{"meta":{"query_hash":"7dc0bf08a709","filters":{"venue":"Evaluation Journal of Australasia"},"cohort_total":19,"direct_labels_cover":1,"predictions_cover":19,"exported":19,"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/7dc0bf08a709","api":"https://metacan.xera.ac/api/v1/cohort?venue=Evaluation+Journal+of+Australasia"},"results":[{"id":"W1548455965","doi":"10.1177/1035719x0700700105","title":"Contribution analysis","year":2007,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":37,"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 for Advancing Health Outcomes","funders":"Australian Agency for International Development","keywords":"Outcome (game theory); Political science; Risk analysis (engineering); Process management; Computer science; Engineering; Business; Economics","score_opus":0.18382612084877567,"score_gpt":0.5447163949946406,"score_spread":0.36089027414586494,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1548455965","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":[0.88567793,0.000083748455,0.10785539,0.0017899828,0.0006822624,0.00014835836,0.0000022347642,0.0000060986417,0.0037539708],"genre_scores_gemma":[0.99771845,0.000007661924,0.0014460967,0.00012391144,0.0002389525,9.997735e-7,0.0000063350985,0.0000031839502,0.00045440017],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9925101,0.00042316323,0.0016129245,0.00015651231,0.005086727,0.0002106083],"domain_scores_gemma":[0.9911426,0.00068796537,0.0013025143,0.0002613361,0.0064304126,0.00017519617],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.062236667,0.000093786424,0.00031066663,0.0012509817,0.00010511519,0.00015438136,0.00035927538,0.000069858725,0.00875173],"category_scores_gemma":[0.004824936,0.000065608605,0.00031883456,0.0024359268,0.00004326834,0.0007320279,0.000016091297,0.00018478069,0.00019188195],"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":[0.00031630194,0.00024153152,0.5999541,0.0000020064483,0.0009112599,0.000017125241,0.0013559841,0.052410845,0.0037652391,0.0042664153,0.013248406,0.32351074],"study_design_scores_gemma":[0.0011486746,0.00018881417,0.95637727,0.0000067776737,0.00069211086,0.000032198936,0.00079663587,0.020324362,0.0026472274,0.0078190975,0.0098742405,0.00009257317],"about_ca_topic_score_codex":0.0000017030579,"about_ca_topic_score_gemma":0.000031143973,"teacher_disagreement_score":0.35642314,"about_ca_system_score_codex":0.00019358945,"about_ca_system_score_gemma":0.00039080004,"threshold_uncertainty_score":0.9921544},"labels":[],"label_agreement":null},{"id":"W1588254628","doi":"10.1177/1035719x0900900107","title":"Assessment, Evaluation and Improvement of University Council Performance","year":2009,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Commonwealth, Australian Politics and Federalism","field":"Social 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":"Centre for International Governance Innovation","funders":"","keywords":"Corporate governance; Government (linguistics); Public administration; Research council; Political science; Quality (philosophy); Public sector; Private sector; Higher education; Accounting; Business; Public relations; Engineering; Finance","score_opus":0.09597426486263874,"score_gpt":0.379544139714682,"score_spread":0.2835698748520432,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1588254628","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":[0.9819774,0.000045122484,0.00006764281,0.0024890727,0.00018761086,0.0002695428,0.0000031381783,0.0000037079596,0.014956757],"genre_scores_gemma":[0.99888444,0.00016919449,0.00037308148,0.000047824447,0.00012360606,3.016801e-7,0.00000145815,0.0000022116192,0.00039786045],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9973569,0.00031821817,0.00033450854,0.00008433813,0.001746419,0.00015959825],"domain_scores_gemma":[0.9971348,0.00005268296,0.00047681006,0.000082513594,0.0021302404,0.0001229472],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009259855,0.00006754586,0.00014507165,0.000085480664,0.00021630815,0.000037273407,0.00012354561,0.000059430582,0.00030472942],"category_scores_gemma":[0.00021555802,0.0000681317,0.000050351282,0.00014266523,0.0000935956,0.00039329016,0.000007031664,0.00013962334,8.9707584e-7],"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":[0.00041576623,0.00094703323,0.0654622,0.00013029455,0.00029611954,0.000013327852,0.037161592,0.0031787136,0.014072391,0.10087772,0.014646078,0.7627988],"study_design_scores_gemma":[0.0024882779,0.0010515639,0.96770275,0.000089194094,0.00030212995,0.00000905026,0.004712306,0.003960953,0.00039923907,0.008056948,0.01107281,0.00015480512],"about_ca_topic_score_codex":0.00013071546,"about_ca_topic_score_gemma":0.00008858226,"teacher_disagreement_score":0.9022405,"about_ca_system_score_codex":0.0008469634,"about_ca_system_score_gemma":0.0025080475,"threshold_uncertainty_score":0.44491687},"labels":[],"label_agreement":null},{"id":"W2231091450","doi":"10.1177/1035719x1501500105","title":"A Three-Way Approach to Evaluating Partnerships: Partnership Survey, Integration Measure and Social Network Analysis","year":2015,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Community Health and Development","field":"Health Professions","cited_by":3,"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":"General partnership; Social network analysis; Government (linguistics); Social Partnership; Process management; Knowledge management; Public relations; Computer science; Business; Management science; Political science; Engineering; Public administration; World Wide Web","score_opus":0.5868224709143645,"score_gpt":0.5260932263135538,"score_spread":0.06072924460081064,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2231091450","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":[0.9598454,0.00032630062,0.02922423,0.0048566107,0.000605193,0.0017281286,0.000006373631,0.000033237367,0.0033744958],"genre_scores_gemma":[0.9879909,0.000004243206,0.010393586,0.00061013753,0.0007005911,0.000073881136,0.000083325234,0.000016983058,0.00012635808],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9900083,0.006110985,0.0013812012,0.0002211594,0.001738018,0.0005403579],"domain_scores_gemma":[0.9938877,0.00081204134,0.00097575976,0.00022791099,0.0035132617,0.0005832982],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.051411156,0.00020671333,0.00057830615,0.00042257388,0.0007807883,0.0000518907,0.000234933,0.00022720892,0.00012809907],"category_scores_gemma":[0.0039172727,0.00017385965,0.00011815988,0.0013871596,0.00004479808,0.00026030187,0.00007648127,0.0010500498,0.00003072342],"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":[0.001631367,0.0002245904,0.8039541,0.00012413964,0.0007846498,0.0000018189719,0.05907925,0.008535952,0.000020876643,0.0013838456,0.07313734,0.05112202],"study_design_scores_gemma":[0.0024841456,0.00039234603,0.9582269,0.00015023621,0.00083527376,0.0000073539973,0.014498148,0.018693306,0.0000024185206,0.0034054392,0.0010627682,0.00024164976],"about_ca_topic_score_codex":0.0001364788,"about_ca_topic_score_gemma":0.0041996525,"teacher_disagreement_score":0.15427276,"about_ca_system_score_codex":0.00060466334,"about_ca_system_score_gemma":0.0023831448,"threshold_uncertainty_score":0.97677183},"labels":[],"label_agreement":null},{"id":"W2376570249","doi":"10.1177/1035719x1501500204","title":"An evaluation of governance mechanisms as determinants of performance: A case study of Australian universities","year":2015,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Corporate Finance and Governance","field":"Business, Management and Accounting","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":"Institute on Governance","funders":"","keywords":"Corporate governance; Premise; Accounting; Context (archaeology); Business; Project governance; Good governance; Public relations; Political science; Finance; Geography","score_opus":0.09357504224514256,"score_gpt":0.33775894519110694,"score_spread":0.24418390294596437,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2376570249","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":[0.9989299,0.000025294252,0.00003867614,0.000044190016,0.00030824606,0.00037820305,0.000005993605,0.0000037828236,0.00026572385],"genre_scores_gemma":[0.9996056,0.0000059584327,0.00013401826,0.000008438996,0.00013698348,0.0000030061035,0.0000021391909,0.0000107142005,0.00009317393],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99694425,0.00009936097,0.00080621406,0.00013694254,0.0018784874,0.00013476399],"domain_scores_gemma":[0.99278337,0.00001666802,0.0030532447,0.00022949601,0.003891257,0.000025937967],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0039672125,0.0001423369,0.00036111253,0.00020166946,0.00004628606,0.000024443983,0.00025292404,0.00006073909,0.0001928511],"category_scores_gemma":[0.00013460896,0.00013365585,0.00007301872,0.00045048102,0.000054494358,0.002247071,0.00002777184,0.000111418725,0.0000051926295],"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":[0.0019844605,0.0036193528,0.7498221,0.00073252065,0.00033757387,0.00052547175,0.010075123,0.14572935,0.006292274,0.0017505769,0.0019713002,0.07715988],"study_design_scores_gemma":[0.014828586,0.0037434956,0.8202638,0.00077649887,0.0023968576,0.0006235689,0.075817294,0.06739541,0.010874907,0.0025675301,0.0002639715,0.0004481031],"about_ca_topic_score_codex":0.00068780413,"about_ca_topic_score_gemma":0.00022131049,"teacher_disagreement_score":0.078333944,"about_ca_system_score_codex":0.00014252636,"about_ca_system_score_gemma":0.00061149755,"threshold_uncertainty_score":0.54503274},"labels":[],"label_agreement":null},{"id":"W2738974322","doi":"10.1177/1035719x0500500203","title":"Measuring Success: Factors Impacting on the Implementation and Use of Performance Measurement within Victoria's Human Services Agencies","year":2005,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Public Policy and Administration Research","field":"Social Sciences","cited_by":16,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for International Governance Innovation","funders":"","keywords":"Accountability; Performance measurement; Government (linguistics); Human services; Human resources; Politics; Service (business); Business; Performance management; Conceptual framework; Knowledge management; Public relations; Process management; Political science; Computer science; Marketing; Sociology","score_opus":0.31872653191797873,"score_gpt":0.4521182143741484,"score_spread":0.13339168245616967,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2738974322","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":[0.9967108,0.000018033013,0.000009430682,0.0022951327,0.00009681995,0.00020704503,0.0000028075647,0.0000048757056,0.00065508817],"genre_scores_gemma":[0.9995568,0.000024035391,0.000070387534,0.00003346445,0.00026869946,0.0000031019104,0.0000013576886,0.0000040886475,0.000038093087],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99674857,0.00050170056,0.00047673463,0.00007820688,0.002017078,0.00017770694],"domain_scores_gemma":[0.99815565,0.0001512235,0.0005985309,0.000077860765,0.0009111719,0.00010554134],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.00802687,0.000079861195,0.00011061454,0.00014166582,0.0005473289,0.00023195769,0.00017773386,0.00003825673,0.0003262256],"category_scores_gemma":[0.00048061408,0.00005501954,0.00004311722,0.0001943057,0.00009730315,0.0012280159,0.000012649301,0.0001690336,0.000001065748],"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":[0.0001913949,0.00029505297,0.54063547,0.00017139978,0.00048045637,0.0000017748281,0.33109638,0.0063972007,0.06299508,0.02571775,0.0010477076,0.03097035],"study_design_scores_gemma":[0.001099364,0.00057485024,0.81415063,0.00029555807,0.00011498333,0.000005601309,0.100092374,0.0010468066,0.076899625,0.0004086875,0.005096776,0.00021475441],"about_ca_topic_score_codex":0.00024446662,"about_ca_topic_score_gemma":0.0010324878,"teacher_disagreement_score":0.27351516,"about_ca_system_score_codex":0.00028950206,"about_ca_system_score_gemma":0.00061818596,"threshold_uncertainty_score":0.42096666},"labels":[],"label_agreement":null},{"id":"W2740844297","doi":"10.1177/1035719x0200200212","title":"Triage: A new group technique gaining recognition in evaluation","year":2002,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Delphi Technique in Research","field":"Social Sciences","cited_by":60,"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; Quebec Automobile Insurance Corporation","funders":"","keywords":"Triage; Context (archaeology); Computer science; Delphi method; Focus group; Process (computing); Delphi; Flexibility (engineering); Process management; Data science; Knowledge management; Artificial intelligence; Engineering; Medicine; Medical emergency","score_opus":0.4213983606392435,"score_gpt":0.5097812029103096,"score_spread":0.08838284227106608,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2740844297","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":[0.8527794,0.0012277245,0.01844119,0.025654962,0.0015686965,0.008389986,0.0000056856775,0.00018895934,0.091743395],"genre_scores_gemma":[0.9897678,0.00018063515,0.009056043,0.000038034847,0.0005269237,0.00008841675,0.000004411554,0.000014231709,0.00032351736],"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.9931431,0.002467484,0.00082972815,0.00018337679,0.0030437151,0.00033259418],"domain_scores_gemma":[0.9975196,0.0003249399,0.0005336656,0.00014896192,0.0012835667,0.00018924373],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.031984944,0.00011248926,0.00021080406,0.00076778565,0.00015239298,0.00012215367,0.00032499226,0.0002035038,0.007205388],"category_scores_gemma":[0.0068565924,0.00011680668,0.000106013606,0.0010546331,0.000097683456,0.00091422873,0.000017776436,0.0005604159,0.000064197695],"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":[0.00007789742,0.00021154097,0.00258538,0.0000096247695,0.000024971261,0.000023558438,0.008310052,0.00031920418,0.0037561143,0.000316928,0.024715193,0.95964956],"study_design_scores_gemma":[0.0356721,0.0053893467,0.13599104,0.005037016,0.0013133131,0.00076438294,0.04859561,0.08461918,0.021895068,0.5971449,0.06062715,0.0029509168],"about_ca_topic_score_codex":0.0001726842,"about_ca_topic_score_gemma":0.00022338062,"teacher_disagreement_score":0.9566986,"about_ca_system_score_codex":0.00096033595,"about_ca_system_score_gemma":0.0006626711,"threshold_uncertainty_score":0.9967752},"labels":[],"label_agreement":null},{"id":"W2789492532","doi":"10.1177/1035719x0700700103","title":"The fate of recommendations","year":2007,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"General Dynamics (Canada)","funders":"","keywords":"Audit; Performance audit; Identification (biology); Business; Government (linguistics); Audit plan; Auditor independence; Parliament; Accounting; Chief audit executive; Joint audit; Public relations; Internal audit; Political science; Law; Politics","score_opus":0.26287728300665364,"score_gpt":0.5593003095300301,"score_spread":0.29642302652337643,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2789492532","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":[0.9394496,0.00023775964,0.019151086,0.02098752,0.0025746278,0.00033379617,0.0000040604355,0.0000058891715,0.017255671],"genre_scores_gemma":[0.9964011,0.00005805738,0.0021770445,0.00007149699,0.00014908184,0.000001332166,0.0000013480427,0.0000036946071,0.0011368867],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99426514,0.0004458823,0.0017066842,0.00009588423,0.003323097,0.00016328832],"domain_scores_gemma":[0.99276984,0.001757742,0.0015379757,0.00026949152,0.0035730905,0.0000918402],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.05582915,0.00007056839,0.0001619728,0.00028151018,0.00019578522,0.00010414078,0.00049185735,0.00003939964,0.0027300534],"category_scores_gemma":[0.003732657,0.000039411076,0.00013422444,0.00063194113,0.0000826511,0.00048533877,0.000022245495,0.00017605584,0.00007279115],"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":[0.00012456987,0.0000885012,0.02319669,0.0000018653271,0.0000601689,0.000001170297,0.0014053069,0.0029749584,0.0019508477,0.0040457835,0.04344601,0.9227041],"study_design_scores_gemma":[0.0024683517,0.00062306266,0.6841198,0.000067442495,0.00013557273,0.000094280964,0.010261884,0.010483724,0.012785486,0.052629706,0.22615927,0.00017138009],"about_ca_topic_score_codex":0.0000014283061,"about_ca_topic_score_gemma":0.000028517457,"teacher_disagreement_score":0.92253274,"about_ca_system_score_codex":0.00007786153,"about_ca_system_score_gemma":0.0004706749,"threshold_uncertainty_score":0.9981816},"labels":[],"label_agreement":null},{"id":"W2790752207","doi":"10.1177/1035719x1701700204","title":"Interactive Logic Models: Using Design and Technology to Explore the Effects of Dynamic Situations on Program Logic","year":2017,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Ontario College of Art and Design; Centre for Advancing Health Outcomes","funders":"","keywords":"Logic model; Computer science; Interactivity; Software engineering; Software; Process (computing); Logic programming; Human–computer interaction; Programming language; World Wide Web","score_opus":0.4434407751028504,"score_gpt":0.5792202090185361,"score_spread":0.1357794339156857,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2790752207","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":[0.89579386,0.000072534174,0.09252,0.009646002,0.00048102488,0.001234435,9.448086e-7,0.000008979317,0.00024222981],"genre_scores_gemma":[0.9770121,0.000030579966,0.022735666,0.000083739775,0.0000342233,0.00003759086,4.9952786e-7,0.000006035834,0.000059595295],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9967277,0.0005618194,0.0007045566,0.00019957397,0.0016522991,0.000154064],"domain_scores_gemma":[0.99499124,0.0009985052,0.0013728607,0.00048341235,0.002074495,0.00007946606],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0070545785,0.00012700603,0.00026594536,0.00054434425,0.00038647107,0.0002558834,0.00066119403,0.00007832239,0.000104781175],"category_scores_gemma":[0.0062539964,0.00007329917,0.00007174731,0.00035793605,0.00017204652,0.0007863135,0.00007797802,0.00025318368,0.0000136247645],"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":[0.00025916207,0.00041356354,0.0016905107,0.000016965849,0.00013310352,0.000008249871,0.005204042,0.25239727,0.011040946,0.002811854,0.00076350797,0.72526085],"study_design_scores_gemma":[0.0010619723,0.0024296073,0.022475405,0.0002225181,0.000141694,0.000056414203,0.0030589742,0.765372,0.0027335281,0.20228429,0.000047931433,0.00011568746],"about_ca_topic_score_codex":0.000001551459,"about_ca_topic_score_gemma":0.000004237009,"teacher_disagreement_score":0.72514516,"about_ca_system_score_codex":0.00012891418,"about_ca_system_score_gemma":0.00036136448,"threshold_uncertainty_score":0.74870694},"labels":[],"label_agreement":null},{"id":"W2791138626","doi":"10.1177/1035719x0500500104","title":"The Development of a Logic Model for the Protection against Family Violence Act: An Incremental Approach","year":2005,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","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":true,"ca_institutions":"Government of Northwest Territories","funders":"","keywords":"Legislation; Jurisdiction; Economic Justice; Process (computing); Domestic violence; Political science; Law; Computer science; Poison control; Medicine; Human factors and ergonomics; Environmental health; Programming language","score_opus":0.37940263289098225,"score_gpt":0.4914204480204588,"score_spread":0.11201781512947656,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2791138626","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":[0.73522466,0.00009898156,0.261471,0.0014758765,0.00020114114,0.0009832239,0.0000018886543,0.0000046118635,0.0005385911],"genre_scores_gemma":[0.9649699,0.000038196547,0.03448948,0.00011985717,0.00014923543,0.000085947824,0.0000029969513,0.000005991021,0.00013842454],"study_design_codex":"design_other","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9948038,0.0003661676,0.0013841132,0.00017182938,0.003092328,0.00018176751],"domain_scores_gemma":[0.9959968,0.00034883482,0.001241942,0.0002869418,0.0020511195,0.00007437767],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.027956976,0.00011887703,0.00017480475,0.00014712285,0.00053727545,0.00017136043,0.00069782994,0.000054198026,0.000077148215],"category_scores_gemma":[0.00083465345,0.00005769844,0.0001188613,0.00030576627,0.0000860327,0.0007649029,0.00003276292,0.0001941498,0.000013154309],"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":[0.000115065624,0.000100750956,0.00030047836,0.0000028003844,0.000028255743,2.0644038e-8,0.0021587186,0.30002734,0.0041022743,0.00019126249,0.0002686378,0.6927044],"study_design_scores_gemma":[0.0008197395,0.00017169586,0.023117507,0.00002628605,0.000038576396,0.0000060217358,0.0044096154,0.96493363,0.0026385842,0.002652424,0.0011120109,0.000073909774],"about_ca_topic_score_codex":7.5628856e-7,"about_ca_topic_score_gemma":0.00002784377,"teacher_disagreement_score":0.69263047,"about_ca_system_score_codex":0.00020520821,"about_ca_system_score_gemma":0.0012530881,"threshold_uncertainty_score":0.9689385},"labels":[],"label_agreement":null},{"id":"W2792013179","doi":"10.1177/1035719x1601600103","title":"Performance measurement as precursor to organizational evaluation capacity building","year":2016,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Public Policy and Administration Research","field":"Social Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"National Circus School","funders":"","keywords":"Transparency (behavior); Performance measurement; Business; Capacity building; Process management; Organizational performance; Knowledge management; Bridge (graph theory); Computer science; Risk analysis (engineering); Management science; Marketing; Engineering; Economics; Computer security; Medicine","score_opus":0.24231114241699608,"score_gpt":0.43398980465486364,"score_spread":0.19167866223786756,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2792013179","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":[0.9620994,0.000015228939,0.0011227397,0.024718622,0.00041136166,0.00051453535,0.0000039078027,0.000012689203,0.011101564],"genre_scores_gemma":[0.9976778,0.000019421092,0.00089423783,0.00011709879,0.00066638505,0.000017977485,9.3961967e-7,0.000007279706,0.00059885887],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.99325377,0.0008689725,0.0004691888,0.00014551095,0.0049787266,0.00028385027],"domain_scores_gemma":[0.99338955,0.000118943106,0.00030805205,0.00011129999,0.0057330867,0.00033909877],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.018850414,0.00009077345,0.00012292418,0.00026253745,0.00035750482,0.00011670555,0.00030124732,0.00007500051,0.0046898695],"category_scores_gemma":[0.011217706,0.00006829209,0.000054534466,0.00053567626,0.000104629915,0.00086428743,0.000016692675,0.00011662932,0.00014851306],"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":[0.00049569726,0.00081653445,0.06335225,0.000051996944,0.00041915526,0.0000053653116,0.025648832,0.003543253,0.17764825,0.08607168,0.0555424,0.5864046],"study_design_scores_gemma":[0.009680821,0.0027221327,0.63638824,0.001338407,0.00054965215,0.00014580191,0.0038823842,0.0042223325,0.113773316,0.048529573,0.17752874,0.001238628],"about_ca_topic_score_codex":0.0000236717,"about_ca_topic_score_gemma":0.00005639883,"teacher_disagreement_score":0.585166,"about_ca_system_score_codex":0.0010960919,"about_ca_system_score_gemma":0.0045493767,"threshold_uncertainty_score":0.9971112},"labels":[],"label_agreement":null},{"id":"W2793638346","doi":"10.1177/1035719x0300300205","title":"Reflections on a decade in the life of the Australasian Evaluation Society: 1990-1999","year":2003,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":1,"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":"Rigour; Professional association; Consolidation (business); Political science; Sociology; Psychology; Public relations; Epistemology","score_opus":0.3166674843876108,"score_gpt":0.5439863815010553,"score_spread":0.2273188971134445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2793638346","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":[0.96114254,0.00016862033,0.00041516774,0.01621635,0.001386323,0.001120911,0.000004963602,0.00000520164,0.019539941],"genre_scores_gemma":[0.9980512,0.00002997059,0.0006854401,0.00077824114,0.00015132796,0.00003290743,0.0000018271037,0.00000907029,0.00026001624],"study_design_codex":"simulation_or_modeling","study_design_gemma":"observational","domain_scores_codex":[0.985623,0.0040258043,0.0018071062,0.00026136165,0.008015937,0.0002667616],"domain_scores_gemma":[0.9944098,0.001050083,0.001600194,0.0006389757,0.0021943734,0.000106565436],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.052213907,0.00018044015,0.00030929624,0.00036807155,0.00027927515,0.00016875261,0.00089052675,0.00013069811,0.0027191648],"category_scores_gemma":[0.008996985,0.0000951842,0.00041749244,0.0021131758,0.00014121438,0.0005697297,0.000020449368,0.0006323112,0.00005246361],"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":[0.0003688242,0.0022087446,0.13704538,0.000037012614,0.0004818566,0.0000042142483,0.053645268,0.48059088,0.007215796,0.032879252,0.17706536,0.10845741],"study_design_scores_gemma":[0.0045153066,0.00082638394,0.8715647,0.00018190563,0.00049075333,0.00013392301,0.026139261,0.022303298,0.0031654933,0.041474797,0.028932681,0.00027153682],"about_ca_topic_score_codex":0.000008911642,"about_ca_topic_score_gemma":0.00008877196,"teacher_disagreement_score":0.7345193,"about_ca_system_score_codex":0.00028063444,"about_ca_system_score_gemma":0.0024138507,"threshold_uncertainty_score":0.99935067},"labels":[],"label_agreement":null},{"id":"W3015092896","doi":"10.1177/1035719x20905056","title":"Thinking with theory as a policy evaluation tool: The case of boarding schools for remote First Nations students","year":2020,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Indigenous and Place-Based Education","field":"Social Sciences","cited_by":20,"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":"Project commissioning; Investment (military); Boarding school; Publishing; Quality (philosophy); Public relations; Policy analysis; Public administration; Economic growth; Sociology; Political science; Economics","score_opus":0.06148909512569058,"score_gpt":0.42209298209277973,"score_spread":0.36060388696708917,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3015092896","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":[0.9726868,0.00009212215,0.0032488452,0.020621011,0.0002572711,0.0015006852,0.000007616394,0.000009620976,0.0015760203],"genre_scores_gemma":[0.9971394,0.000032429092,0.0015648981,0.00022167798,0.00092590955,0.000015403466,0.000006663598,0.000011143081,0.00008247047],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.9966158,0.00095571234,0.00047919902,0.00012325704,0.0016258052,0.00020021558],"domain_scores_gemma":[0.9959178,0.0007854778,0.0007802992,0.0001139824,0.0022981032,0.0001043249],"candidate_categories":["sts"],"consensus_categories":[],"category_scores_codex":[0.016059747,0.0000985836,0.0001566834,0.00018773896,0.0013052329,0.00020873522,0.00032482014,0.000070890994,0.00026907935],"category_scores_gemma":[0.0069225985,0.000071293354,0.00010723796,0.00062728726,0.000078923724,0.00046288862,0.000011717495,0.00020697087,0.000007043452],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0009706323,0.00027067904,0.009890073,0.00012481656,0.00065184454,0.000027061993,0.77611274,0.051015347,0.00024266257,0.081837796,0.0027396667,0.0761167],"study_design_scores_gemma":[0.014731878,0.0046195174,0.029641725,0.0019071365,0.005673839,0.0010378344,0.5854094,0.0309385,0.002743344,0.30115184,0.020931037,0.0012139558],"about_ca_topic_score_codex":0.0005355421,"about_ca_topic_score_gemma":0.0010977517,"teacher_disagreement_score":0.21931405,"about_ca_system_score_codex":0.0006278813,"about_ca_system_score_gemma":0.0062515168,"threshold_uncertainty_score":0.99999493},"labels":[],"label_agreement":null},{"id":"W3151837118","doi":"10.1177/1035719x20982087","title":"Book review: Qualitative inquiry: Thematic, narrative and arts-based perspectives (2nd edition)","year":2020,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Participatory Visual Research Methods","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":false,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Centre for Advancing Health Outcomes","funders":"","keywords":"Thematic map; Narrative; The arts; Narrative inquiry; Epistemology; Sociology; Visual arts; Art; Literature; Philosophy; Geography; Cartography","score_opus":0.6949894939658819,"score_gpt":0.6938884711478448,"score_spread":0.001101022818037145,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3151837118","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":[0.24979119,0.19662201,0.036963444,0.5053477,0.000958514,0.004184057,0.000018842564,0.00009374291,0.00602049],"genre_scores_gemma":[0.9794704,0.0045093284,0.006027923,0.007964731,0.0015837624,0.00006315616,0.0000028529844,0.000022802411,0.00035505262],"study_design_codex":"qualitative","study_design_gemma":"qualitative","domain_scores_codex":[0.991137,0.0062383623,0.0005296428,0.00017214408,0.0016915469,0.00023125383],"domain_scores_gemma":[0.99608755,0.001133868,0.00053242553,0.000081158396,0.0017531462,0.00041184554],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.013154943,0.0001155373,0.00033021538,0.00009407393,0.0003008642,0.000072369854,0.00017406899,0.00005613718,0.004574106],"category_scores_gemma":[0.02229311,0.00009790864,0.000101171136,0.00036837487,0.0006651182,0.0008075439,0.000015511128,0.00031563162,0.000018175815],"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":[0.00007856995,0.0000766528,0.00006834738,0.000119199496,0.00007763011,0.000010242614,0.770768,0.000049395032,0.00026821363,0.0016614974,0.22480023,0.002022051],"study_design_scores_gemma":[0.0017385415,0.0013042347,0.0021454091,0.001368161,0.0003120594,0.000008852086,0.89265555,0.0012781654,0.0011831617,0.0042612054,0.09343783,0.00030680595],"about_ca_topic_score_codex":0.0000038178864,"about_ca_topic_score_gemma":0.0000031356371,"teacher_disagreement_score":0.72967917,"about_ca_system_score_codex":0.00020482196,"about_ca_system_score_gemma":0.0011508625,"threshold_uncertainty_score":0.99633586},"labels":[],"label_agreement":null},{"id":"W3154034046","doi":"10.1177/1035719x211008263","title":"Thinking with complexity in evaluation: A case study review","year":2021,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","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":"Impact","funders":"","keywords":"Context (archaeology); Management science; Social complexity; Computer science; Complexity management; Exploratory research; Knowledge management; Sociology; Social science; Engineering; Management","score_opus":0.5494958181463374,"score_gpt":0.5867373010000473,"score_spread":0.03724148285370987,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3154034046","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":[0.986816,0.0023626813,0.00044425696,0.006238922,0.0003958947,0.0010740968,0.0000016689064,0.0000056724425,0.0026607893],"genre_scores_gemma":[0.99647003,0.00012872653,0.0025820786,0.0005141794,0.000114066104,0.000029474224,0.0000045269217,0.000008114734,0.00014878133],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9838009,0.0038527367,0.0019411156,0.0003379382,0.009859649,0.00020766737],"domain_scores_gemma":[0.98900443,0.0005260991,0.0012371781,0.00052397145,0.008572862,0.00013545254],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.07097194,0.0001652614,0.00052028015,0.000373412,0.00016852785,0.00027866373,0.00041409294,0.000044165485,0.013533339],"category_scores_gemma":[0.004135991,0.0001121453,0.00012035682,0.0018157378,0.00005778907,0.0010034977,0.000053949545,0.0003892168,0.000058722962],"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":[0.00019653211,0.0024703469,0.56338626,0.0001319776,0.00030185122,0.0061930716,0.01717302,0.039237913,0.00008147398,0.00054752885,0.014206309,0.35607368],"study_design_scores_gemma":[0.016470566,0.0023773268,0.7433393,0.0030463333,0.0015919012,0.039998394,0.08924558,0.07460197,0.00019332535,0.024561448,0.0038879823,0.0006858894],"about_ca_topic_score_codex":0.000014059578,"about_ca_topic_score_gemma":0.00092943787,"teacher_disagreement_score":0.3553878,"about_ca_system_score_codex":0.00031594088,"about_ca_system_score_gemma":0.0030742483,"threshold_uncertainty_score":0.9873684},"labels":[],"label_agreement":null},{"id":"W4221079982","doi":"10.1177/1035719x221080575","title":"Evaluation in the field of early childhood development: A scoping review","year":2022,"lang":"en","type":"review","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":9,"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 Alberta","funders":"Social Sciences and Humanities Research Council of Canada","keywords":"Bridging (networking); Field (mathematics); Evaluation methods; Inclusion (mineral); Program evaluation; Early childhood education; Impact evaluation; Capacity building; Early childhood; Psychology; Political science; Management science; Computer science; Engineering; Medicine; Pedagogy; Social psychology; Developmental psychology","score_opus":0.43272754631129584,"score_gpt":0.5948855563214568,"score_spread":0.16215801001016095,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4221079982","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":[0.00034536462,0.9935125,0.0000532131,0.0005351694,0.000594214,0.003216775,0.0000032836244,0.0000026062457,0.0017368468],"genre_scores_gemma":[0.0011153144,0.9977635,0.00036883727,0.00027966936,0.00013590099,0.0002551552,0.000023460396,0.000015224918,0.000042909418],"study_design_codex":"design_other","study_design_gemma":"not_applicable","domain_scores_codex":[0.9690884,0.008993531,0.005662453,0.00038367877,0.0156013435,0.00027062325],"domain_scores_gemma":[0.98736346,0.0041399472,0.0056365337,0.0006414216,0.0021439781,0.00007463207],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.14245601,0.00030046926,0.0016162171,0.0010558168,0.00015471768,0.00014117334,0.0016276438,0.00013681651,0.01974411],"category_scores_gemma":[0.016715068,0.00017935617,0.00059419544,0.0023198358,0.000034417673,0.0006032061,0.000089330875,0.00087854173,0.00006764579],"study_design_candidate":"design_other","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":true,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000035718422,0.000089232235,0.0000419497,0.0023546817,0.000062713945,0.0000017458914,0.0012855429,0.00011988315,3.677255e-8,0.000013662501,0.0015460566,0.9944809],"study_design_scores_gemma":[0.0014371711,0.000761228,0.0043515293,0.30046332,0.0034060976,0.00023853108,0.0008437159,0.00023102539,0.0000032914986,0.0011626822,0.68663865,0.00046277823],"about_ca_topic_score_codex":0.00000228786,"about_ca_topic_score_gemma":0.000016269947,"teacher_disagreement_score":0.99401814,"about_ca_system_score_codex":0.00041671158,"about_ca_system_score_gemma":0.013194819,"threshold_uncertainty_score":0.99239945},"labels":[],"label_agreement":null},{"id":"W4291209168","doi":"10.1177/1035719x221119841","title":"Developmental evaluation during the COVID-19 pandemic: Practice-based learnings from projects in British Columbia, Canada","year":2022,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","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":true,"ca_institutions":"Centre for Advancing Health Outcomes; University of British Columbia","funders":"","keywords":"Context (archaeology); Leverage (statistics); Knowledge management; Embeddedness; Coronavirus disease 2019 (COVID-19); Pandemic; Ambidexterity; Stakeholder engagement; Public relations; Process management; Political science; Business; Sociology; Computer science","score_opus":0.21498372054834972,"score_gpt":0.4556638389750601,"score_spread":0.24068011842671036,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4291209168","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":[0.99262136,0.00013966222,0.00010522997,0.0052861045,0.0006094721,0.00085116207,0.000027015885,0.0000075688695,0.00035241348],"genre_scores_gemma":[0.9966317,0.000009663759,0.0006851096,0.0018093532,0.00010913449,0.00012173714,0.000029178647,0.000011077609,0.0005930625],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9825811,0.004611411,0.0015414143,0.00034804584,0.010640408,0.00027764027],"domain_scores_gemma":[0.9939653,0.002084328,0.001756648,0.00019647877,0.0018094039,0.00018778397],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.047283057,0.0001145097,0.00026500373,0.0002239635,0.00084804447,0.0006601074,0.0006871485,0.00004665624,0.025781836],"category_scores_gemma":[0.015271136,0.00012389424,0.00008621834,0.0011050525,0.00005502074,0.00082896935,0.000090001035,0.00071140676,0.000011364208],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_system_candidate":true,"about_ca_system_consensus":true,"study_design_scores_codex":[0.000098634235,0.000095578325,0.755562,0.000003474899,0.000036303485,0.00003409886,0.0023920743,0.19053322,0.0002508311,9.151627e-7,0.015503768,0.03548911],"study_design_scores_gemma":[0.004445442,0.00014405901,0.86159396,0.00002348682,0.000093604045,0.000578614,0.044354916,0.0410526,0.000023451206,0.00053001876,0.04697265,0.0001872208],"about_ca_topic_score_codex":0.34233454,"about_ca_topic_score_gemma":0.9006011,"teacher_disagreement_score":0.5582665,"about_ca_system_score_codex":0.005977111,"about_ca_system_score_gemma":0.032188762,"threshold_uncertainty_score":0.9978388},"labels":[],"label_agreement":null},{"id":"W4309592929","doi":"10.1177/1035719x221139858","title":"A Culturally Adaptive Approach to First Nations evaluation consulting","year":2022,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"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":"Reflexivity; Context (archaeology); Corporate governance; Public relations; Cultural diversity; Work (physics); Sociology; Political science; Business; Social science; Engineering; Geography; Law","score_opus":0.28401913605717927,"score_gpt":0.4882824620841202,"score_spread":0.20426332602694092,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4309592929","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":[0.8740892,0.00028330748,0.019504068,0.015312221,0.0028983522,0.0034630485,0.000056057885,0.000038421746,0.08435536],"genre_scores_gemma":[0.9879614,0.0000053668027,0.010433683,0.00026935624,0.00021420179,0.00019551197,0.000022236376,0.000011312178,0.00088691106],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.98499095,0.0019330255,0.001637395,0.0003609538,0.010811838,0.0002658497],"domain_scores_gemma":[0.989737,0.0011162744,0.0014360001,0.0003438887,0.007142915,0.00022395159],"candidate_categories":["metaresearch","insufficient_payload"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.04182841,0.00017413107,0.00031662942,0.0011067004,0.0011816105,0.0002719942,0.00078561914,0.000045740613,0.009048968],"category_scores_gemma":[0.010564907,0.00013825123,0.00019729121,0.0023825592,0.000045111166,0.0009522168,0.0001293889,0.00041268006,0.00014374165],"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":[0.0001444164,0.00030029408,0.003950591,0.0000036742804,0.00008029316,0.0000013609207,0.009229834,0.913559,0.00029420046,0.002475537,0.026881816,0.04307898],"study_design_scores_gemma":[0.004577788,0.001213417,0.11117091,0.000051340125,0.00036394515,0.00025820217,0.04228308,0.7860399,0.0002399749,0.0066941,0.046673615,0.00043371206],"about_ca_topic_score_codex":0.0000064161013,"about_ca_topic_score_gemma":0.00004165652,"teacher_disagreement_score":0.12751909,"about_ca_system_score_codex":0.0009207907,"about_ca_system_score_gemma":0.0018138336,"threshold_uncertainty_score":0.99776953},"labels":[],"label_agreement":null},{"id":"W4378373311","doi":"10.1177/1035719x231179984","title":"Advancing an ethical imperative for collaborative approaches to evaluation with low incidence and underserved communities: Insights from a DeafBlind Support Services pilot program evaluation","year":2023,"lang":"en","type":"article","venue":"Evaluation Journal of Australasia","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 Alberta","funders":"","keywords":"Transformative learning; Vocational education; Recreation; Service (business); Participatory evaluation; Psychology; Public relations; Medical education; Sociology; Political science; Medicine; Pedagogy; Business","score_opus":0.4662869316973545,"score_gpt":0.508740093857817,"score_spread":0.04245316216046252,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4378373311","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":[0.9858291,0.00006709424,0.004033766,0.0040510925,0.00029994166,0.0054620043,0.000037594134,0.00003700111,0.00018237831],"genre_scores_gemma":[0.9834609,0.000023253551,0.014534305,0.00038571656,0.00020095473,0.0010090844,0.00033791084,0.000025424228,0.000022425333],"study_design_codex":"simulation_or_modeling","study_design_gemma":"simulation_or_modeling","domain_scores_codex":[0.9862161,0.0037934538,0.0015304354,0.0005060865,0.0075936057,0.00036030618],"domain_scores_gemma":[0.9842301,0.0020311181,0.0012991222,0.00049358694,0.011572286,0.0003737862],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03836672,0.00031644193,0.0005463703,0.00087214226,0.0005649667,0.0008134449,0.00057905045,0.00019490166,0.00065258186],"category_scores_gemma":[0.00203345,0.00022997885,0.00007790538,0.002000052,0.00015466067,0.0026576768,0.00007832156,0.0005164151,0.000034974946],"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":[0.0030140039,0.00078452844,0.022234382,0.00007936016,0.00042758437,0.00000470391,0.15214032,0.501812,0.0025272951,0.0007726454,0.001210678,0.31499252],"study_design_scores_gemma":[0.0045193722,0.005859018,0.099463515,0.0002059728,0.00038617157,0.000012817579,0.11402939,0.76057744,0.0008247614,0.013595279,0.00024993805,0.00027629832],"about_ca_topic_score_codex":0.000042952568,"about_ca_topic_score_gemma":0.0038496882,"teacher_disagreement_score":0.31471622,"about_ca_system_score_codex":0.00046360848,"about_ca_system_score_gemma":0.0034088006,"threshold_uncertainty_score":0.9902038},"labels":[{"model":"gpt","categories":[],"domain":null,"study_design":"qualitative","genre":"methods","about_ca_system":false,"about_ca_topic":false,"confidence":"high"},{"model":"grok","categories":[],"domain":null,"study_design":"qualitative","genre":"methods","about_ca_system":false,"about_ca_topic":true,"confidence":"high"},{"model":"opus","categories":[],"domain":null,"study_design":"theoretical_or_conceptual","genre":"commentary","about_ca_system":false,"about_ca_topic":false,"confidence":"medium"}],"label_agreement":"split"},{"id":"W4386880443","doi":"10.1177/1035719x231200050","title":"A scoping review of commissioning practices used in the evaluation of Indigenous health and wellbeing programs: Protocol article","year":2023,"lang":"en","type":"review","venue":"Evaluation Journal of Australasia","topic":"Indigenous Health, Education, and Rights","field":"Social Sciences","cited_by":5,"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":"National Health and Medical Research Council","keywords":"Project commissioning; Indigenous; Publishing; Best practice; Public relations; Commission; Political science; Law","score_opus":0.3515823254558067,"score_gpt":0.5814701198182733,"score_spread":0.22988779436246665,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4386880443","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":[0.0018497754,0.88271457,0.0000028770266,0.00036046118,0.00017046212,0.11447928,0.0000013607585,0.000006471457,0.00041472955],"genre_scores_gemma":[0.00019494654,0.99703354,0.00026193657,0.00001249053,0.00022343807,0.0022211156,0.000010152006,0.000016423519,0.000025934713],"study_design_codex":"design_other","study_design_gemma":"systematic_review","domain_scores_codex":[0.9836188,0.010559839,0.0025120848,0.0001876032,0.002784116,0.00033754425],"domain_scores_gemma":[0.9881775,0.00082522654,0.009074114,0.00020488996,0.0015789393,0.00013935794],"candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.07057003,0.00017280814,0.0010659863,0.00035682414,0.0010073484,0.00006869298,0.0003501672,0.00015477538,0.00005871813],"category_scores_gemma":[0.0011171293,0.00010953849,0.00017058413,0.0011242611,0.00014198113,0.00039889812,0.000003056817,0.00046340082,0.0000023215046],"study_design_candidate":"systematic_review","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":[0.0000100443685,0.00030872642,0.00025515477,0.21209851,0.00007360882,0.0000012663816,0.29282728,0.000010910923,8.2822645e-8,0.0002974251,0.000035258454,0.49408174],"study_design_scores_gemma":[0.0006667453,0.00048750092,0.00017977598,0.8423845,0.00091606437,0.000039156268,0.013330777,0.000025934301,0.0000012036015,0.000393018,0.14141172,0.00016356031],"about_ca_topic_score_codex":0.0022853476,"about_ca_topic_score_gemma":0.020196576,"teacher_disagreement_score":0.63028604,"about_ca_system_score_codex":0.00052114483,"about_ca_system_score_gemma":0.088314034,"threshold_uncertainty_score":0.9976823},"labels":[],"label_agreement":null}]}