainty_score":0.9999906},"labels":[],"label_agreement":null},{"id":"W2154807644","doi":"10.1177/1356389009341729","title":"Evaluating Service Organization Models","year":2009,"lang":"en","type":"article","venue":"Evaluation","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Public Health Agency of Canada; Université de Montréal; École Nationale d'Administration Publique","funders":"Canadian Institutes of Health Research; U.S. Public Health Service; Ministère de la Santé et des Services sociaux","keywords":"Computer science; Relevance (law); Set (abstract data type); Management science; Context (archaeology); Service (business); Contrast (vision); Data science; Artificial intelligence; Engineering","score_opus":0.8545316963181638,"score_gpt":0.762880564368063,"score_spread":0.09165113195010088,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2154807644","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.91941416,0.000013758967,0.017813854,0.04261166,0.0005045898,0.0022400457,0.000008066542,0.0002218022,0.017172044],"genre_scores_gemma":[0.9398211,0.0000040358123,0.0052471072,0.054173373,0.0003136694,0.000095444455,0.00006574368,0.0000134196525,0.0002660676],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9959498,0.0014510325,0.0006907174,0.00027848012,0.0012017423,0.00042825253],"domain_scores_gemma":[0.9961576,0.00052673934,0.0003614034,0.0003000595,0.0025216315,0.00013255325],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.009270842,0.00010058965,0.0001244037,0.00013532411,0.00095652154,0.000017408085,0.00016811094,0.00009602867,0.0030809913],"category_scores_gemma":[0.00416914,0.00009936996,0.000012818961,0.0013167068,0.000009134707,0.0006690601,0.00003015368,0.00020758617,0.0015034208],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00009085436,0.00017749344,0.015431541,0.00035245446,0.000028031956,0.000001031145,0.23034911,0.08856766,0.062386,0.05749112,0.063773274,0.48135144],"study_design_scores_gemma":[0.002411178,0.00016024997,0.12767209,0.00010222881,0.00005129729,0.0000014694112,0.0026838786,0.8183572,0.00034140755,0.04596051,0.0020015654,0.00025691962],"about_ca_topic_score_codex":0.000092176546,"about_ca_topic_score_gemma":0.00007178946,"teacher_disagreement_score":0.72978956,"about_ca_system_score_codex":0.00051608396,"about_ca_system_score_gemma":0.0014859799,"threshold_uncertainty_score":0.999274},"labels":[],"label_agreement":null},{"id":"W2155601053","doi":"10.1177/1356389004046292","title":"What Counts is not Falling... but Landing","year":2004,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Douglas Mental Health University Institute; Douglas College","funders":"Canadian Institutes of Health Research; U.S. Public Health Service","keywords":"Strengths and weaknesses; Process management; Falling (accident); Process (computing); Management science; Computer science; Operations research; Political science; Business; Psychology; Engineering; Medicine; Environmental health","score_opus":0.36712162891403827,"score_gpt":0.545401205321608,"score_spread":0.1782795764075697,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2155601053","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.92755425,0.00027927163,0.006742318,0.009161123,0.003438715,0.00042069584,0.0000063643515,0.00004777552,0.052349467],"genre_scores_gemma":[0.99431556,0.000074281306,0.000447267,0.0014050642,0.00013809718,0.000013810451,0.000011164338,0.0000042245697,0.0035905119],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9970722,0.000080479724,0.00026106922,0.00018754521,0.0023031516,0.00009552032],"domain_scores_gemma":[0.99907464,0.000108738735,0.00011256081,0.00022441888,0.00044530744,0.000034349123],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.005076817,0.00005653532,0.000077223165,0.00010844682,0.00010164884,0.00031791552,0.00015293533,0.000038686205,0.009722084],"category_scores_gemma":[0.00042916366,0.000043299173,0.000036302517,0.00023856878,0.000015740961,0.00072246196,0.000018467708,0.000054187076,0.00785042],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000040870516,0.000106230436,0.0068528005,0.0000071623153,0.000029563618,0.0000014235293,0.0075694574,0.015738226,0.0033912587,0.007198231,0.029173158,0.92989165],"study_design_scores_gemma":[0.004603031,0.00023121231,0.11026843,0.00018699931,0.000106086874,0.000011026098,0.0066615185,0.3872756,0.020782955,0.23567007,0.23353624,0.0006668172],"about_ca_topic_score_codex":0.000007615105,"about_ca_topic_score_gemma":0.000015394933,"teacher_disagreement_score":0.9292248,"about_ca_system_score_codex":0.00012192347,"about_ca_system_score_gemma":0.00023405065,"threshold_uncertainty_score":0.99292207},"labels":[],"label_agreement":null},{"id":"W2157111091","doi":"10.1177/1356389010380001","title":"To Be or Not to Be a Profession: Pros, Cons and Challenges for Evaluation","year":2010,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"École Nationale d'Administration Publique; Université Laval","funders":"","keywords":"Professionalization; Institutionalisation; Certification; Charter; Political science; Public relations; Accreditation; Valuation (finance); Quality (philosophy); Sociology; Engineering ethics; Public administration; Law; Accounting; Business; Engineering","score_opus":0.6053607172548194,"score_gpt":0.6135973810690398,"score_spread":0.00823666381422039,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2157111091","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.79473704,0.000077760444,0.0013335758,0.19377165,0.0015355375,0.006384036,0.000033159653,0.000042387277,0.002084868],"genre_scores_gemma":[0.9816914,0.000031661108,0.00793967,0.005393663,0.00030901915,0.0026318098,0.00004603757,0.000017702278,0.0019390256],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.99355215,0.00054636714,0.0006621632,0.00069810695,0.004267258,0.00027397982],"domain_scores_gemma":[0.99454087,0.0011910578,0.00024414496,0.00062086707,0.0031648166,0.00023823256],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.032717694,0.00017851626,0.00024907006,0.00036657625,0.0003688266,0.0002379841,0.0003112247,0.00013259394,0.0035911864],"category_scores_gemma":[0.014972149,0.00012306368,0.000051224604,0.00051089254,0.00004082708,0.0005212143,0.00010860922,0.00014390281,0.00019810378],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0005851957,0.00012769984,0.0006174099,0.000027536666,0.000017848599,1.7389266e-7,0.016419306,0.00047202804,0.011196504,0.0037353323,0.021085361,0.9457156],"study_design_scores_gemma":[0.005952696,0.0017382143,0.18162662,0.000082208506,0.00022010591,0.00001019036,0.013685881,0.3799471,0.008680734,0.012741659,0.3946256,0.00068901037],"about_ca_topic_score_codex":0.0000064786745,"about_ca_topic_score_gemma":0.0016139788,"teacher_disagreement_score":0.9450266,"about_ca_system_score_codex":0.00009894803,"about_ca_system_score_gemma":0.0010592877,"threshold_uncertainty_score":0.99731964},"labels":[],"label_agreement":null},{"id":"W2171507080","doi":"10.1177/1356389011430371","title":"Evaluation models and evaluation use","year":2012,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":59,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Sherbrooke; Université de Montréal","funders":"Canadian Institutes of Health Research; U.S. Public Health Service","keywords":"Context (archaeology); Management science; Psychological intervention; Process (computing); Perspective (graphical); Field (mathematics); Evaluation methods; Impact evaluation; Systematic review; Affect (linguistics); Program evaluation; Computer science; Knowledge management; Process management; Psychology; Political science; Engineering; Artificial intelligence; Medicine; MEDLINE","score_opus":0.6626932025379391,"score_gpt":0.5990196685991209,"score_spread":0.06367353393881814,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2171507080","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9668698,0.0015645671,0.0068127415,0.0008694028,0.0014973313,0.0022817894,0.000008145826,0.00005124658,0.020044986],"genre_scores_gemma":[0.9968998,0.000045826982,0.0014463653,0.00027284288,0.00042087593,0.00048293208,0.000109272034,0.000017682662,0.00030444245],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.97816,0.0040842025,0.0009581088,0.0005798193,0.015796611,0.0004212246],"domain_scores_gemma":[0.9907592,0.0012668835,0.00049075537,0.0007430733,0.006521757,0.0002183534],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.12947914,0.0002182865,0.00024430937,0.00047638564,0.0003357339,0.0005037873,0.0002409877,0.00013676015,0.0077032982],"category_scores_gemma":[0.012562397,0.00017696261,0.000079074096,0.00078560994,0.00006728505,0.0062922225,0.00008957772,0.00014819198,0.0006465879],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00002554527,0.0001227148,0.016240744,0.0000028776585,0.000031931464,3.5881353e-8,0.0036830197,0.047439437,0.00043990175,0.004128042,0.004689848,0.9231959],"study_design_scores_gemma":[0.0011752428,0.000037207457,0.16239081,0.000010815628,0.00033469076,0.0000036638567,0.00065824087,0.77540463,0.000173651,0.057667717,0.0019632413,0.00018010789],"about_ca_topic_score_codex":0.000022014376,"about_ca_topic_score_gemma":0.00008454061,"teacher_disagreement_score":0.9230158,"about_ca_system_score_codex":0.0005734608,"about_ca_system_score_gemma":0.0008599508,"threshold_uncertainty_score":0.9957552},"labels":[],"label_agreement":null},{"id":"W2321004893","doi":"10.1177/1356389014540705","title":"Evaluations, international agencies and censorship: A field doer’s viewpoint","year":2014,"lang":"en","type":"article","venue":"Evaluation","topic":"Qualitative Research Methods and Ethics","field":"Social Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Field (mathematics); Sociology; Action (physics); Perspective (graphical); Epistemology; Computer science; Mathematics","score_opus":0.27986036555218136,"score_gpt":0.6030785631938922,"score_spread":0.32321819764171084,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2321004893","genre_codex":"other","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23680189,0.0010555728,0.14163768,0.28030106,0.0019604983,0.0016641133,0.000010531544,0.00015049364,0.33641815],"genre_scores_gemma":[0.97791713,0.00054264616,0.017479332,0.0014430055,0.00057202135,0.000089681074,0.00001615212,0.000007476909,0.0019325765],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9934885,0.0047238544,0.000166369,0.00017044126,0.001289187,0.00016164439],"domain_scores_gemma":[0.9969833,0.0021005985,0.000068074296,0.0001014936,0.00066527375,0.00008128759],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.033752076,0.0000556797,0.000078203426,0.000082046805,0.00039556276,0.00010892201,0.00012197621,0.00007754052,0.0013742243],"category_scores_gemma":[0.03156105,0.000051181054,0.000029909652,0.000136881,0.0002044099,0.00022221275,0.00003266329,0.00014388384,0.000056760207],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00001626498,0.000020785294,0.0008124181,0.000013297551,0.000020283058,1.3965905e-7,0.04588155,0.00002928202,0.00060226413,0.43412837,0.0030586116,0.51541674],"study_design_scores_gemma":[0.00058969756,0.00013965972,0.015381028,0.000060397953,0.00004257388,5.813944e-7,0.015458889,0.034110375,0.0004020259,0.6363581,0.2972437,0.00021295546],"about_ca_topic_score_codex":0.00030875596,"about_ca_topic_score_gemma":0.00039096622,"teacher_disagreement_score":0.7411152,"about_ca_system_score_codex":0.00014270113,"about_ca_system_score_gemma":0.00025278077,"threshold_uncertainty_score":0.99953866},"labels":[],"label_agreement":null},{"id":"W2324664192","doi":"10.1177/1356389014564248","title":"The institutionalization of evaluation matters: Updating the <i>International Atlas of Evaluation</i> 10 years later","year":2015,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":114,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université Laval","funders":"","keywords":"Institutionalisation; Political science; Regional science; Public administration; Geography; Law","score_opus":0.2102689892953537,"score_gpt":0.4898436435508309,"score_spread":0.2795746542554772,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2324664192","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9080491,0.0011072233,0.003734284,0.026082715,0.0046652746,0.003868282,0.000049803853,0.000038477603,0.052404813],"genre_scores_gemma":[0.9978589,0.00002755371,0.00040923254,0.00048209072,0.00022402499,0.00023464009,0.00025694477,0.000011003814,0.00049565325],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9822151,0.0020715038,0.0015023838,0.00038353654,0.013633656,0.00019379103],"domain_scores_gemma":[0.986153,0.001104237,0.0013322355,0.00073282607,0.010616667,0.000061013943],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.071788594,0.00014952618,0.00019909994,0.00024275074,0.000268019,0.00024386842,0.0006609633,0.000080972466,0.004859765],"category_scores_gemma":[0.010870377,0.000090525355,0.00011841597,0.00086915575,0.00022376583,0.00089328474,0.00013517065,0.000115561365,0.00036903162],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00027431693,0.00021323812,0.014640591,0.000009817303,0.00019039285,2.5562437e-7,0.0063946275,0.41197252,0.0020958793,0.014199504,0.13778523,0.4122236],"study_design_scores_gemma":[0.0019615155,0.00007998525,0.052228473,0.00003409219,0.0001868774,0.000004835899,0.0024469874,0.86182547,0.0015806805,0.03365036,0.045869738,0.00013100257],"about_ca_topic_score_codex":0.000038117512,"about_ca_topic_score_gemma":0.0001409408,"teacher_disagreement_score":0.4498529,"about_ca_system_score_codex":0.00042021775,"about_ca_system_score_gemma":0.0020985885,"threshold_uncertainty_score":0.9974615},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"observational","genre":"empirical","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"agree"},{"id":"W2605706618","doi":"10.1177/1356389017697620","title":"Evaluability assessment of a small NGO in water-based development","year":2017,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Waterloo","funders":"Mitacs","keywords":"Accountability; Context (archaeology); Political science; Public relations; Water development; Face (sociological concept); Qualitative research; Baseline (sea); Psychology; Sociology; Water resources; Geography; Social science","score_opus":0.4462252381466189,"score_gpt":0.5827030978504076,"score_spread":0.13647785970378873,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2605706618","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9783971,0.00001108007,0.0060170037,0.001561852,0.00037060858,0.00079419516,0.0000018044892,0.000010699163,0.012835661],"genre_scores_gemma":[0.9893357,0.0000011084741,0.010132974,0.00008154628,0.000027979153,0.00016896622,0.000025564303,0.000006244187,0.0002198929],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9935225,0.00083791156,0.0011472378,0.00046289215,0.0037911467,0.00023831133],"domain_scores_gemma":[0.99638367,0.00034759982,0.0005687114,0.001271999,0.0013631391,0.0000648564],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.04582828,0.00013442134,0.00028529888,0.00030582037,0.00031477617,0.0002223635,0.0007757697,0.00007334439,0.0038971938],"category_scores_gemma":[0.0028727537,0.00009035523,0.00007416601,0.00015335137,0.00009105515,0.00041467822,0.00013077489,0.000117579344,0.000153893],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000044410113,0.0003992253,0.6240411,0.000021212876,0.0000127869025,5.595517e-7,0.0015256848,0.040380202,0.0033028491,0.00029814697,0.000095128235,0.32987872],"study_design_scores_gemma":[0.001043278,0.00005030922,0.6518721,0.000021292131,0.000009286129,1.3771889e-7,0.00009806313,0.3366754,0.0061176424,0.0032073467,0.0008218672,0.00008328992],"about_ca_topic_score_codex":0.00006404194,"about_ca_topic_score_gemma":0.0008674607,"teacher_disagreement_score":0.32979542,"about_ca_system_score_codex":0.00039145295,"about_ca_system_score_gemma":0.0017292053,"threshold_uncertainty_score":0.9970134},"labels":[],"label_agreement":null},{"id":"W2606109332","doi":"10.1177/1356389017697615","title":"Network-based approaches for evaluating ambient assisted living (AAL) technologies","year":2017,"lang":"en","type":"article","venue":"Evaluation","topic":"Technology Use by Older Adults","field":"Social Sciences","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University; University of Toronto","funders":"Economic and Social Research Council","keywords":"Assisted living; Perspective (graphical); Activities of daily living; Computer science; Cognition; Bridge (graph theory); Independent living; Function (biology); Living lab; Human–computer interaction; Knowledge management; Psychology; Gerontology; Medicine; Artificial intelligence","score_opus":0.19563511380443538,"score_gpt":0.399164352199533,"score_spread":0.20352923839509762,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2606109332","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9084874,0.00073775655,0.038244233,0.029996896,0.0017006745,0.0058822315,0.00001037524,0.0026571532,0.012283273],"genre_scores_gemma":[0.9807859,0.0000048082743,0.017917925,0.000048883812,0.00016903858,0.00082451,0.000011945994,0.000017818156,0.00021919252],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9977641,0.00032933467,0.0002544614,0.0004043672,0.00079188735,0.00045585973],"domain_scores_gemma":[0.99791247,0.0004904544,0.00046124012,0.000707788,0.0003952976,0.00003276233],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.008403167,0.00013482879,0.00017362025,0.00010378937,0.0029754988,0.0002621843,0.00071426097,0.00030517095,0.00004049478],"category_scores_gemma":[0.018200617,0.00013821195,0.0000778642,0.00015395408,0.00044703382,0.000328816,0.00011516553,0.0001541098,0.000018266697],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000025557598,0.00011400818,0.022015512,0.000027982765,0.000038051898,4.2619504e-7,0.0026906182,0.0030064422,0.0004483521,0.013685265,0.00196997,0.9559778],"study_design_scores_gemma":[0.002450905,0.00038986624,0.3194819,0.00056965684,0.0003982462,0.0000010513886,0.010473699,0.5799877,0.0020603468,0.07731885,0.006052908,0.00081486173],"about_ca_topic_score_codex":0.0001854625,"about_ca_topic_score_gemma":0.0019942992,"teacher_disagreement_score":0.95516294,"about_ca_system_score_codex":0.00031306932,"about_ca_system_score_gemma":0.000419843,"threshold_uncertainty_score":0.9983225},"labels":[],"label_agreement":null},{"id":"W2742194300","doi":"10.1177/1356389017715719","title":"Using online mining techniques to inform formative evaluations: An analysis of YouTube video comments about chronic pain","year":2017,"lang":"en","type":"article","venue":"Evaluation","topic":"Focus Groups and Qualitative Methods","field":"Social Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Economic and Social Research Council; Centre National de la Recherche Scientifique; McGill University","keywords":"Judgement; Formative assessment; Social media; Psychology; Applied psychology; Computer science; Public relations; World Wide Web; Political science","score_opus":0.3269929441657317,"score_gpt":0.5957748917242113,"score_spread":0.26878194755847956,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2742194300","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8785248,0.000083311555,0.11489079,0.0010659625,0.00014147097,0.0010381807,0.000060043858,0.00004193442,0.0041535357],"genre_scores_gemma":[0.9171794,0.000011910149,0.082224905,0.00021364586,0.00012820587,0.00006440662,0.00011875843,0.00000779746,0.000050947794],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9953574,0.0026773268,0.0004804406,0.00018735508,0.0010839287,0.00021356212],"domain_scores_gemma":[0.9973332,0.00041996897,0.00059920235,0.0004095848,0.0011488955,0.00008909162],"candidate_categories":["metaresearch","sts"],"consensus_categories":[],"category_scores_codex":[0.030545456,0.00010853819,0.0002789803,0.00039919943,0.00135189,0.0001425155,0.00033531754,0.000079434736,0.00027186156],"category_scores_gemma":[0.0076190853,0.000103238875,0.00010166556,0.00063823245,0.00018060558,0.0012353122,0.000060795064,0.000068544534,0.0000025410554],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000022130325,0.00010938614,0.0081911,0.000016006086,0.00026788478,7.9386126e-8,0.18451492,0.0027044113,0.0010541042,0.0034175136,0.00007792812,0.7996245],"study_design_scores_gemma":[0.00053826807,0.00054122874,0.19237867,0.00019681353,0.0014224243,5.684851e-8,0.056936007,0.73941845,0.0013931422,0.0036000821,0.0032326246,0.00034223936],"about_ca_topic_score_codex":0.0032440159,"about_ca_topic_score_gemma":0.008981107,"teacher_disagreement_score":0.7992823,"about_ca_system_score_codex":0.00070742826,"about_ca_system_score_gemma":0.00035880806,"threshold_uncertainty_score":0.9999482},"labels":[],"label_agreement":null},{"id":"W2800395496","doi":"10.1177/1356389018763242","title":"The evaluation of social innovation: A review and integration of the current empirical knowledge base","year":2018,"lang":"en","type":"review","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":75,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa","funders":"","keywords":"Empirical research; Knowledge management; Knowledge base; Social innovation; Scale (ratio); Social learning; Perspective (graphical); Public relations; Computer science; Political science","score_opus":0.735361020503894,"score_gpt":0.6846952553233857,"score_spread":0.05066576518050825,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2800395496","genre_codex":"review","genre_gemma":"review","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"review","genre_consensus":"review","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.00014783848,0.99107516,0.00016549493,0.0010946924,0.0011065527,0.0037372555,0.00002689948,0.000006465999,0.0026396567],"genre_scores_gemma":[0.0020497027,0.9965651,0.000041487958,0.00006971597,0.00030366136,0.00061240286,0.00011480591,0.000015048337,0.00022808951],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9819608,0.008051935,0.0030800223,0.0005149441,0.0062072347,0.00018508136],"domain_scores_gemma":[0.98232067,0.0015576759,0.0034916,0.0008422759,0.011757849,0.000029923061],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.08993141,0.0002714268,0.00095171126,0.00033570552,0.0004428288,0.0001018824,0.0007263231,0.00012299456,0.0008066486],"category_scores_gemma":[0.018286293,0.00012663752,0.00028981417,0.0036579224,0.0003287219,0.00026726292,0.00019542979,0.00033856486,0.00007970074],"study_design_candidate":"not_applicable","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000033094568,0.000052527623,0.000010198817,0.001352405,0.000027472895,2.1335878e-9,0.00034433434,0.0000012537071,5.2038274e-7,0.0009825772,0.018810531,0.9784149],"study_design_scores_gemma":[0.00038427,0.0000958495,0.0017016963,0.0071949116,0.0032132156,0.0000019288075,0.00007722396,0.010307582,0.000005541917,0.006293852,0.9705553,0.00016864366],"about_ca_topic_score_codex":0.0000014479458,"about_ca_topic_score_gemma":0.000060101356,"teacher_disagreement_score":0.9782462,"about_ca_system_score_codex":0.0003575836,"about_ca_system_score_gemma":0.004849209,"threshold_uncertainty_score":0.9899831},"labels":[],"label_agreement":null},{"id":"W2801790340","doi":"10.1177/1356389018765487","title":"Theory-based evaluations: Framing the existence of a new theory in evaluation and the rise of the 5th generation","year":2018,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":57,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre Intégré Universitaire de Santé et de Services Sociaux du Saguenay–Lac-Saint-Jean; Université de Sherbrooke; University of Victoria","funders":"","keywords":"Epistemology; Framing (construction); Complementarity (molecular biology); Realism; Sociology; Computer science; Philosophy","score_opus":0.28830792381595954,"score_gpt":0.5359311604701663,"score_spread":0.24762323665420677,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2801790340","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96020854,0.0012979815,0.026404109,0.004957583,0.00064242544,0.0030336266,0.000005543953,0.000008794125,0.0034413922],"genre_scores_gemma":[0.9981953,0.000018662413,0.0006308205,0.0004261561,0.00017538865,0.00020625164,0.0000071777968,0.00000878782,0.0003314902],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.98093605,0.011423248,0.0013131201,0.00038277914,0.005752627,0.00019215721],"domain_scores_gemma":[0.9910332,0.0038688737,0.0010831376,0.0011110315,0.0028686158,0.000035146444],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.12298959,0.00015409176,0.00024806117,0.00022443909,0.00043036835,0.00012472861,0.0006983181,0.00008327469,0.0017113321],"category_scores_gemma":[0.021416653,0.000069813956,0.00011409014,0.0013736546,0.00085816195,0.0004040151,0.00009762762,0.00016725733,0.00002819772],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00050857814,0.00012141381,0.0123663675,0.000012434962,0.000047216385,2.7144022e-8,0.03323808,0.039657474,0.0041656904,0.1441031,0.000825192,0.76495445],"study_design_scores_gemma":[0.0021884877,0.0000692762,0.10523258,0.000044547756,0.00014188897,5.495108e-7,0.0021916716,0.6534495,0.003167833,0.23328385,0.00016444045,0.00006540613],"about_ca_topic_score_codex":0.000078720295,"about_ca_topic_score_gemma":0.000550226,"teacher_disagreement_score":0.764889,"about_ca_system_score_codex":0.00013061018,"about_ca_system_score_gemma":0.00177896,"threshold_uncertainty_score":0.99920124},"labels":[],"label_agreement":null},{"id":"W2913663231","doi":"10.1177/1356389019827035","title":"Combining internal and external evaluations within a multilevel evaluation framework: Computational text analysis of lessons from the Asian Development Bank","year":2019,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Simon Fraser University","funders":"Lee Kuan Yew School of Public Policy, National University of Singapore; Whitney and Betty MacMillan Center for International and Area Studies; Yale University","keywords":"Macro; Psychological intervention; Micro level; Scenario analysis; Political science; Computer science; Process management; Psychology; Business; Economics; Finance; Microeconomics","score_opus":0.23131795111349623,"score_gpt":0.5268569038509368,"score_spread":0.2955389527374405,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2913663231","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86836123,0.00018722734,0.12593582,0.002374948,0.00055713125,0.0010760957,0.000041421725,0.000016815926,0.0014493029],"genre_scores_gemma":[0.97846675,0.000004444834,0.020460628,0.00040655572,0.00006133234,0.00013802675,0.0002860427,0.000012482489,0.0001637499],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9881575,0.0018628136,0.0015288497,0.0006789352,0.007554268,0.00021763478],"domain_scores_gemma":[0.9914562,0.003700934,0.0011900467,0.00064768374,0.0029106722,0.00009443977],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.028343145,0.00022424606,0.0004417207,0.00065050396,0.00035500125,0.00033284345,0.0005607258,0.00012312042,0.009750921],"category_scores_gemma":[0.0042338576,0.00015981836,0.0001540067,0.0014077937,0.00011082527,0.0005883819,0.0001599292,0.0002821353,0.0002558053],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000057703735,0.00010779811,0.17847937,0.0000025431675,0.0005105262,1.0276973e-7,0.019573038,0.3695074,0.0003901339,0.003953611,0.00015101292,0.42726675],"study_design_scores_gemma":[0.0006807516,0.00003299894,0.47326377,0.00005620252,0.00047290899,5.5681335e-7,0.0015411783,0.5043526,0.00009519552,0.019336695,0.00006954982,0.00009759523],"about_ca_topic_score_codex":0.00010668513,"about_ca_topic_score_gemma":0.00042452302,"teacher_disagreement_score":0.42716914,"about_ca_system_score_codex":0.00028342596,"about_ca_system_score_gemma":0.0013497504,"threshold_uncertainty_score":0.9911543},"labels":[],"label_agreement":null},{"id":"W2973640600","doi":"10.1177/1356389019870213","title":"Rapid impact evaluation","year":2019,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Summative assessment; Impact evaluation; Formative assessment; Counterfactual thinking; Credibility; Stakeholder; Stakeholder engagement; Impact assessment; Ex-ante; Salience (neuroscience); Program evaluation; Theory of change; Process management; Legitimacy; Computer science; Management science; Knowledge management; Psychology; Business; Political science; Public relations; Economics; Social psychology; Artificial intelligence","score_opus":0.3381382875021545,"score_gpt":0.5894724107344181,"score_spread":0.2513341232322636,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2973640600","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.8863718,0.00028897144,0.00055439863,0.0009726254,0.0012486198,0.0015233598,0.0000050410285,0.00003679221,0.10899839],"genre_scores_gemma":[0.99724853,0.000012762353,0.00022987991,0.0002972326,0.00016141542,0.0001095081,0.00009319406,0.000011159888,0.0018363221],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9881693,0.0014467104,0.00071780395,0.000521596,0.008897135,0.0002474397],"domain_scores_gemma":[0.9952343,0.00057084224,0.00039116936,0.0008702323,0.002838637,0.00009480512],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.04635462,0.00015459766,0.00021967468,0.00039141305,0.00012441498,0.00028017693,0.00039328306,0.00008488656,0.17223346],"category_scores_gemma":[0.0031270778,0.00010883505,0.0001618908,0.00092568423,0.000023749735,0.0010793101,0.00005027773,0.0001221302,0.018877393],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000042894277,0.000059049464,0.052304924,0.000002071212,0.000024129013,8.1862545e-8,0.0006016649,0.038213342,0.001619306,0.0005133395,0.011670101,0.8949491],"study_design_scores_gemma":[0.0011409542,0.00015091256,0.34132248,0.0000070984884,0.0000395946,0.0000019114352,0.0002934946,0.6352287,0.00031824218,0.012769017,0.0086043505,0.00012326778],"about_ca_topic_score_codex":0.000017152603,"about_ca_topic_score_gemma":0.00001884264,"teacher_disagreement_score":0.8948258,"about_ca_system_score_codex":0.00047147486,"about_ca_system_score_gemma":0.0011256437,"threshold_uncertainty_score":0.9819786},"labels":[],"label_agreement":null},{"id":"W3020197475","doi":"10.1177/1356389020911060","title":"Evaluators in the Anthropocene","year":2020,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":30,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria; Université de Sherbrooke","funders":"","keywords":"Anthropocene; Biosphere; Environmental ethics; Action (physics); State (computer science); Natural (archaeology); Call to action; Earth system science; Work (physics); Non-human; Environmental planning; Environmental resource management; Political science; Ecology; History; Geography; Archaeology; Environmental science; Business; Law; Computer science; Biology; Engineering","score_opus":0.4706248224375595,"score_gpt":0.5890246793657072,"score_spread":0.11839985692814775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3020197475","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.86963075,0.00021198446,0.0021305499,0.077694185,0.0004110751,0.00093008863,0.0000029298742,0.00002579192,0.048962638],"genre_scores_gemma":[0.9939139,0.000010993906,0.00021109692,0.0056004,0.00015977872,0.000057573816,0.000010209631,0.000004281208,0.000031793596],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.992666,0.001463494,0.00053505384,0.00031193008,0.004875513,0.00014800612],"domain_scores_gemma":[0.99841696,0.0005746899,0.00015630231,0.00034349237,0.00045728445,0.000051275198],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.022130672,0.000083176004,0.0001222252,0.000104978906,0.00011860603,0.00016696916,0.0005318873,0.000036730922,0.008981529],"category_scores_gemma":[0.0050010183,0.000047837697,0.00005152376,0.0012963549,0.000044363584,0.00038888564,0.000034640423,0.00012642698,0.001813462],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00007130217,0.00014928132,0.15976445,0.0000054643765,0.000011594583,0.0000027001577,0.019898431,0.023159439,0.0004392241,0.0050968975,0.06141278,0.72998846],"study_design_scores_gemma":[0.0008116145,0.00012036655,0.34387082,0.0000043045447,0.000017072876,0.0000013464153,0.0031195544,0.6204537,0.00019345488,0.015438916,0.015872426,0.00009642368],"about_ca_topic_score_codex":0.000007869057,"about_ca_topic_score_gemma":0.00002358612,"teacher_disagreement_score":0.729892,"about_ca_system_score_codex":0.000059039667,"about_ca_system_score_gemma":0.00031491774,"threshold_uncertainty_score":0.9989637},"labels":[],"label_agreement":null},{"id":"W3095230013","doi":"10.1177/1356389020952462","title":"Evaluation for planetary health","year":2020,"lang":"en","type":"article","venue":"Evaluation","topic":"Health Systems, Economic Evaluations, Quality of Life","field":"Economics, Econometrics and Finance","cited_by":23,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Dialogic; Psychological intervention; Field (mathematics); Management science; Human health; Natural (archaeology); Work (physics); Engineering ethics; Sociology; Computer science; Knowledge management; Psychology; Engineering; Medicine; Geography","score_opus":0.6745489268580942,"score_gpt":0.510194529002474,"score_spread":0.16435439785562023,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3095230013","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.0695763,0.018596621,0.12011585,0.7610228,0.0025427924,0.015000485,0.0012861202,0.00030753503,0.01155151],"genre_scores_gemma":[0.91248226,0.000071344715,0.009649054,0.07295333,0.0017123094,0.0012781438,0.0016944631,0.000059041417,0.00010007858],"study_design_codex":"not_applicable","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99430877,0.0010156828,0.0031586052,0.0006800091,0.0004598142,0.00037709152],"domain_scores_gemma":[0.9964787,0.0005902127,0.0019906997,0.00033366753,0.00035023325,0.00025651048],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.06950216,0.0001738053,0.0006676952,0.00016150894,0.00024807657,0.00006725151,0.00019474672,0.00010891543,0.0018806664],"category_scores_gemma":[0.012598245,0.00023470339,0.000116134106,0.00020295702,0.000025417194,0.00048109124,0.000021384332,0.00011127088,0.0023645575],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00012896,0.00021889768,0.010145408,0.001125237,0.00035228403,2.151368e-7,0.016440911,0.027659692,0.000026537806,0.09049026,0.7956921,0.057719495],"study_design_scores_gemma":[0.002416734,0.0003049995,0.013845842,0.000035342546,0.000040460283,0.000001215823,0.0005822197,0.861375,0.000007882184,0.04395243,0.07713798,0.00029989745],"about_ca_topic_score_codex":0.0002001344,"about_ca_topic_score_gemma":0.00007653697,"teacher_disagreement_score":0.84290594,"about_ca_system_score_codex":0.001018502,"about_ca_system_score_gemma":0.0007919565,"threshold_uncertainty_score":0.9990317},"labels":[],"label_agreement":null},{"id":"W3100567947","doi":"10.1177/1356389020969714","title":"Applying the eight principles of developmental evaluation: Complex Care Hub as a case example","year":2020,"lang":"en","type":"article","venue":"Evaluation","topic":"Geriatric Care and Nursing Homes","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Calgary; Alberta Health Services","funders":"Alberta Health Services","keywords":"Context (archaeology); Health care; Psychology; Nursing; Medical education; Medicine; Political science","score_opus":0.33880428037014854,"score_gpt":0.46695976281898927,"score_spread":0.12815548244884073,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3100567947","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.963082,0.0009783532,0.00024195341,0.0022756117,0.0006167031,0.006779311,0.0000392503,0.00006159301,0.02592522],"genre_scores_gemma":[0.99595785,0.000006586933,0.0010433953,0.00044939254,0.00034726484,0.0017157055,0.0003805242,0.000016113769,0.000083176354],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9970916,0.0010244678,0.0004893175,0.00024342212,0.0009430657,0.00020816228],"domain_scores_gemma":[0.9981916,0.00035428067,0.00027926257,0.00019161668,0.00090934115,0.000073899835],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001990613,0.000117215495,0.00018188676,0.000050261395,0.0008095994,0.000011389939,0.00012485313,0.00008409953,0.0036842145],"category_scores_gemma":[0.0006785116,0.00008925438,0.000054833796,0.0003355626,0.000038986134,0.0000834987,0.00008196729,0.00018804178,0.00021340964],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00030055738,0.00009033765,0.017380955,0.00074873434,0.00016128637,0.00002787653,0.5420187,0.0012675737,0.00497896,0.0018077946,0.02727435,0.40394285],"study_design_scores_gemma":[0.0062141526,0.00035970806,0.08136202,0.00026924477,0.0009798006,0.000118925025,0.6239989,0.045604125,0.0015662119,0.00094072084,0.23802009,0.0005661307],"about_ca_topic_score_codex":0.0007415551,"about_ca_topic_score_gemma":0.00045455332,"teacher_disagreement_score":0.40337673,"about_ca_system_score_codex":0.00040570786,"about_ca_system_score_gemma":0.0013271887,"threshold_uncertainty_score":0.99722654},"labels":[],"label_agreement":null},{"id":"W3109011900","doi":"10.1177/1356389020969721","title":"How to normalize reflexive evaluation? Navigating between legitimacy and integrity","year":2020,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":13,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Athena Sustainable Materials Institute","funders":"","keywords":"Reflexivity; Normalization (sociology); Legitimacy; Epistemology; Process management; Political science; Sociology; Computer science; Knowledge management; Business; Social science; Politics","score_opus":0.4709169939777773,"score_gpt":0.5835940233142233,"score_spread":0.11267702933644597,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3109011900","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.87594014,0.00012315765,0.01538944,0.09956587,0.00027194584,0.0017560774,0.000024250947,0.000062961015,0.0068661342],"genre_scores_gemma":[0.9910709,0.00000512286,0.0054757413,0.00252878,0.000520779,0.00016720436,0.00007833718,0.0000127523235,0.00014035469],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99123955,0.0013454193,0.0007007883,0.0007060523,0.005751748,0.00025646584],"domain_scores_gemma":[0.9953752,0.00083984074,0.0003654601,0.00039804343,0.0026971472,0.0003242723],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.021389624,0.0001893298,0.00030038232,0.0001226058,0.0003133693,0.0008232196,0.0003897599,0.00008834487,0.0018289536],"category_scores_gemma":[0.015351594,0.00015441205,0.0000683358,0.0010590411,0.000055660596,0.0013333109,0.00017090337,0.00035467767,0.00048671014],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000035866626,0.000017364478,0.0557517,0.000008186939,0.000024232175,3.7058626e-7,0.007895164,0.0016599097,0.00082665664,0.0006424487,0.0048118946,0.9283262],"study_design_scores_gemma":[0.0019677796,0.0004507978,0.29791245,0.00006506221,0.00018860192,0.000002531093,0.0058112303,0.6530857,0.0019115765,0.019898057,0.01827625,0.0004299731],"about_ca_topic_score_codex":0.000023655692,"about_ca_topic_score_gemma":0.000034373235,"teacher_disagreement_score":0.92789626,"about_ca_system_score_codex":0.00018705009,"about_ca_system_score_gemma":0.00052497035,"threshold_uncertainty_score":0.9990835},"labels":[],"label_agreement":null},{"id":"W3139167773","doi":"10.1177/1356389020978501","title":"Developing an ethical rationale for collaborative approaches to evaluation","year":2021,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":14,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; University of Victoria","funders":"","keywords":"Reflexivity; Dialogic; Reciprocity (cultural anthropology); Inclusion (mineral); Sociology; Engineering ethics; Context (archaeology); Politics; Set (abstract data type); Epistemology; Political science; Computer science; Pedagogy; Social science; Engineering; Law","score_opus":0.8319796326770158,"score_gpt":0.5935660075730588,"score_spread":0.23841362510395703,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3139167773","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.39298645,0.00021336725,0.49785814,0.088489756,0.0013642856,0.0054619038,0.000051786414,0.00006025925,0.013514049],"genre_scores_gemma":[0.8962305,0.0000042983092,0.09628579,0.0036638852,0.0003477308,0.0019804297,0.0007365388,0.000015261294,0.000735526],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9908017,0.002177653,0.0007442344,0.00070348685,0.0053600417,0.00021289944],"domain_scores_gemma":[0.9890243,0.0012880238,0.00024423926,0.0004926798,0.00882321,0.00012758622],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.03849379,0.00014313185,0.00021656291,0.00023321777,0.00041523846,0.00045629634,0.00026152912,0.00017043752,0.0027223597],"category_scores_gemma":[0.01617557,0.00012519157,0.00006461582,0.0015033045,0.000029771461,0.0007276449,0.000055702785,0.00014580127,0.0003018478],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00014234944,0.00016497994,0.001168629,0.000012098754,0.000047200963,3.4375728e-7,0.013327864,0.12036316,0.00223159,0.2259679,0.0102342125,0.6263397],"study_design_scores_gemma":[0.0013444033,0.0001508736,0.019606424,0.000017027838,0.00006712306,0.0000022470026,0.007813538,0.7781933,0.006716229,0.16035031,0.025492057,0.00024649446],"about_ca_topic_score_codex":0.0000019457038,"about_ca_topic_score_gemma":0.0003546394,"teacher_disagreement_score":0.6578301,"about_ca_system_score_codex":0.00048005217,"about_ca_system_score_gemma":0.0055643823,"threshold_uncertainty_score":0.9981893},"labels":[],"label_agreement":null},{"id":"W4220896026","doi":"10.1177/13563890221086306","title":"Contribution of practical wisdom to resolving ethical issues","year":2022,"lang":"en","type":"article","venue":"Evaluation","topic":"Ethics in medical practice","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Engineering ethics; Action (physics); Face (sociological concept); Ethical issues; Epistemology; Management science; Sociology; Psychology; Social science; Economics; Philosophy; Engineering","score_opus":0.22857419154794673,"score_gpt":0.6319014456775595,"score_spread":0.40332725412961273,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4220896026","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.23581381,0.00039578718,0.025064588,0.6995928,0.0040647355,0.0040298286,0.00002727657,0.00015990398,0.030851262],"genre_scores_gemma":[0.9838829,0.000011879561,0.004733655,0.010109261,0.00036193608,0.00047352287,0.000054410215,0.00001378008,0.0003586562],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","domain_scores_codex":[0.97595745,0.020186523,0.00065919367,0.00026490228,0.0026354592,0.00029644754],"domain_scores_gemma":[0.9836323,0.014183693,0.00035849505,0.00031549815,0.0013882645,0.0001217205],"candidate_categories":["metaresearch","sts","research_integrity","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.060678046,0.00007254013,0.00018343677,0.00009306802,0.0013854047,0.000004675782,0.00013776128,0.0003083112,0.008353434],"category_scores_gemma":[0.22614668,0.00007404418,0.00003303874,0.00035586327,0.000078712495,0.00012404154,0.00030194997,0.0063054147,0.00031066433],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012804974,0.0004115839,0.0054439916,0.00016563611,0.00007291382,0.000010282327,0.03596713,0.0012336635,0.015425528,0.650818,0.28294426,0.006226486],"study_design_scores_gemma":[0.004401531,0.0012450852,0.027856234,0.00034732046,0.00041175805,0.000012947654,0.034238968,0.047838878,0.0010348206,0.2831739,0.59899,0.00044852035],"about_ca_topic_score_codex":0.00016764329,"about_ca_topic_score_gemma":0.000054300803,"teacher_disagreement_score":0.7480691,"about_ca_system_score_codex":0.0005519039,"about_ca_system_score_gemma":0.0010079496,"threshold_uncertainty_score":0.99991465},"labels":[],"label_agreement":null},{"id":"W4232707640","doi":"10.1177/1356389015580672","title":"Canadian Evaluation Society 36th Conference 2015: <i>Evaluation for the world we want</i>","year":2015,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Political science; Engineering ethics; Environmental ethics; Engineering; Philosophy","score_opus":0.5302600051866425,"score_gpt":0.5674058189989668,"score_spread":0.03714581381232429,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4232707640","genre_codex":"commentary","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13269062,0.008225206,0.099003814,0.52670425,0.020051008,0.040938277,0.00040924747,0.00032022138,0.17165737],"genre_scores_gemma":[0.988943,0.00005232011,0.0016921755,0.0015171826,0.0005712643,0.0019274878,0.00034780856,0.000024450641,0.004924334],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9848996,0.0018146978,0.0010592225,0.00075406546,0.010949274,0.00052318414],"domain_scores_gemma":[0.98264664,0.0016269572,0.00062838773,0.0010486497,0.013670123,0.00037922416],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.10284872,0.00027745165,0.000293268,0.00039046168,0.0007794687,0.00073437643,0.0008270883,0.00014125841,0.008080318],"category_scores_gemma":[0.0058870814,0.00018920278,0.00022603494,0.0014449529,0.00010658174,0.0011519067,0.00006524509,0.00022631479,0.00093806477],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000047078054,0.00004300964,0.0014455611,0.0000043595796,0.000052160398,5.803412e-8,0.004731103,0.04406756,0.000070555434,0.0019658764,0.43225658,0.5153161],"study_design_scores_gemma":[0.0022282267,0.00008234342,0.0055739363,0.000020361025,0.00032755276,0.00000212914,0.0041039926,0.81583184,0.00017250289,0.06983407,0.101601176,0.00022184412],"about_ca_topic_score_codex":0.0031160417,"about_ca_topic_score_gemma":0.19679792,"teacher_disagreement_score":0.8562524,"about_ca_system_score_codex":0.0023562675,"about_ca_system_score_gemma":0.013129653,"threshold_uncertainty_score":0.99983984},"labels":[],"label_agreement":null},{"id":"W4236952314","doi":"10.1177/1356389014543594","title":"The 35 <sup>th</sup> Canadian Evaluation Society Annual Conference: ‘Celebrating Contributions to Canadian Evaluation’","year":2014,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Political science; Library science; Computer science","score_opus":0.11994209496311921,"score_gpt":0.48091336084166886,"score_spread":0.36097126587854966,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4236952314","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.6864747,0.00060153333,0.01442343,0.1682095,0.003862756,0.0136148995,0.0009919959,0.00017016236,0.11165103],"genre_scores_gemma":[0.991741,0.000013114426,0.0006584528,0.0040244334,0.00066621415,0.0011583513,0.0006346882,0.000026398626,0.0010773398],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.98186386,0.0044854693,0.0014767008,0.0009888655,0.010090612,0.0010944754],"domain_scores_gemma":[0.97303504,0.0017792442,0.00051471824,0.0012894372,0.022234645,0.0011469338],"candidate_categories":["metaresearch","metaepi_narrow","sts","scholarly_communication","insufficient_payload"],"consensus_categories":["metaresearch","insufficient_payload"],"category_scores_codex":[0.10661256,0.0003648871,0.00036650998,0.0006108574,0.0036310782,0.0015279854,0.0010581884,0.0002782805,0.010138531],"category_scores_gemma":[0.03579605,0.0002818306,0.00022382228,0.0019161478,0.000148187,0.0011237062,0.00008084836,0.00041344052,0.0022907138],"study_design_candidate":"simulation_or_modeling","study_design_consensus":null,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.000017496859,0.000044415472,0.0041008443,0.0000031463162,0.00007685674,1.9599233e-7,0.012221045,0.19990969,0.000083395666,0.009469066,0.18569079,0.5883831],"study_design_scores_gemma":[0.001173952,0.00013273885,0.030391058,0.000025156316,0.00016740515,0.0000028963184,0.007012899,0.82836527,0.00005420078,0.018017689,0.11434115,0.00031560884],"about_ca_topic_score_codex":0.13999821,"about_ca_topic_score_gemma":0.8704617,"teacher_disagreement_score":0.7304635,"about_ca_system_score_codex":0.0050721145,"about_ca_system_score_gemma":0.024638223,"threshold_uncertainty_score":0.9999634},"labels":[],"label_agreement":null},{"id":"W4236954017","doi":"10.1177/135638900300900406","title":"Evaluability Assessment: A Tool for Incorporating Evaluation in Social Change Programmes","year":2003,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":34,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Université de Montréal; University of Calgary","funders":"","keywords":"Participatory evaluation; Accountability; Process management; Promotion (chess); Process (computing); Social change; Monitoring and evaluation; Theory of change; Citizen journalism; Plan (archaeology); Political science; Theme (computing); Politics; Public relations; Management science; Knowledge management; Sociology; Business; Engineering; Computer science; Public administration","score_opus":0.4539931329436821,"score_gpt":0.5881362452183663,"score_spread":0.13414311227468417,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4236954017","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9666018,0.00014199883,0.008847391,0.0023480614,0.0009464109,0.009916979,0.000018141756,0.000057110727,0.011122104],"genre_scores_gemma":[0.9837788,0.0000022154602,0.008671657,0.00028212252,0.00028671673,0.0067166216,0.00014313804,0.000017798486,0.00010098233],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9859606,0.0043205954,0.0015095995,0.00087435526,0.006894502,0.0004403371],"domain_scores_gemma":[0.9941632,0.0010792287,0.0007343861,0.0004998351,0.0034611674,0.00006217799],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.11808656,0.00022942794,0.0003616562,0.00047265718,0.00048393814,0.0003664751,0.0003139536,0.00016008824,0.0024369794],"category_scores_gemma":[0.013360029,0.00019872378,0.00016141747,0.0014786106,0.00007076483,0.0013693722,0.00004091121,0.00018983333,0.000081349084],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00005073347,0.00034721973,0.14789762,0.000021944626,0.000016173582,1.5099911e-7,0.0032805898,0.00228102,0.00026125592,0.020239754,0.00075291377,0.8248506],"study_design_scores_gemma":[0.0026764646,0.000147888,0.36203307,0.000015061472,0.000064256164,8.96538e-7,0.00092909427,0.4717394,0.00010413649,0.15933119,0.00274651,0.00021201151],"about_ca_topic_score_codex":0.000022984923,"about_ca_topic_score_gemma":0.00032365287,"teacher_disagreement_score":0.8246386,"about_ca_system_score_codex":0.0009796759,"about_ca_system_score_gemma":0.0018175447,"threshold_uncertainty_score":0.99847496},"labels":[],"label_agreement":null},{"id":"W4285595442","doi":"10.1177/13563890221107044","title":"A theory-based approach to designing interventions for Planetary Health","year":2022,"lang":"en","type":"article","venue":"Evaluation","topic":"Climate Change and Health Impacts","field":"Environmental Science","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Psychological intervention; Action (physics); Management science; Theory of change; Engineering ethics; Field (mathematics); Human health; Public relations; Psychology; Computer science; Sociology; Political science; Medicine; Engineering","score_opus":0.285646917889852,"score_gpt":0.4283617954686286,"score_spread":0.1427148775787766,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4285595442","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.13564824,0.0004284053,0.8358927,0.01201845,0.00048582637,0.0074872435,0.00026075874,0.000116610114,0.0076617687],"genre_scores_gemma":[0.97432476,0.000001272754,0.017321218,0.005893349,0.000047071673,0.0014726628,0.0007464048,0.0000127307785,0.00018054062],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99885947,0.00030451128,0.00015544432,0.00016469526,0.00032229128,0.00019357482],"domain_scores_gemma":[0.9996346,0.00008288798,0.000072286995,0.00010842858,0.000006057993,0.000095699965],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.004045608,0.0000504266,0.000072942166,0.000048419708,0.0004224768,0.000010751442,0.00008226652,0.0000100237685,0.003145109],"category_scores_gemma":[0.00009355352,0.00005216978,0.00004111558,0.0001492916,0.00000884565,0.000046997135,0.00004698201,0.000052246243,0.00006341536],"study_design_candidate":"simulation_or_modeling","study_design_consensus":"simulation_or_modeling","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034640735,0.000738608,0.003975273,0.0003200574,0.000011786518,3.3081574e-7,0.01079228,0.61118114,0.0012117672,0.000951214,0.13793205,0.23253909],"study_design_scores_gemma":[0.0034479764,0.0034469557,0.08286986,0.00015376916,0.000104817394,0.000009457073,0.0040107504,0.8612765,0.00024476845,0.020539897,0.023377165,0.00051808235],"about_ca_topic_score_codex":0.00008023588,"about_ca_topic_score_gemma":0.00003639471,"teacher_disagreement_score":0.8386765,"about_ca_system_score_codex":0.0004819632,"about_ca_system_score_gemma":0.000041847772,"threshold_uncertainty_score":0.99776614},"labels":[],"label_agreement":null},{"id":"W4319439088","doi":"10.1177/13563890221149452","title":"The use of mediation analysis in evaluation of complex health interventions","year":2023,"lang":"en","type":"article","venue":"Evaluation","topic":"Health Policy Implementation Science","field":"Health Professions","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"McMaster University","funders":"Economic and Social Research Council; Bill and Melinda Gates Foundation","keywords":"Mediation; Causal inference; Intervention (counseling); Context (archaeology); Psychological intervention; Randomized controlled trial; Quality (philosophy); Psychology; Inference; Applied psychology; Medicine; Nursing; Computer science; Sociology; Social science","score_opus":0.9517514639979373,"score_gpt":0.7979085226517668,"score_spread":0.1538429413461705,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4319439088","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9763778,0.00004833482,0.0024822648,0.01712788,0.000377259,0.0032538283,0.000102873426,0.000033923687,0.00019587482],"genre_scores_gemma":[0.998071,0.0000717038,0.0003795004,0.00049915625,0.000035821617,0.00054676575,0.00030988702,0.0000066292528,0.000079518635],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9855596,0.009487711,0.0021124734,0.00021211703,0.0022539548,0.0003741533],"domain_scores_gemma":[0.99246633,0.0038842706,0.0014434842,0.0003684872,0.001770612,0.00006682647],"candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.051974718,0.00007003259,0.0002673905,0.0009312326,0.00042993829,0.000007936819,0.00014975635,0.000052563333,0.0007519985],"category_scores_gemma":[0.013277061,0.000057568373,0.00010293028,0.0046960027,0.00006837781,0.00026347314,0.000053824355,0.00013920834,0.000070564405],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006652805,0.00013028135,0.4931772,0.0007248699,0.00025111425,7.234584e-8,0.06450524,0.16677411,0.00084659574,0.003822561,0.043644544,0.22605687],"study_design_scores_gemma":[0.0005854354,0.00002911665,0.62199265,0.000048408667,0.000079188525,1.5074779e-8,0.0018540057,0.37380522,0.0000060873667,0.0006883151,0.0008844316,0.000027109334],"about_ca_topic_score_codex":0.0014322314,"about_ca_topic_score_gemma":0.017723378,"teacher_disagreement_score":0.22602975,"about_ca_system_score_codex":0.00066293374,"about_ca_system_score_gemma":0.0017982272,"threshold_uncertainty_score":0.9950345},"labels":[],"label_agreement":null},{"id":"W4322744890","doi":"10.1177/13563890231156954","title":"How can climate change and its interaction with other compounding risks be considered in evaluation? Experiences from Vietnam","year":2023,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":7,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Waterloo; University of Alberta; University of Guelph","funders":"Australian Centre for International Agricultural Research; Canadian Institutes of Health Research; Consortium of International Agricultural Research Centers","keywords":"Climate change; Context (archaeology); Environmental resource management; Environmental planning; Public relations; Political science; Business; Psychology; Geography; Environmental science; Ecology","score_opus":0.6069973684791174,"score_gpt":0.5550850423223465,"score_spread":0.051912326156770994,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4322744890","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.98864734,0.00021124794,0.00006017375,0.008013335,0.0006210676,0.0013147092,0.000025868892,0.000050128667,0.001056123],"genre_scores_gemma":[0.9979693,0.000059622038,0.00016088401,0.00041839213,0.00013970585,0.001066888,0.00010029376,0.000013297775,0.00007157778],"study_design_codex":"observational","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9942534,0.0010937816,0.00047277985,0.0005499427,0.0033711542,0.00025895282],"domain_scores_gemma":[0.9976649,0.00077997963,0.00039332622,0.0002911684,0.00079598324,0.0000746429],"candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.012224016,0.00016667346,0.00023838088,0.00054199743,0.0002643324,0.0005525851,0.0001716471,0.000071229595,0.0017989558],"category_scores_gemma":[0.0014324833,0.00012541586,0.000030339063,0.0010773874,0.00005479471,0.0013700718,0.000063193,0.00012861073,0.000101135876],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00034267554,0.00016150929,0.42566052,0.000023309529,0.000073409195,0.000005096957,0.20220369,0.007643127,0.004074518,0.00093844725,0.0012696432,0.35760403],"study_design_scores_gemma":[0.0013632437,0.00008871122,0.29396734,0.00007427349,0.000038131624,0.0000021937772,0.06110349,0.64018095,0.00041135814,0.0018695529,0.00073482445,0.00016592156],"about_ca_topic_score_codex":0.00019236241,"about_ca_topic_score_gemma":0.004061178,"teacher_disagreement_score":0.63253784,"about_ca_system_score_codex":0.0002063037,"about_ca_system_score_gemma":0.00020057004,"threshold_uncertainty_score":0.99911356},"labels":[],"label_agreement":null},{"id":"W4382811084","doi":"10.1177/13563890231185167","title":"How to develop a realist programme theory using Margaret Archer’s structure–agency–culture framework: The case of adolescent accountability for sexual and reproductive health in urban resource-constrained settings","year":2023,"lang":"en","type":"article","venue":"Evaluation","topic":"Social Policy and Reform Studies","field":"Social Sciences","cited_by":9,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"Fonds Wetenschappelijk Onderzoek","keywords":"Structure and agency; Agency (philosophy); Sociology; Critical realism (philosophy of perception); Reproductive health; Accountability; Context (archaeology); Resource (disambiguation); Epistemology; Social science; Political science; Realism; Population","score_opus":0.12633726763033404,"score_gpt":0.4286265921107111,"score_spread":0.30228932448037704,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4382811084","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95100325,0.00016923952,0.0003047844,0.044924606,0.000120024146,0.0032346374,0.00012444069,0.000042021995,0.000077008386],"genre_scores_gemma":[0.99737775,0.000011840701,0.0015224009,0.00040640394,0.0004958624,0.00009324743,0.000027241917,0.000008814497,0.000056431603],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.99775237,0.0010527652,0.0002364363,0.00032190804,0.0002922316,0.00034430547],"domain_scores_gemma":[0.99884033,0.00033731613,0.00019008576,0.00015795059,0.00041501224,0.000059331283],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.010278387,0.000102320955,0.00018221678,0.00006351496,0.00081277586,0.000056618872,0.00009183773,0.0001102141,0.0000019669858],"category_scores_gemma":[0.011524854,0.00007003339,0.000024247975,0.000971991,0.00041520674,0.00008644098,0.000059315244,0.00016604457,9.775976e-8],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00010843672,0.000029089644,0.0021923226,0.00012337841,0.000036690464,0.0000011812483,0.80608636,0.00003084126,0.00009882263,0.01428701,0.001271476,0.17573436],"study_design_scores_gemma":[0.0004926591,0.00013917215,0.056801174,0.00012361522,0.00006180759,0.000007653717,0.6854084,0.00050651334,0.000066689776,0.25447312,0.001686366,0.00023280918],"about_ca_topic_score_codex":0.002992732,"about_ca_topic_score_gemma":0.010758481,"teacher_disagreement_score":0.24018611,"about_ca_system_score_codex":0.0003197999,"about_ca_system_score_gemma":0.0005342268,"threshold_uncertainty_score":0.9968015},"labels":[],"label_agreement":null},{"id":"W4389600041","doi":"10.1177/13563890231207115","title":"Interorganizational evaluation capacity building in the public, health and community sectors","year":2023,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"University of Ottawa","funders":"","keywords":"Capacity building; Work (physics); Business; Public sector; Public relations; Joint (building); Knowledge management; Political science; Economic growth; Economics; Computer science; Engineering","score_opus":0.6482669341035596,"score_gpt":0.5566919701817297,"score_spread":0.09157496392182995,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4389600041","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9787655,0.000053296244,0.0013916023,0.017591268,0.0003688876,0.0008044613,0.000005759971,0.000037282414,0.0009819315],"genre_scores_gemma":[0.9980484,0.000018916431,0.00024897474,0.0013174136,0.00006584506,0.00010835988,0.00015255022,0.0000068437334,0.000032668922],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9819601,0.011599226,0.00073426455,0.00029529235,0.00514667,0.00026440824],"domain_scores_gemma":[0.9960219,0.0020249512,0.00028661243,0.00045423702,0.0011335921,0.00007870915],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.13825716,0.00010317632,0.00014684604,0.0005444381,0.00074288796,0.00037753413,0.00042742686,0.000049805996,0.0009792569],"category_scores_gemma":[0.015215927,0.00007137193,0.000028885903,0.0025190008,0.00006713567,0.00072711764,0.00009688259,0.00033907578,0.00016766952],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000017224129,0.00028075426,0.36624292,0.000022278027,0.000024382842,2.886971e-7,0.12129401,0.006656781,0.0004019657,0.011059261,0.016398873,0.4776013],"study_design_scores_gemma":[0.00052227866,0.00007558499,0.5369964,0.0000115540315,0.000005485696,0.000003178308,0.012528984,0.42126077,0.000015589412,0.026414683,0.0020947664,0.00007069174],"about_ca_topic_score_codex":0.00009899289,"about_ca_topic_score_gemma":0.0009051754,"teacher_disagreement_score":0.47753057,"about_ca_system_score_codex":0.0003440436,"about_ca_system_score_gemma":0.0008450732,"threshold_uncertainty_score":0.99993396},"labels":[],"label_agreement":null},{"id":"W4391389032","doi":"10.1177/13563890241227433","title":"Planetary health: Creating rapid impact assessment tools","year":2024,"lang":"en","type":"article","venue":"Evaluation","topic":"Climate Change and Health Impacts","field":"Environmental Science","cited_by":4,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Future Earth; University of Victoria","funders":"","keywords":"Environmental planning; Environmental science","score_opus":0.17725812000104163,"score_gpt":0.4650921412998606,"score_spread":0.28783402129881897,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4391389032","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9333403,0.0024217793,0.00040233118,0.0068833106,0.00064367446,0.0011321165,0.000060464172,0.00017001109,0.054945983],"genre_scores_gemma":[0.99723697,0.00022140284,0.00083302613,0.00092337566,0.00022640347,0.000039277707,0.00040259343,0.0000127681005,0.00010418772],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.998586,0.00014871852,0.00021302393,0.00021666849,0.00055050687,0.0002850819],"domain_scores_gemma":[0.99953943,0.00011578907,0.000052857173,0.00013900304,0.0000064671895,0.0001464239],"candidate_categories":["insufficient_payload"],"consensus_categories":["insufficient_payload"],"category_scores_codex":[0.002159669,0.00009354243,0.00010398804,0.0000397567,0.00015164784,0.00010471897,0.00005779441,0.00003573826,0.025697878],"category_scores_gemma":[0.00006806966,0.0000761088,0.000038194503,0.00017219066,0.000017362301,0.00037015355,0.000034080203,0.00011269639,0.0011591322],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000008929371,0.000035727568,0.012405724,0.000073412295,0.000010551463,0.0000044629073,0.0026806202,0.0030192104,0.00083539373,0.00004350806,0.071619175,0.9092633],"study_design_scores_gemma":[0.00023031914,0.0003526102,0.7356357,0.00015124312,0.000022140148,0.000013901999,0.00015782095,0.25540048,0.00004241902,0.0008334493,0.007025675,0.00013425603],"about_ca_topic_score_codex":0.0006441802,"about_ca_topic_score_gemma":0.00012054,"teacher_disagreement_score":0.909129,"about_ca_system_score_codex":0.001024102,"about_ca_system_score_gemma":0.000120709876,"threshold_uncertainty_score":0.9996186},"labels":[],"label_agreement":null},{"id":"W4401276261","doi":"10.1177/13563890241265859","title":"Post-truth and pathways for evaluators","year":2024,"lang":"en","type":"article","venue":"Evaluation","topic":"Viral Infections and Outbreaks Research","field":"Medicine","cited_by":6,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Victoria","funders":"","keywords":"Disinformation; Transformative learning; Skepticism; Post truth; Politics; Phenomenon; Narrative; Fake news; Sociology; Premise; Political science; Public relations; Epistemology; Social media; Law; Media studies","score_opus":0.0724581220496297,"score_gpt":0.4204130039843808,"score_spread":0.3479548819347511,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4401276261","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9882017,0.002002641,0.0016447824,0.0017975856,0.0004250094,0.0011237899,0.00002264626,0.00008774912,0.0046940898],"genre_scores_gemma":[0.99845797,0.000035134864,0.00031499838,0.0000900956,0.00024691655,0.0001634519,0.000079099176,0.000012817295,0.0005995122],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9992577,0.00004231968,0.00009980902,0.00016749695,0.00032420197,0.000108477536],"domain_scores_gemma":[0.99943215,0.0001102061,0.000009107429,0.000079576464,0.0003106681,0.00005830261],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0012930207,0.00005444,0.00007406735,0.0001336657,0.00007002419,0.00005171041,0.000012794575,0.000044024076,0.00033781576],"category_scores_gemma":[0.00044100016,0.000042068314,0.000045645447,0.0001305541,0.000021994576,0.00008850563,0.000011733312,0.000071514005,0.00006716524],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000117164,0.00005730118,0.0006183671,0.00031891983,0.000085925734,0.00000388722,0.0005471959,0.000016660195,0.0633704,0.009785876,0.0028846064,0.9221937],"study_design_scores_gemma":[0.0029094575,0.002383613,0.059738904,0.0002941301,0.00070428435,0.00013233138,0.00049500173,0.8774742,0.017067116,0.016584704,0.021977732,0.00023851283],"about_ca_topic_score_codex":0.000014975769,"about_ca_topic_score_gemma":0.000004314631,"teacher_disagreement_score":0.92195517,"about_ca_system_score_codex":0.000084406885,"about_ca_system_score_gemma":0.00022529675,"threshold_uncertainty_score":0.36988455},"labels":[],"label_agreement":null},{"id":"W4404873315","doi":"10.1177/13563890241289937","title":"Artificial intelligence and big data-driven evaluation research and practices: A systematic literature review","year":2024,"lang":"en","type":"article","venue":"Evaluation","topic":"Artificial Intelligence in Healthcare and Education","field":"Medicine","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"École Nationale d'Administration Publique","funders":"","keywords":"Big data; Systematic review; Data science; Management science; Psychology; Political science; Engineering ethics; Computer science; MEDLINE; Engineering; Data mining","score_opus":0.7863038003097949,"score_gpt":0.6347017197247322,"score_spread":0.15160208058506264,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4404873315","genre_codex":"review","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.092129566,0.8570651,0.0017899356,0.03568389,0.0019334183,0.010491723,0.000025248457,0.00011083315,0.0007703073],"genre_scores_gemma":[0.92738026,0.069392115,0.0004575901,0.0003185905,0.0011889407,0.0005365992,0.00061115157,0.000020061576,0.000094705116],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.99604267,0.0011106428,0.00065997423,0.0005814983,0.0014028265,0.0002024096],"domain_scores_gemma":[0.9961234,0.0011180784,0.00018431965,0.0006118819,0.0018414836,0.00012086837],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.020490304,0.00011985505,0.000251804,0.00030525762,0.00018437122,0.00028833657,0.00010091278,0.00011681102,0.00010430493],"category_scores_gemma":[0.014636051,0.00009511757,0.000020300733,0.0009663841,0.00009481312,0.0005338216,0.00007143068,0.000413344,0.00010206124],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000043573807,0.0000900688,0.00013307983,0.12236121,0.000055643875,0.000009223583,0.004534952,0.000009416062,0.00024699196,0.0013564399,0.0022038098,0.8689556],"study_design_scores_gemma":[0.000024145687,0.00049203646,0.00031565834,0.19571333,0.0014035371,0.00029514448,0.003087177,0.7681053,0.00050561206,0.028908474,0.00096055627,0.00018902475],"about_ca_topic_score_codex":0.0000676982,"about_ca_topic_score_gemma":0.000103686354,"teacher_disagreement_score":0.86876655,"about_ca_system_score_codex":0.0002062029,"about_ca_system_score_gemma":0.00083001924,"threshold_uncertainty_score":0.9936641},"labels":[],"label_agreement":null},{"id":"W4417434618","doi":"10.1177/13563890251395056","title":"Unraveling the complexities of learning in community development evaluation","year":2025,"lang":"en","type":"article","venue":"Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Ottawa; University of Victoria","funders":"","keywords":"Neoliberalism (international relations); Community development; Focus (optics); Learning community; Key (lock)","score_opus":0.42157615803294757,"score_gpt":0.5579690605544919,"score_spread":0.13639290252154435,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4417434618","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"codex-gemma-dda1882f352a","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.95451057,0.00019094293,0.005377234,0.0015573854,0.00029190467,0.0007068303,5.7461085e-7,0.000013354647,0.03735122],"genre_scores_gemma":[0.9985002,0.000004754017,0.00061782345,0.00015479901,0.000016140611,0.000111831985,0.000030444955,0.00000300246,0.000560992],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9904492,0.0048194914,0.0009976852,0.0001783032,0.003415149,0.0001401655],"domain_scores_gemma":[0.9955357,0.0019774619,0.0003388283,0.00039182996,0.0017412308,0.000014971858],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.07685577,0.00008753845,0.00017055348,0.0004170769,0.0004758031,0.00009192168,0.0004231305,0.00004472172,0.0012217795],"category_scores_gemma":[0.006692639,0.000059756727,0.00003903186,0.0010754216,0.00007520949,0.00023545219,0.0001068656,0.00029811624,0.00005459689],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000318763,0.000097460674,0.075849764,0.000014891731,0.000020006373,2.6345841e-8,0.01805038,0.19258171,0.0005148027,0.006109304,0.00045888353,0.7062709],"study_design_scores_gemma":[0.0005510733,0.000020486954,0.47523436,0.000055435175,0.000017694823,1.3705703e-7,0.012632826,0.47916502,0.0010437927,0.029435996,0.0017940751,0.00004908723],"about_ca_topic_score_codex":0.00010298693,"about_ca_topic_score_gemma":0.001097403,"teacher_disagreement_score":0.7062218,"about_ca_system_score_codex":0.00029723707,"about_ca_system_score_gemma":0.0013660745,"threshold_uncertainty_score":0.99969125},"labels":[],"label_agreement":null}]}