{"meta":{"page":1,"per_page":50,"max_per_page":100,"total":73,"total_is_capped":false,"direct_labels_cover":1,"predictions_cover":73,"direct_label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline (scores rank; they never assert a category)","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12","author_layer_release":"2026-06-26"},"query_hash":"3e0a82fc70b4","filters":{"venue":"New Directions for Evaluation"}},"results":[{"id":"W2101260272","doi":"10.1002/ev.313","title":"Policy implementation: Implications for evaluation","year":2009,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":96,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University; Douglas Mental Health University Institute","funders":"","keywords":"Accountability; Toolbox; Corporate governance; Context (archaeology); Process (computing); Government (linguistics); Sustainability; Political science; Public relations; Process management; Public administration; Sociology; Computer science; Business; Management; Economics","authors":[{"name":"Amy DeGroff","is_ca":false},{"name":"Margaret Cargo","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3635462965884845,"gpt":0.6388161775626578,"spread":0.2752698809741734,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.4540865,0.003371701,0.00587556,0.007350595,0.005036974,0.02814998,0.007697774,0.01594864,0.02746675],"category_scores_gemma":[0.6958904,0.001073708,0.002346772,0.01272551,0.01443326,0.02382317,0.0084104,0.00875962,0.002057092],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.03223608,"about_ca_system_score_gemma":0.07962222,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01190211,"about_ca_topic_score_gemma":0.009014539,"domain_scores_codex":[0.43743,0.5136371,0.01552967,0.004082157,0.0244753,0.004845687],"domain_scores_gemma":[0.1960635,0.7134812,0.01667934,0.01441785,0.05197144,0.007386603],"domain_codex":"methods","domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.002295794,0.001701429,0.007585606,0.01691569,0.0006862576,0.0004469728,0.003865322,0.008790869,0.000160149,0.4755763,0.08701752,0.3949581],"study_design_scores_gemma":[0.001920436,0.0017028,0.006551351,0.07645462,0.001055226,0.000226955,0.02260603,0.02247754,0.001068961,0.7217121,0.1439165,0.00030742],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.01166026,0.07371935,0.05988561,0.7118945,0.008755189,0.0105992,0.002014727,0.0005922159,0.1208791],"genre_scores_gemma":[0.6265844,0.06230487,0.1592458,0.09050732,0.004668756,0.04422284,0.001246258,0.0003387362,0.01088101],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.4540865,"threshold_uncertainty_score":0.6732086,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1543074450","doi":"10.1002/ev.20076","title":"Framing the Capacity to Do and Use Evaluation","year":2014,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":96,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université du Québec en Outaouais; University of Ottawa","funders":"","keywords":"Framing (construction); Conceptualization; Situated; Construct (python library); Capacity building; Organizational change; Sociology; Management science; Knowledge management; Public relations; Computer science; Political science; Economics","authors":[{"name":"J. Bradley Cousins","is_ca":true},{"name":"Swee C. Goh","is_ca":true},{"name":"Catherine Elliott","is_ca":true},{"name":"Isabelle Bourgeois","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3372561403528904,"gpt":0.5217974321371192,"spread":0.1845412917842288,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.04145863,0.0008836851,0.0005808765,0.004957188,0.005061521,0.01802886,0.002036938,0.0052275,0.007687629],"category_scores_gemma":[0.06174754,0.0006856648,0.0005207575,0.002663179,0.05970605,0.01736624,0.01152164,0.006538274,0.0005711409],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01662704,"about_ca_system_score_gemma":0.01525394,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.009178755,"about_ca_topic_score_gemma":0.004943858,"domain_scores_codex":[0.9557816,0.03491345,0.001254122,0.001649959,0.004195326,0.00220559],"domain_scores_gemma":[0.9009835,0.07559327,0.003817906,0.005274771,0.01106453,0.003266112],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000007568898,0.0000116773,0.0002991858,0.00004508108,0.0000028513,0.00003228566,0.00427847,0.0004400515,0.0000804563,0.9882413,0.001094232,0.005466822],"study_design_scores_gemma":[0.00001659371,0.00002561242,0.0007192267,0.0007522631,0.00001166331,0.00006353661,0.01261406,0.002646409,0.0006455964,0.9133756,0.06908432,0.00004520083],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.06639417,0.00757985,0.2284453,0.1445264,0.001252479,0.0005096663,0.0001746878,0.000211323,0.5509061],"genre_scores_gemma":[0.9661865,0.00170588,0.02214491,0.002494883,0.0002771133,0.0004177698,0.00005099316,0.00006223343,0.006659765],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9585414,"threshold_uncertainty_score":0.2192568,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2140465263","doi":"10.1002/ev.240","title":"Going through the process: An examination of the operationalization of process use in empirical research on evaluation","year":2007,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":74,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Operationalization; Process (computing); Empirical research; Empirical examination; Process management; Research methodology; Psychology; Management science; Computer science; Sociology; Epistemology; Business; Actuarial science; Engineering","authors":[{"name":"Courtney Amo","is_ca":true},{"name":"J. Bradley Cousins","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.6748434947289444,"gpt":0.6733453144103931,"spread":0.001498180318551312,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1162731,0.0007493747,0.0008917744,0.0117999,0.004081478,0.01359049,0.002098031,0.002649886,0.001891506],"category_scores_gemma":[0.3796244,0.0006970566,0.001487867,0.01176674,0.02897124,0.02055711,0.00711375,0.004382303,0.0001493087],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00840466,"about_ca_system_score_gemma":0.00929725,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006004802,"about_ca_topic_score_gemma":0.003690885,"domain_scores_codex":[0.8412625,0.1203737,0.01115009,0.003143049,0.02169799,0.002372661],"domain_scores_gemma":[0.3469426,0.5728718,0.0316855,0.01900832,0.02767095,0.001820814],"domain_codex":null,"domain_gemma":"methods","domain_candidate":"methods","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002561738,0.0004901417,0.08066736,0.001598976,0.0002029227,0.0002054672,0.2881257,0.002244,0.001877612,0.4763933,0.0007293436,0.147209],"study_design_scores_gemma":[0.0001262885,0.001225228,0.1566842,0.007287987,0.0004199436,0.0007805972,0.2978917,0.03007906,0.007073355,0.4659566,0.03211258,0.0003624251],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.6563143,0.003953152,0.2713367,0.008134892,0.000135219,0.001487806,0.0001180423,0.0002794402,0.05824037],"genre_scores_gemma":[0.9708866,0.0005134114,0.02737309,0.0002435374,0.00001620595,0.0005777811,0.00003887075,0.00005325877,0.0002973245],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.883727,"threshold_uncertainty_score":0.614918,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2101892546","doi":"10.1002/ev.316","title":"Knowledge theories can inform evaluation practice: What can a complexity lens add?","year":2009,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":60,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Competence (human resources); Fidelity; Knowledge management; Computer science; Process (computing); Context (archaeology); Complex adaptive system; Paradigm shift; Psychology; Epistemology; Artificial intelligence; Social psychology","authors":[{"name":"Penelope Hawe","is_ca":true},{"name":"Lyndal Bond","is_ca":false},{"name":"Helen Butler","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3755486202484554,"gpt":0.565178999495196,"spread":0.1896303792467406,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.3522426,0.003306118,0.006747243,0.03062395,0.0079228,0.04666286,0.007092794,0.01122383,0.01507694],"category_scores_gemma":[0.4993799,0.002246881,0.003868798,0.01305716,0.06751481,0.08432464,0.02345114,0.01721628,0.001830223],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02655771,"about_ca_system_score_gemma":0.02743994,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.007586419,"about_ca_topic_score_gemma":0.0102052,"domain_scores_codex":[0.6143786,0.3359211,0.01381965,0.003515593,0.02908884,0.003276369],"domain_scores_gemma":[0.1858568,0.7434679,0.01043797,0.02107557,0.03487414,0.004287641],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0001804955,0.000189028,0.003326719,0.006996929,0.0003104134,0.0001841564,0.008726928,0.002308187,0.0001153044,0.7663354,0.0201533,0.1911732],"study_design_scores_gemma":[0.000087857,0.0001250516,0.001092166,0.01506418,0.0001476913,0.0001217967,0.009821593,0.003024222,0.0002517103,0.9247552,0.04537641,0.0001322275],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.008178153,0.1077695,0.2630899,0.5545104,0.006585613,0.001374525,0.0003544737,0.0006124529,0.05752506],"genre_scores_gemma":[0.4621877,0.07241422,0.394902,0.05125061,0.008398079,0.006358791,0.0002357275,0.0004527334,0.003800199],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.3522426,"threshold_uncertainty_score":0.7988002,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2066945648","doi":"10.1002/ev.315","title":"Knowledge translation: Implications for evaluation","year":2009,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":49,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Canadian Institutes of Health Research","funders":"","keywords":"Knowledge translation; Computer science; Knowledge management; Field (mathematics); Translation (biology); Sustainability","authors":[{"name":"Colleen Davison","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5370852562015647,"gpt":0.6053385358954619,"spread":0.06825327969389727,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.3484131,0.002971364,0.006076108,0.01081293,0.006256331,0.02650989,0.006897055,0.01511735,0.02383847],"category_scores_gemma":[0.6634713,0.001090195,0.002738133,0.01682822,0.03192297,0.03764268,0.01087729,0.009222144,0.002351412],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02912114,"about_ca_system_score_gemma":0.04479538,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01101193,"about_ca_topic_score_gemma":0.005171733,"domain_scores_codex":[0.5179054,0.4376667,0.01344768,0.005042946,0.02253345,0.003403878],"domain_scores_gemma":[0.1837628,0.7369837,0.01341164,0.01534541,0.04593169,0.004564777],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0008893514,0.000388031,0.00222179,0.009355695,0.0003481417,0.0004486917,0.004393993,0.004227235,0.00009590934,0.6467705,0.06249544,0.2683651],"study_design_scores_gemma":[0.0003823108,0.0001928626,0.001195325,0.01653476,0.0002135924,0.0001723362,0.008403186,0.007655465,0.0003110711,0.9227388,0.04208197,0.0001183686],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.007774075,0.09006819,0.08714192,0.6963973,0.009139501,0.003093526,0.0007306919,0.0004375451,0.1052173],"genre_scores_gemma":[0.7018377,0.05530583,0.1452956,0.06968587,0.008317306,0.01178001,0.0006288679,0.0003054665,0.006843332],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.3484131,"threshold_uncertainty_score":0.8035227,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2948770661","doi":"10.1002/ev.20365","title":"Sustainability‐Ready Evaluation: A Call to Action","year":2019,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":30,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Canadian Evaluation Society","funders":"","keywords":"Sustainability; Transformative learning; Sustainability science; Sustainability organizations; Relevance (law); Process management; Sustainable development; Social sustainability; Environmental resource management; Checklist; Natural resource; Action (physics); Business; Engineering ethics; Management science; Environmental planning; Computer science; Risk analysis (engineering); Political science; Sociology; Engineering; Psychology; Economics; Ecology","authors":[{"name":"Andy Rowe","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3058992048136394,"gpt":0.5825784932667877,"spread":0.2766792884531483,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.6195292,0.002738876,0.005974706,0.007788901,0.01369883,0.04218663,0.01029457,0.04015231,0.01429921],"category_scores_gemma":[0.5939354,0.002202036,0.0047507,0.005039988,0.04743597,0.04938367,0.02784846,0.05155265,0.003248448],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02999712,"about_ca_system_score_gemma":0.1774504,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.009849535,"about_ca_topic_score_gemma":0.01095528,"domain_scores_codex":[0.3474703,0.5337562,0.0307584,0.007477756,0.07155301,0.008984444],"domain_scores_gemma":[0.1656388,0.6587371,0.01511685,0.031805,0.1000885,0.0286138],"domain_codex":"methods","domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002505973,0.0007914006,0.001716862,0.004881787,0.0001617525,0.0004010921,0.008684715,0.001867567,0.0005565382,0.2335049,0.4462766,0.3009061],"study_design_scores_gemma":[0.0002856501,0.0003942348,0.002126624,0.02445687,0.0001266564,0.0004148532,0.01672015,0.002403154,0.0007515468,0.4055137,0.5463141,0.0004924672],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.0007430742,0.006136144,0.02500995,0.952478,0.004246571,0.0007155819,0.00005346606,0.0003053509,0.01031191],"genre_scores_gemma":[0.1020285,0.01735385,0.3152071,0.5364714,0.009351172,0.006784605,0.0003746433,0.0007095642,0.01171921],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.6195292,"threshold_uncertainty_score":0.4691883,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1986224130","doi":"10.1002/ev.133","title":"Background and history of the Joint Committee's Program Evaluation Standards","year":2004,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":27,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Queen's University","funders":"","keywords":"Joint (building); Context (archaeology); Political science; Library science; Engineering ethics; Computer science; Engineering; History; Archaeology; Civil engineering","authors":[{"name":"Donald B. Yarbrough","is_ca":false},{"name":"Lyn M. Shulha","is_ca":true},{"name":"Flora A Caruthers","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3548191126107928,"gpt":0.5240969201650518,"spread":0.169277807554259,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1708321,0.001137705,0.001510802,0.01792486,0.01120726,0.02390107,0.005849861,0.007421306,0.005899972],"category_scores_gemma":[0.2029462,0.00190257,0.001403974,0.02078603,0.01951945,0.006872911,0.0059042,0.01598931,0.001859508],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.1155207,"about_ca_system_score_gemma":0.2878299,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.6719794,"about_ca_topic_score_gemma":0.6261027,"domain_scores_codex":[0.8385463,0.04336849,0.01376343,0.007406555,0.08941179,0.007503518],"domain_scores_gemma":[0.6699486,0.07358265,0.00935027,0.01953854,0.2199178,0.00766213],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00003120284,0.0000706906,0.001046182,0.0005757432,0.00002667048,0.00005890675,0.002904352,0.001602531,0.000308692,0.7112929,0.1492247,0.1328574],"study_design_scores_gemma":[0.000009573116,0.00003037576,0.002824103,0.002540347,0.00002139563,0.00004607184,0.000884995,0.0006134872,0.0004379679,0.04216322,0.9503427,0.00008585276],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.00448101,0.08445513,0.111445,0.3109683,0.0105205,0.001740404,0.002128599,0.001080728,0.4731803],"genre_scores_gemma":[0.2599772,0.123507,0.2162232,0.07889466,0.01229955,0.004690323,0.004365489,0.002322374,0.2977202],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6719794,"threshold_uncertainty_score":0.9034573,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3203797720","doi":"10.1002/ev.20478","title":"An introduction to policy advocacy evaluation: The concepts, history, and literature of the field","year":2021,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":25,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Impact","funders":"","keywords":"Field (mathematics); Adaptation (eye); Policy development; Engineering ethics; Management science; Evaluation methods; Computer science; Political science; Public administration; Psychology","authors":[{"name":"Jared Raynor","is_ca":false},{"name":"Julia Coffman","is_ca":false},{"name":"Sarah Stachowiak","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.12003077569205,"gpt":0.5220412336486495,"spread":0.4020104579565995,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02413883,0.001069984,0.001230957,0.008949646,0.004066679,0.01300165,0.001574761,0.005846349,0.008969664],"category_scores_gemma":[0.03859368,0.0007639079,0.001239422,0.008137804,0.0140061,0.0115489,0.004717198,0.01070697,0.00257341],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.00736585,"about_ca_system_score_gemma":0.01022524,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003155035,"about_ca_topic_score_gemma":0.004004639,"domain_scores_codex":[0.9788159,0.01493328,0.001692483,0.0008532145,0.003273228,0.0004319236],"domain_scores_gemma":[0.9158549,0.07394528,0.002183668,0.001629936,0.005074736,0.001311441],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00005984923,0.0001673986,0.0009841742,0.003398828,0.00002932879,0.0002011593,0.004357541,0.0008645074,0.0004041214,0.607442,0.1627096,0.2193815],"study_design_scores_gemma":[0.0000106865,0.00006617831,0.0008214323,0.009038984,0.00001010298,0.0003432119,0.001494419,0.0005413572,0.0001567873,0.203991,0.783473,0.00005295092],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"review","genre_gemma":"review","genre_scores_codex":[0.002149277,0.4931101,0.1120968,0.2651191,0.02442179,0.0009063494,0.0007069637,0.0004099239,0.1010797],"genre_scores_gemma":[0.06920272,0.6011276,0.1878337,0.06786297,0.0373251,0.003385243,0.0005965391,0.0005737994,0.03209233],"genre_candidate":"review","genre_consensus":"review","teacher_disagreement_score":0.02413883,"threshold_uncertainty_score":0.1276599,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1929107339","doi":"10.1002/ev.20078","title":"Cross‐Case Analysis and Implications for Research, Theory, and Practice","year":2014,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":25,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université du Québec en Outaouais; University of Ottawa","funders":"","keywords":"Leverage (statistics); Theory of change; Organizational change; Affect (linguistics); Thematic analysis; Context (archaeology); Order (exchange); Evaluation methods; Organization development; Knowledge management; Management science; Public relations; Sociology; Computer science; Political science; Management; Business; Qualitative research; Social science; Engineering; Economics","authors":[{"name":"J. Bradley Cousins","is_ca":true},{"name":"Isabelle Bourgeois","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4689937918488332,"gpt":0.6700118587679306,"spread":0.2010180669190974,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.06023709,0.00063436,0.001013411,0.007248352,0.006627776,0.008287329,0.003093448,0.002132547,0.011353],"category_scores_gemma":[0.1150928,0.0005686231,0.0008177477,0.009334709,0.005971221,0.01047199,0.005172702,0.002305213,0.0007305607],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01041882,"about_ca_system_score_gemma":0.004112858,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005784805,"about_ca_topic_score_gemma":0.007603145,"domain_scores_codex":[0.9115549,0.07308399,0.00290638,0.001988372,0.008336085,0.002130197],"domain_scores_gemma":[0.7665394,0.2087662,0.004003285,0.006596324,0.01313541,0.0009593751],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","study_design_scores_codex":[0.0002914833,0.002692953,0.04455267,0.001903174,0.000225161,0.002554465,0.2124325,0.004057533,0.0009415005,0.4973651,0.01226045,0.2207229],"study_design_scores_gemma":[0.0001018365,0.0005747923,0.04062124,0.006580276,0.0001810851,0.001441648,0.57342,0.01707375,0.002725068,0.2486636,0.1084819,0.0001348229],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"methods","genre_scores_codex":[0.5561095,0.01125931,0.185009,0.02581008,0.0007840673,0.006658667,0.0006463791,0.000215245,0.2135077],"genre_scores_gemma":[0.9089054,0.003552183,0.07459771,0.001146503,0.0001116504,0.00367031,0.0003797244,0.00005227379,0.007584261],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.06023709,"threshold_uncertainty_score":0.318568,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1992053221","doi":"10.1002/ev.325","title":"Enhancing disaster and emergency preparedness, response, and recovery through evaluation","year":2010,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":24,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Social Sciences and Humanities Research Council","funders":"","keywords":"Preparedness; Disaster response; Emergency management; Disaster preparedness; Emergency response; Process management; Evaluation methods; Disaster recovery; Disaster planning; Computer science; Business; Risk analysis (engineering); Computer security; Public relations; Political science; Medical emergency; Poison control; Human factors and ergonomics; Engineering; Medicine; Reliability engineering","authors":[{"name":"Liesel A. Ritchie","is_ca":false},{"name":"Wayne MacDonald","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1637329626934722,"gpt":0.5114703559569701,"spread":0.3477373932634978,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05634214,0.0005652581,0.0007059893,0.002700732,0.001690956,0.009153529,0.0009931433,0.001391173,0.003180807],"category_scores_gemma":[0.04876889,0.0002053618,0.0005154315,0.001859372,0.005216806,0.00513807,0.004670579,0.001644118,0.0002692824],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.005687425,"about_ca_system_score_gemma":0.01430773,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.002107243,"about_ca_topic_score_gemma":0.002484366,"domain_scores_codex":[0.9258814,0.06546298,0.001304873,0.0005805253,0.005559944,0.001210231],"domain_scores_gemma":[0.9468051,0.04152282,0.003234548,0.001423435,0.005011362,0.002002747],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0003429461,0.0008738523,0.009667818,0.002164637,0.0001103345,0.0001487195,0.004548889,0.01235526,0.0008582681,0.6403917,0.01082173,0.3177159],"study_design_scores_gemma":[0.0006951861,0.002478592,0.01904065,0.01259424,0.0004505431,0.0005422341,0.02126828,0.06841223,0.007942213,0.7217634,0.1445262,0.0002863564],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.1458093,0.01593145,0.3406585,0.09418296,0.0005817396,0.002806809,0.0002097448,0.0005202076,0.3992993],"genre_scores_gemma":[0.9131467,0.004007331,0.07818524,0.001452302,0.00008948085,0.0005127036,0.00004352311,0.00001944826,0.002543337],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.05634214,"threshold_uncertainty_score":0.2979693,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2097486283","doi":"10.1002/ev.20045","title":"Rethinking the Relationship Among Monitoring, Evaluation, and Results‐Based Management: Observations From Canada","year":2013,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Public Policy and Administration Research","field":"Social Sciences","cited_by":24,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Accountability; Government (linguistics); Monitoring and evaluation; Dual (grammatical number); Performance management; Process management; Political science; Public administration; Business; Marketing","authors":[{"name":"Robert Lahey","is_ca":false},{"name":"Steffen Bohni Nielsen","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2437606918516431,"gpt":0.4279213200316784,"spread":0.1841606281800354,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011466,0.0004011864,0.0006440621,0.00302773,0.01596274,0.006742591,0.002367823,0.001556251,0.00188733],"category_scores_gemma":[0.03386444,0.0006261637,0.0004197457,0.008812846,0.005541604,0.00191954,0.003534056,0.003706168,0.0001738915],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.1647457,"about_ca_system_score_gemma":0.2025025,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.997729,"about_ca_topic_score_gemma":0.9988122,"domain_scores_codex":[0.9887639,0.002744188,0.000459452,0.0006673866,0.004996939,0.002368086],"domain_scores_gemma":[0.9546742,0.01561244,0.002243574,0.001009875,0.02328623,0.003173626],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.0005439863,0.0003982697,0.3567989,0.0004687259,0.0001097757,0.00228217,0.4727893,0.001918228,0.002420214,0.02762776,0.03199057,0.1026522],"study_design_scores_gemma":[0.00005242232,0.0001277655,0.440015,0.0007114069,0.00009533844,0.0002165499,0.4324707,0.003592517,0.001840055,0.002091483,0.1185394,0.0002473549],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9407244,0.00244807,0.001250696,0.02229459,0.00007430742,0.0001934352,0.0008819214,0.00004920736,0.03208338],"genre_scores_gemma":[0.9920089,0.001127971,0.0010757,0.0012018,0.00001251051,0.00002566819,0.0001544796,0.00003710049,0.004355716],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.8352543,"threshold_uncertainty_score":0.9687766,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1999716079","doi":"10.1002/ev.243","title":"Infusing evaluative thinking as process use: The case of the International Development Research Centre (IDRC)","year":2007,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":21,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Accountability; Process (computing); Political science; Sociology; Process management; Management; Business; Computer science","authors":[{"name":"Fred Carden","is_ca":false},{"name":"Sarah Earl","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.453763683231604,"gpt":0.6131986765418254,"spread":0.1594349933102213,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.07510892,0.0007956385,0.0008646993,0.003141938,0.03268916,0.03302682,0.003755132,0.008534928,0.004342149],"category_scores_gemma":[0.0624779,0.000721127,0.0008868957,0.003313278,0.0562201,0.01120705,0.01054665,0.01330531,0.0005061783],"about_ca_system_candidate":true,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.05935515,"about_ca_system_score_gemma":0.04600045,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.2607748,"about_ca_topic_score_gemma":0.2470183,"domain_scores_codex":[0.8941929,0.08088566,0.00110162,0.002074852,0.01335931,0.008385685],"domain_scores_gemma":[0.9013883,0.07718405,0.003407351,0.004654597,0.009119047,0.004246599],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","study_design_scores_codex":[0.00009712376,0.0001665049,0.002739242,0.0001505,0.00001732002,0.004585906,0.3968565,0.001442252,0.0004710708,0.5642857,0.007429224,0.02175868],"study_design_scores_gemma":[0.00015784,0.0002281624,0.004181713,0.001334328,0.00009082045,0.002409658,0.5062369,0.008259964,0.003150814,0.1302867,0.3434628,0.0002003085],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.310509,0.002867549,0.04023669,0.1087948,0.00034735,0.0006891305,0.00007226004,0.0003077982,0.5361754],"genre_scores_gemma":[0.9752088,0.0005779939,0.007978215,0.002624129,0.00004526568,0.0001691585,0.0000121253,0.0001147532,0.01326967],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9406449,"threshold_uncertainty_score":0.5185138,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2070292114","doi":"10.1002/ev.393","title":"Internal evaluation, historically speaking","year":2011,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":20,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Centre for Advancing Health Outcomes; University of British Columbia","funders":"","keywords":"Context (archaeology); Conservatism; Speculation; Government (linguistics); Function (biology); Evaluation methods; Internal validity; Political science; Public administration; Sociology; Public relations; Law; Business; History","authors":[{"name":"Sandra Mathison","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4596669989734262,"gpt":0.5336734648069541,"spread":0.07400646583352793,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02384737,0.0006609487,0.0007017195,0.004428807,0.003039782,0.01359882,0.001118431,0.001252656,0.004834597],"category_scores_gemma":[0.04598076,0.0002663316,0.0004646526,0.003956595,0.01508999,0.006884928,0.00503629,0.002636632,0.0009644075],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01294272,"about_ca_system_score_gemma":0.009247742,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003514384,"about_ca_topic_score_gemma":0.00288354,"domain_scores_codex":[0.9694826,0.01654276,0.001715231,0.002024217,0.009029339,0.001205939],"domain_scores_gemma":[0.9543013,0.0184663,0.004126866,0.003221247,0.01851786,0.001366437],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00007345891,0.0001028194,0.005645331,0.0007370454,0.00002870709,0.0001204233,0.007165432,0.0008467087,0.000432552,0.6817874,0.02084003,0.2822201],"study_design_scores_gemma":[0.00005598283,0.0003247225,0.01880724,0.009207484,0.0001098496,0.0008572637,0.01723104,0.003523801,0.004679136,0.3831814,0.561885,0.0001371185],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.04296228,0.04430625,0.06992962,0.05119978,0.002611591,0.0003343757,0.0002332645,0.0003409967,0.7880818],"genre_scores_gemma":[0.9067696,0.01796104,0.0194126,0.005924142,0.001268406,0.000359186,0.0001827987,0.0002093843,0.04791286],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9761527,"threshold_uncertainty_score":0.1261185,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3011548789","doi":"10.1002/ev.20398","title":"Finding the Impact: Methods for Assessing the Contribution of Collective Impact to Systems and Population Change in a Multi‐Site Study","year":2020,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Impact","funders":"","keywords":"Scale (ratio); Context (archaeology); Interim; Population; Work (physics); SPARK (programming language); Management science; Computer science; Political science; Sociology; Engineering","authors":[{"name":"Sarah Stachowiak","is_ca":true},{"name":"Jewlya Lynn","is_ca":false},{"name":"Terri Akey","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5624400587745525,"gpt":0.6718887697007279,"spread":0.1094487109261754,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1834699,0.001609687,0.001954047,0.009344366,0.00260282,0.006816916,0.003377112,0.001828048,0.00645298],"category_scores_gemma":[0.3033651,0.001093946,0.003634382,0.009802483,0.005649056,0.004720349,0.006933912,0.003378124,0.0006333433],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004073071,"about_ca_system_score_gemma":0.004909451,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005815359,"about_ca_topic_score_gemma":0.007199498,"domain_scores_codex":[0.7363665,0.2366756,0.007560759,0.006175041,0.01242048,0.0008015945],"domain_scores_gemma":[0.4712714,0.4623832,0.02331467,0.0221431,0.01933436,0.001553385],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.001277818,0.001483436,0.1639545,0.008790137,0.005448068,0.0005072349,0.0555578,0.02240628,0.002963448,0.1292372,0.01710087,0.5912732],"study_design_scores_gemma":[0.001577588,0.006384911,0.2350095,0.01153538,0.00494394,0.0006945196,0.07999187,0.1922801,0.008601093,0.3595571,0.09800996,0.001414132],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"empirical","genre_scores_codex":[0.06755281,0.001413083,0.8995656,0.001598169,0.0004194983,0.01359371,0.00155678,0.0005557582,0.01374458],"genre_scores_gemma":[0.1782054,0.0004577779,0.7900923,0.0003270775,0.00006835969,0.02892352,0.0003779718,0.0001497588,0.0013979],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.1834699,"threshold_uncertainty_score":0.970293,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1781948974","doi":"10.1002/ev.20111","title":"How Does Accreditation Fit Into the Picture?","year":2015,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Accounting Education and Careers","field":"Business, Management and Accounting","cited_by":17,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Victoria","funders":"","keywords":"Accreditation; Blank; Computer science; Public relations; Political science; Law; Engineering","authors":[{"name":"James C. McDavid","is_ca":true},{"name":"Irene Huse","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.06017099692544828,"gpt":0.3194472065074327,"spread":0.2592762095819845,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0816688,0.0008534237,0.002054173,0.0097895,0.006907308,0.02863356,0.002834766,0.01106089,0.0100566],"category_scores_gemma":[0.1841661,0.0007493458,0.001305825,0.007753699,0.02697164,0.06093667,0.007664841,0.01864472,0.002372281],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01282076,"about_ca_system_score_gemma":0.02150136,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01284018,"about_ca_topic_score_gemma":0.01030974,"domain_scores_codex":[0.9288412,0.04479235,0.003861535,0.003111777,0.01576034,0.003632693],"domain_scores_gemma":[0.8256001,0.1085766,0.006768341,0.006301424,0.04571233,0.007041019],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00006507406,0.0001226057,0.005320643,0.001252681,0.00008242324,0.0001265725,0.00332362,0.0006271497,0.0001127911,0.7279004,0.1201998,0.1408662],"study_design_scores_gemma":[0.00002542652,0.0001106928,0.008014156,0.00915292,0.00007206246,0.0002761525,0.02126612,0.001576472,0.0002513822,0.6532779,0.3058355,0.0001413677],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.003226737,0.0429379,0.009411978,0.8938583,0.004723117,0.00004612456,0.0001152481,0.0001492221,0.04553144],"genre_scores_gemma":[0.6057115,0.07987799,0.02323384,0.2616233,0.01451108,0.0003616931,0.0004073542,0.0007313827,0.01354198],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.0816688,"threshold_uncertainty_score":0.4319111,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2013250096","doi":"10.1002/ev.20007","title":"When one must go: The Canadian experience with strategic review and judging program value","year":2012,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Valuation (finance); Function (biology); Government (linguistics); Value (mathematics); Citizen journalism; Public relations; Process (computing); Value for money; Public administration; Business; Political science; Economics; Public economics; Computer science; Accounting; Law","authors":[{"name":"François Dumaine","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.38211926976493,"gpt":0.524848788433648,"spread":0.142729518668718,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.09253305,0.0005500506,0.001054967,0.005499267,0.02775499,0.02435435,0.003408269,0.003487343,0.004857157],"category_scores_gemma":[0.1252026,0.0005727977,0.0005773462,0.009955995,0.02187442,0.00622281,0.006727309,0.006237126,0.0004621652],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.18534,"about_ca_system_score_gemma":0.3177511,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9346248,"about_ca_topic_score_gemma":0.9671484,"domain_scores_codex":[0.8704042,0.06674352,0.003536206,0.003643278,0.04592741,0.009745293],"domain_scores_gemma":[0.816703,0.09481774,0.00465567,0.004585266,0.06610148,0.01313684],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","study_design_scores_codex":[0.0002795569,0.0002088669,0.01164888,0.0009037015,0.00008893257,0.0006647203,0.1208501,0.00226804,0.0008110615,0.4096863,0.09912063,0.3534692],"study_design_scores_gemma":[0.00008693124,0.0001821595,0.01823564,0.001563088,0.00008567153,0.0002553582,0.09156796,0.002389026,0.0014439,0.06471133,0.8190959,0.0003829771],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.1272884,0.02479639,0.01740785,0.2371969,0.001854232,0.0006175029,0.0002712269,0.0002646928,0.5903028],"genre_scores_gemma":[0.9389402,0.01231819,0.01565463,0.008676007,0.0002673057,0.0001756927,0.0001200229,0.0001508871,0.02369713],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9074669,"threshold_uncertainty_score":0.9448901,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4388306673","doi":"10.1002/ev.20564","title":"Artificial intelligence and the future of evaluation education: Possibilities and prototypes","year":2023,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":15,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Fraser Health","funders":"","keywords":"Chatbot; Evaluation methods; Engineering ethics; Literacy; Computer science; Program evaluation; Management science; Psychology; Pedagogy; Artificial intelligence; Political science; Engineering","authors":[{"name":"John M. LaVelle","is_ca":false},{"name":"Tian Ford","is_ca":false},{"name":"María Montenegro","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2331504610129478,"gpt":0.5285796286386468,"spread":0.295429167625699,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.04652652,0.000528688,0.000700078,0.00203292,0.002924613,0.01477628,0.002962243,0.004164731,0.006922479],"category_scores_gemma":[0.03292329,0.0004145784,0.0006387548,0.0011818,0.01958588,0.01695871,0.004686157,0.004338104,0.0006021291],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.005388487,"about_ca_system_score_gemma":0.005827419,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003153087,"about_ca_topic_score_gemma":0.001776822,"domain_scores_codex":[0.9748905,0.02119292,0.0006471915,0.0005376567,0.001770107,0.0009616725],"domain_scores_gemma":[0.9442628,0.04624326,0.000890327,0.002895908,0.003544377,0.002163364],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002207172,0.0005669452,0.001591782,0.0005291066,0.00001503308,0.0002057377,0.007126331,0.002406661,0.0005081817,0.8734421,0.004766324,0.1086211],"study_design_scores_gemma":[0.000217884,0.0008176088,0.001272784,0.002882975,0.00002787644,0.0004103971,0.01422166,0.01802012,0.002040655,0.7973178,0.1626463,0.0001237952],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.1317136,0.05004778,0.1831253,0.2710562,0.001886824,0.0007982326,0.0001336076,0.001150916,0.3600874],"genre_scores_gemma":[0.8873028,0.008848689,0.09124804,0.003200431,0.0002922638,0.0005289186,0.00006772172,0.00008733855,0.008423749],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.04652652,"threshold_uncertainty_score":0.2460587,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2033749959","doi":"10.1002/ev.1190","title":"Evaluative inquiry in university‐school professional learning partnerships","year":2000,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":12,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Queen's University","funders":"","keywords":"Professional learning community; Professional development; Pedagogy; Psychology; Sociology","authors":[{"name":"Lyn M. Shulha","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3781556627691831,"gpt":0.5350127287443462,"spread":0.1568570659751632,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1256486,0.000383734,0.0006862238,0.003030009,0.01199347,0.01835717,0.002119464,0.003382769,0.005174762],"category_scores_gemma":[0.1748495,0.0004674642,0.0003127631,0.003423545,0.02249571,0.01207593,0.01696897,0.003813579,0.0003569609],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01039858,"about_ca_system_score_gemma":0.02525286,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001998315,"about_ca_topic_score_gemma":0.003189683,"domain_scores_codex":[0.7794903,0.2090036,0.001513865,0.001365266,0.005219125,0.003407876],"domain_scores_gemma":[0.6830457,0.2788817,0.00903766,0.006044535,0.009991335,0.01299908],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","study_design_scores_codex":[0.0003326424,0.001613138,0.01373926,0.0007785853,0.000048869,0.0006572384,0.1470829,0.003712404,0.0002661088,0.6782577,0.01068067,0.1428305],"study_design_scores_gemma":[0.000317253,0.0007261592,0.005964408,0.001684141,0.00003420753,0.0003450912,0.336419,0.007433502,0.001111652,0.5543759,0.09150466,0.00008394744],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.5466221,0.008619864,0.08265789,0.07620462,0.0007158782,0.001518171,0.00005446498,0.0002433922,0.2833637],"genre_scores_gemma":[0.9855182,0.0008479455,0.009908015,0.0006330446,0.00005026175,0.0005923273,0.00001013092,0.00001356531,0.002426445],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1256486,"threshold_uncertainty_score":0.6645012,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4224927296","doi":"10.1002/ev.20494","title":"Evaluation policy and organizational evaluation capacity building: A study of international aid agency evaluation policies","year":2022,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":11,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Agency (philosophy); Pace; Evaluation methods; Program evaluation; Policy analysis; Capacity building; Public relations; Management science; Knowledge management; Process management; Business; Political science; Computer science; Public administration; Economics; Sociology; Economic growth","authors":[{"name":"Hind Al Hudib","is_ca":true},{"name":"J. Bradley Cousins","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3393749749893569,"gpt":0.529793832087028,"spread":0.1904188570976711,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0828068,0.000308885,0.0004697217,0.003294694,0.01113347,0.01318018,0.002181374,0.002357321,0.003479148],"category_scores_gemma":[0.1544629,0.0006378003,0.000312365,0.00492015,0.01583365,0.01031512,0.007108046,0.005545232,0.0002852233],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02980999,"about_ca_system_score_gemma":0.03140177,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.02502399,"about_ca_topic_score_gemma":0.0144709,"domain_scores_codex":[0.9159221,0.07088581,0.00183486,0.001686611,0.004270644,0.005400045],"domain_scores_gemma":[0.7736757,0.1698823,0.02197598,0.008099886,0.01809195,0.008274129],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.0003491015,0.002480082,0.128908,0.0003977111,0.00006919956,0.0005781027,0.4578199,0.004647769,0.0006774105,0.3280821,0.006954959,0.06903557],"study_design_scores_gemma":[0.00008995754,0.0003837722,0.07272197,0.0007132109,0.000033393,0.0001673505,0.8311166,0.007592757,0.0008959815,0.03598158,0.050205,0.0000985444],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.9334589,0.0003877122,0.005339326,0.01137705,0.00003200043,0.0002961906,0.0000453042,0.00002902988,0.04903444],"genre_scores_gemma":[0.9979745,0.0001135365,0.0007457923,0.0003035121,0.000006481695,0.0001469429,0.00000950154,0.000006473416,0.0006933911],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.0828068,"threshold_uncertainty_score":0.4379295,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2008070236","doi":"10.1002/ev.299","title":"Reflections on the dilemmas of conducting environmental evaluations","year":2009,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":10,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Impact","funders":"","keywords":"Cognitive reframing; Multidisciplinary approach; Asset (computer security); Context (archaeology); Set (abstract data type); Face (sociological concept); Computer science; Management science; Theory of change; Work (physics); Sociology; Engineering ethics; Psychology; Social science; Social psychology; Economics; Engineering","authors":[{"name":"Hallie Preskill","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5676539202947379,"gpt":0.5954441216399692,"spread":0.02779020134523125,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.5950261,0.001021566,0.002511029,0.00386544,0.01427812,0.03957076,0.008001775,0.02092451,0.005247561],"category_scores_gemma":[0.6618935,0.001829971,0.002032722,0.004007165,0.06077994,0.03714037,0.01743752,0.03868965,0.001452724],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02578024,"about_ca_system_score_gemma":0.04382318,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.007999527,"about_ca_topic_score_gemma":0.007478223,"domain_scores_codex":[0.2785682,0.6316417,0.02159431,0.009080901,0.05169005,0.00742483],"domain_scores_gemma":[0.149106,0.7760755,0.008158681,0.01102368,0.04733505,0.008301123],"domain_codex":"methods","domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002574074,0.0002598056,0.002214369,0.001814513,0.0001224989,0.0007188728,0.1176356,0.002155673,0.0003235474,0.518586,0.2539023,0.1020094],"study_design_scores_gemma":[0.0002369321,0.0002549067,0.001698059,0.008625879,0.00005147261,0.0006319467,0.1268068,0.003782362,0.0009956168,0.3978921,0.4587492,0.0002745647],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.005429896,0.003965845,0.01082807,0.96035,0.002373127,0.0002984233,0.00004304525,0.00009310204,0.0166186],"genre_scores_gemma":[0.5010082,0.009382629,0.08882508,0.3766243,0.006138661,0.005013986,0.00008852689,0.0008024023,0.01211621],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.5950261,"threshold_uncertainty_score":0.4994049,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2891656379","doi":"10.1002/ev.20346","title":"The Evolving Market for Systematic Evaluation in Canada","year":2018,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":10,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of Ottawa","funders":"","keywords":"Accreditation; Professionalization; Context (archaeology); Supply and demand; Government (linguistics); Supply side; Business; Exploratory research; Affect (linguistics); Public relations; Marketing; Economics; Political science; Economic growth; Sociology; Social science","authors":[{"name":"Robert Lahey","is_ca":false},{"name":"Catherine Elliott","is_ca":true},{"name":"Sarah Heath","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1834456181021273,"gpt":0.4959395639226441,"spread":0.3124939458205168,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.02246012,0.0003545073,0.0006082545,0.005326409,0.01264898,0.01817266,0.003258982,0.00353325,0.01637782],"category_scores_gemma":[0.04423783,0.0006729079,0.0006476109,0.006498698,0.008398074,0.003511337,0.004061796,0.003842383,0.0007719639],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.3124715,"about_ca_system_score_gemma":0.496807,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.9862328,"about_ca_topic_score_gemma":0.9915679,"domain_scores_codex":[0.9721674,0.003403787,0.0007688976,0.001756662,0.01585431,0.006048918],"domain_scores_gemma":[0.8691022,0.02664624,0.00414524,0.00211498,0.07055154,0.02743978],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"observational","study_design_scores_codex":[0.0002821017,0.0002435487,0.03377812,0.0007929663,0.00004902462,0.001052269,0.01021193,0.002144793,0.001660758,0.4913378,0.2053912,0.2530555],"study_design_scores_gemma":[0.00008565362,0.0001252266,0.08197357,0.00169771,0.00004229241,0.0002412947,0.0222571,0.005979823,0.001023865,0.02618947,0.8601332,0.0002509431],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1453054,0.02256622,0.006338318,0.5465443,0.001308289,0.0004284469,0.002056516,0.0003717842,0.2750808],"genre_scores_gemma":[0.9082901,0.009976946,0.007400031,0.02703498,0.0003220742,0.0001114317,0.0005634564,0.0001576543,0.04614318],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9775399,"threshold_uncertainty_score":0.7974356,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1557717928","doi":"10.1002/ev.20082","title":"The Value in Validity","year":2014,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"123 Certification (Canada)","funders":"","keywords":"Beauty; Prioritization; Economic Justice; Value (mathematics); Sociology; Balance (ability); Management science; Psychology; Social psychology; Epistemology; Computer science; Political science; Law; Economics","authors":[{"name":"James C. Griffith","is_ca":false},{"name":"Bianca Montrosse‐Moorhead","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3622601772662026,"gpt":0.5450118219863205,"spread":0.1827516447201178,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1356496,0.0009395056,0.001784645,0.005638207,0.007040419,0.02183406,0.002986968,0.005960945,0.005097989],"category_scores_gemma":[0.212938,0.0007036093,0.001061446,0.003892685,0.1162281,0.02183366,0.01295385,0.01029559,0.0007903806],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01743762,"about_ca_system_score_gemma":0.01624304,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.006053765,"about_ca_topic_score_gemma":0.003205834,"domain_scores_codex":[0.7128092,0.2263263,0.007062119,0.008287271,0.04133485,0.004180312],"domain_scores_gemma":[0.6576813,0.2816871,0.007613819,0.02238187,0.02671572,0.003920265],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00001837513,0.000009838239,0.0004605115,0.0001090282,0.00001587459,0.00002733467,0.001432879,0.0002691621,0.00003965101,0.9836407,0.003337322,0.01063931],"study_design_scores_gemma":[0.00001293375,0.0000195057,0.0003078637,0.0006784673,0.00001152324,0.0000489033,0.001174336,0.0005626293,0.0002011643,0.9700435,0.02691978,0.00001939432],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"other","genre_gemma":"methods","genre_scores_codex":[0.02108645,0.02910824,0.1352781,0.3181277,0.003577662,0.0003395003,0.0002060828,0.0001372509,0.4921389],"genre_scores_gemma":[0.9234975,0.006707184,0.03781762,0.01909066,0.002161998,0.0004995837,0.00006927065,0.0001726512,0.009983559],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.1356496,"threshold_uncertainty_score":0.7173922,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4214553715","doi":"10.1002/ev.20486","title":"How does teaching with cases support the development of evaluation competencies?","year":2021,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Adaptability; Craft; Evaluation methods; Professional development; Medical education; Psychology; Computer science; Knowledge management; Pedagogy; Engineering; Medicine; Management","authors":[{"name":"Isabelle Bourgeois","is_ca":true},{"name":"Lauren Wildschut","is_ca":false},{"name":"Marla Steinberg","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2469212988547501,"gpt":0.4932914012784128,"spread":0.2463701024236627,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.03330185,0.0002597346,0.0003596317,0.002055777,0.001166839,0.006728659,0.00158917,0.001389151,0.005625243],"category_scores_gemma":[0.2272477,0.0003073086,0.0002615741,0.001161307,0.002230731,0.00766698,0.003206693,0.0009818648,0.0008581926],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003022715,"about_ca_system_score_gemma":0.004198757,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001874897,"about_ca_topic_score_gemma":0.002945815,"domain_scores_codex":[0.9534466,0.03686998,0.001295854,0.001233471,0.005646756,0.001507373],"domain_scores_gemma":[0.7016456,0.2482553,0.01787158,0.009741073,0.01459912,0.007887281],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.0004222716,0.003364269,0.1768193,0.0007187786,0.00005133732,0.00141216,0.05586551,0.001609218,0.00326692,0.02959008,0.01223501,0.7146452],"study_design_scores_gemma":[0.0004984447,0.004058716,0.3153518,0.006357336,0.0003319883,0.005376136,0.2348783,0.04325372,0.02195491,0.1273674,0.2401996,0.0003716876],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8903887,0.0008037935,0.03429039,0.0131345,0.0001838513,0.0004574522,0.00007500233,0.0003351806,0.0603313],"genre_scores_gemma":[0.9887673,0.0003303178,0.00963357,0.000279714,0.0000307558,0.00005957308,0.00002532066,0.0000179842,0.0008555561],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.03330185,"threshold_uncertainty_score":0.1761191,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2278381810","doi":"10.1002/ev.20158","title":"Assessing the Practice Impact of Research on Evaluation","year":2015,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Evaluation methods; Program evaluation; Psychology; Applied psychology; Computer science; Political science; Public administration; Reliability engineering; Engineering","authors":[{"name":"J. Bradley Cousins","is_ca":true},{"name":"Kateryna Svensson","is_ca":true},{"name":"Barbara Szijarto","is_ca":true},{"name":"Celine Pinsent","is_ca":true},{"name":"Caroline Andrew","is_ca":true},{"name":"John Sylvestre","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.8085447410014557,"gpt":0.7537715169629834,"spread":0.05477322403847229,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":["metaresearch"],"category_scores_codex":[0.5486119,0.001986606,0.003279007,0.0151602,0.004930666,0.02324456,0.004394612,0.006886628,0.008151031],"category_scores_gemma":[0.8136303,0.001286655,0.002442716,0.01451774,0.02218051,0.02561414,0.0182868,0.009095642,0.0009189603],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.03119584,"about_ca_system_score_gemma":0.03684049,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00488868,"about_ca_topic_score_gemma":0.005183019,"domain_scores_codex":[0.2069506,0.6366068,0.02846944,0.01004545,0.1133498,0.004578064],"domain_scores_gemma":[0.04767953,0.8491367,0.03297289,0.02345814,0.04469252,0.002060209],"domain_codex":"methods","domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.000448033,0.0004106002,0.04314274,0.01454305,0.001755158,0.0002256001,0.01012744,0.005406774,0.0006142105,0.4498789,0.01090856,0.462539],"study_design_scores_gemma":[0.0004047686,0.003091814,0.0814416,0.0845484,0.002935428,0.0005632733,0.0378311,0.01535678,0.004885525,0.643161,0.1254424,0.0003378557],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1190571,0.127783,0.1389456,0.3180472,0.0054543,0.005183393,0.0009308378,0.0003815254,0.284217],"genre_scores_gemma":[0.8945893,0.0239605,0.06256986,0.013038,0.001237666,0.002449601,0.0002093139,0.0001148477,0.001830904],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.4513881,"threshold_uncertainty_score":0.5566419,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2067449005","doi":"10.1002/ev.326","title":"Real‐time evaluation in humanitarian emergencies","year":2010,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Social Sciences and Humanities Research Council","funders":"","keywords":"Formative assessment; Credibility; Evaluation methods; Program evaluation; Field (mathematics); Emergency management; Process management; Computer science; Monitoring and evaluation; Risk analysis (engineering); Political science; Business; Psychology; Engineering; Public administration","authors":[{"name":"Emery Brusset","is_ca":false},{"name":"John Cosgrave","is_ca":false},{"name":"Wayne MacDonald","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2369325683844146,"gpt":0.526100458986193,"spread":0.2891678906017783,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03589217,0.0004357836,0.0004483508,0.001333014,0.000623549,0.003095842,0.0005930071,0.0007894002,0.003559923],"category_scores_gemma":[0.06829165,0.0001903189,0.0003821329,0.0008453292,0.001427992,0.001899782,0.001396281,0.000943501,0.0005234301],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001617866,"about_ca_system_score_gemma":0.001820673,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.001039204,"about_ca_topic_score_gemma":0.0009825756,"domain_scores_codex":[0.9339301,0.0601411,0.001401458,0.0006246209,0.003451001,0.0004516545],"domain_scores_gemma":[0.9371894,0.04898204,0.004181448,0.001630723,0.006864717,0.001151733],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00435597,0.001727219,0.01150532,0.003326416,0.0002504732,0.0006754783,0.00626591,0.05308495,0.004820307,0.0716366,0.0184058,0.8239456],"study_design_scores_gemma":[0.003274831,0.02196015,0.05349191,0.01255002,0.0007146026,0.00345716,0.02663531,0.2357225,0.05583174,0.2248116,0.3606821,0.0008680683],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4876526,0.03789692,0.3022404,0.01455366,0.00218475,0.002966085,0.0003976052,0.0009528152,0.1511552],"genre_scores_gemma":[0.9533927,0.00314715,0.03760653,0.0003982829,0.0001541704,0.000478807,0.0000837616,0.000038256,0.00470033],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.9641078,"threshold_uncertainty_score":0.1898182,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4293180673","doi":"10.1002/ev.20490","title":"The importance of implementation: Putting evaluation policy to work","year":2022,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"McGill University","funders":"","keywords":"Work (physics); Government (linguistics); White paper; Public administration; Public relations; Early adopter; Program evaluation; Evaluation methods; Policy analysis; Business; Political science; Marketing","authors":[{"name":"Leslie A. Fierro","is_ca":true},{"name":"Alana R. Kinarsky","is_ca":false},{"name":"Carlos Echeverria‐Estrada","is_ca":false},{"name":"Nadia Sabat Bass","is_ca":false},{"name":"Christina A. Christie","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2980319978182459,"gpt":0.5961758186242162,"spread":0.2981438208059703,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.6139299,0.001659125,0.003465646,0.007433902,0.02352889,0.06750646,0.008175779,0.02836321,0.01000971],"category_scores_gemma":[0.632082,0.002360868,0.002358122,0.006064767,0.06942561,0.078748,0.02497714,0.04061349,0.002009988],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.05244821,"about_ca_system_score_gemma":0.2472461,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01588233,"about_ca_topic_score_gemma":0.01142473,"domain_scores_codex":[0.2811696,0.6168371,0.02221506,0.0110542,0.04943572,0.01928831],"domain_scores_gemma":[0.1488161,0.7270937,0.01447463,0.0255168,0.06298053,0.02111813],"domain_codex":"methods","domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002266633,0.000620101,0.004405128,0.003292124,0.0001359611,0.0002627535,0.03435894,0.001570654,0.0004852971,0.637095,0.06867458,0.2488727],"study_design_scores_gemma":[0.0003818253,0.0006357049,0.00533867,0.0291268,0.0002374036,0.0002660223,0.0906264,0.002505256,0.0021639,0.5402797,0.3281412,0.0002971124],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.004683529,0.006241201,0.02783417,0.9123735,0.002421157,0.0006969701,0.00003898644,0.0001876437,0.045523],"genre_scores_gemma":[0.608847,0.01369932,0.1147159,0.242719,0.003301472,0.004469274,0.0001651305,0.0006058203,0.01147704],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.6139299,"threshold_uncertainty_score":0.4760932,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1534140708","doi":"10.1002/ev.20088","title":"House With a View: Validity and Evaluative Argument","year":2014,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"123 Certification (Canada)","funders":"","keywords":"CLARITY; Argument (complex analysis); Style (visual arts); Epistemology; Narrative; Evaluation methods; Psychology; Computer science; Sociology; Linguistics; Philosophy; History","authors":[{"name":"Bianca Montrosse‐Moorhead","is_ca":true},{"name":"James C. Griffith","is_ca":false},{"name":"Pamela Pokorny","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.240140513715607,"gpt":0.5046303264375429,"spread":0.2644898127219358,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1004092,0.0006014818,0.001210546,0.002569561,0.006011734,0.01902502,0.002526195,0.007074974,0.006318588],"category_scores_gemma":[0.2555495,0.0007013882,0.0007399375,0.002142657,0.04504205,0.02666344,0.006939893,0.01051741,0.0008116884],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.008304209,"about_ca_system_score_gemma":0.007916982,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003517228,"about_ca_topic_score_gemma":0.002354083,"domain_scores_codex":[0.8421661,0.1307555,0.002637188,0.00410498,0.01819811,0.002138176],"domain_scores_gemma":[0.7050767,0.2628883,0.00541187,0.01021667,0.01483076,0.001575751],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00002512148,0.00002312422,0.0001770521,0.0001096693,0.00001696981,0.00007809785,0.006656816,0.0001303073,0.00005472607,0.967012,0.01663014,0.009085909],"study_design_scores_gemma":[0.00004446672,0.00005653198,0.000658752,0.001617282,0.00003361566,0.00008609361,0.01256525,0.001368654,0.000664743,0.8410231,0.1418375,0.00004384234],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"commentary","genre_gemma":"methods","genre_scores_codex":[0.01973745,0.01664108,0.06048966,0.5924214,0.007104952,0.0003522162,0.00009211645,0.0001178921,0.3030433],"genre_scores_gemma":[0.8483149,0.008094476,0.03570234,0.07275525,0.004866804,0.0008826451,0.00009471743,0.0003112113,0.02897766],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8995908,"threshold_uncertainty_score":0.5310211,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2038633574","doi":"10.1002/ev.1171","title":"Translating evaluation findings into “policy language”","year":2000,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Affect (linguistics); Welfare; Order (exchange); Policy learning; Policy analysis; Program evaluation; Language policy; Computer science; Management science; Political science; Public administration; Sociology; Pedagogy; Economics; Machine learning; Law; Finance","authors":[{"name":"Kenneth Cabatoff","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1976021849984889,"gpt":0.5590826968278899,"spread":0.3614805118294011,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.156343,0.0007346299,0.001301599,0.006424267,0.00250929,0.01765371,0.002182998,0.001788223,0.008775703],"category_scores_gemma":[0.3000719,0.0005081587,0.0006727458,0.004739382,0.00950209,0.007992976,0.005105081,0.005592985,0.001613213],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02586271,"about_ca_system_score_gemma":0.05310005,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.03829459,"about_ca_topic_score_gemma":0.02790044,"domain_scores_codex":[0.819006,0.1429754,0.006996983,0.002139971,0.02665624,0.002225277],"domain_scores_gemma":[0.6446816,0.245495,0.006467561,0.01373955,0.08783863,0.001777791],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0001413578,0.0003259135,0.002437763,0.0036428,0.0001479298,0.0002697577,0.01692639,0.005534031,0.001890888,0.6580023,0.07408161,0.2365994],"study_design_scores_gemma":[0.0002309244,0.0003251755,0.006708131,0.02674374,0.00029881,0.0001270104,0.04214897,0.008441988,0.01148432,0.3370012,0.5662746,0.0002152735],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.02907343,0.01163706,0.2213113,0.3360745,0.008373706,0.001925659,0.002165633,0.001470447,0.3879682],"genre_scores_gemma":[0.6926892,0.02017939,0.2197573,0.03662055,0.001676575,0.002992674,0.001162451,0.0009087534,0.02401309],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.156343,"threshold_uncertainty_score":0.8268305,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3196514161","doi":"10.1002/ev.20464","title":"Supporting systems transformation through design‐driven evaluation","year":2021,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Complex Systems and Decision Making","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Canadian Economics Association","funders":"","keywords":"Cognitive reframing; Computer science; Process (computing); Process management; Service (business); New product development; Knowledge management; Product (mathematics); Value (mathematics); Management science; Engineering management; Systems engineering; Engineering; Business","authors":[{"name":"Cameron D. Norman","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4219147858864458,"gpt":0.5201704823327796,"spread":0.09825569644633386,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05638894,0.001290834,0.0008962627,0.002214781,0.001404008,0.00742562,0.002333204,0.001925057,0.006768286],"category_scores_gemma":[0.1059714,0.0006905892,0.0008494824,0.001125718,0.005295665,0.005204346,0.005736333,0.002523958,0.000821056],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.005306484,"about_ca_system_score_gemma":0.01149596,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00293236,"about_ca_topic_score_gemma":0.002308627,"domain_scores_codex":[0.9308261,0.05489292,0.00199571,0.001699875,0.009439312,0.001146079],"domain_scores_gemma":[0.8685998,0.1015959,0.004243325,0.01074142,0.01336797,0.001451581],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0004135444,0.0009170544,0.004804641,0.001896457,0.0001828572,0.0002709607,0.00313423,0.1348314,0.006732847,0.5249842,0.006987066,0.3148448],"study_design_scores_gemma":[0.0006934353,0.0008458593,0.001130661,0.001342877,0.0001115708,0.0001326366,0.001584921,0.5220158,0.01877415,0.3881384,0.06509055,0.0001391086],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.01735178,0.0003303882,0.9492179,0.002622828,0.000155205,0.002240465,0.0001463962,0.00142946,0.02650553],"genre_scores_gemma":[0.3721265,0.0003154668,0.6220856,0.0005233823,0.00004025236,0.002436262,0.0002561475,0.0002393074,0.001977141],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.05638894,"threshold_uncertainty_score":0.2982167,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2055999462","doi":"10.1002/ev.392","title":"Internal evaluation a quarter‐century later: A conversation with Arnold J. Love","year":2011,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Conversation; Evaluation methods; Perspective (graphical); Association (psychology); Quarter (Canadian coin); Sociology; Psychology; Computer science; Artificial intelligence; History; Engineering","authors":[{"name":"Boris B. Volkov","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2229234698309984,"gpt":0.4641575431803136,"spread":0.2412340733493152,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.09419937,0.0006579602,0.001872243,0.002343828,0.0124392,0.02158207,0.001646522,0.01517172,0.004861577],"category_scores_gemma":[0.119316,0.0008502112,0.001356166,0.002471659,0.026157,0.01842818,0.009645966,0.04370635,0.001711575],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02585227,"about_ca_system_score_gemma":0.02190256,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.009212085,"about_ca_topic_score_gemma":0.0111806,"domain_scores_codex":[0.8968788,0.07444954,0.003813156,0.003709535,0.01665396,0.00449504],"domain_scores_gemma":[0.8701782,0.08354291,0.003248955,0.002432587,0.0236155,0.01698193],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"qualitative","study_design_scores_codex":[0.00005611399,0.0001596456,0.000728461,0.0004044189,0.0000277892,0.0007254373,0.04633964,0.0001501543,0.0002220844,0.1414363,0.7515459,0.05820408],"study_design_scores_gemma":[0.0000157348,0.00004968336,0.0005336078,0.001858749,0.0000119451,0.0005781096,0.0198112,0.0001683184,0.0001758784,0.01959663,0.9571233,0.00007682951],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"other","genre_scores_codex":[0.0008269026,0.01930354,0.0007554638,0.9642933,0.00766812,0.0000100522,0.000006820345,0.00001112215,0.007124709],"genre_scores_gemma":[0.08146647,0.0453897,0.002549585,0.8189036,0.01543609,0.0001211309,0.00002529127,0.0002109306,0.03589711],"genre_candidate":"other","genre_consensus":null,"teacher_disagreement_score":0.09419937,"threshold_uncertainty_score":0.4981798,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1994122017","doi":"10.1002/ev.333","title":"Evaluation of disaster and emergency management: Do no harm, but do better","year":2010,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Disaster Response and Management","field":"Health Professions","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Social Sciences and Humanities Research Council","funders":"","keywords":"Emergency management; Harm; Medical emergency; Business; Medicine; Psychology; Political science; Social psychology","authors":[{"name":"Liesel A. Ritchie","is_ca":false},{"name":"Wayne MacDonald","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08876767161420009,"gpt":0.4590029695750237,"spread":0.3702352979608236,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1110325,0.001257142,0.002437476,0.003216917,0.003135049,0.01502196,0.001319836,0.005115042,0.0043353],"category_scores_gemma":[0.2272376,0.0005495306,0.0008899677,0.001717295,0.01203492,0.01462227,0.006827022,0.009057412,0.001236612],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.007203163,"about_ca_system_score_gemma":0.01628825,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003294929,"about_ca_topic_score_gemma":0.004787893,"domain_scores_codex":[0.8534281,0.114472,0.007581585,0.001453055,0.02185172,0.001213547],"domain_scores_gemma":[0.7455902,0.1735506,0.01178911,0.011917,0.04802983,0.009123298],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0007054266,0.0004888189,0.00746424,0.008559809,0.0006180037,0.0002591996,0.005495972,0.00107964,0.0005714139,0.1206338,0.2495498,0.6045738],"study_design_scores_gemma":[0.0007101531,0.001202922,0.01278909,0.07143743,0.0008827653,0.001685547,0.02609636,0.003351689,0.001491835,0.4161351,0.4638299,0.0003873442],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"commentary","genre_scores_codex":[0.006478129,0.08460758,0.01165479,0.8437529,0.01111512,0.0003037924,0.00009247159,0.0001861423,0.04180898],"genre_scores_gemma":[0.507943,0.1016249,0.05558021,0.3053511,0.01390276,0.0009337173,0.0001506365,0.0002929766,0.01422063],"genre_candidate":"commentary","genre_consensus":"commentary","teacher_disagreement_score":0.8889675,"threshold_uncertainty_score":0.5872029,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2139159602","doi":"10.1002/ev.92","title":"PIE à la Mode: Mainstreaming Evaluation and Accountability in Each Program in Every County of a Statewide School Readiness Initiative","year":2003,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":6,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of British Columbia","funders":"","keywords":"Accountability; Mandate; Mainstreaming; Legislature; Program evaluation; Plan (archaeology); Mainstream; Benchmarking; Quality (philosophy); Public administration; Political science; Public relations; Medical education; Psychology; Business; Pedagogy; Special education; Medicine","authors":[{"name":"Abraham Wandersman","is_ca":false},{"name":"Paul Flaspohler","is_ca":false},{"name":"April Ace","is_ca":false},{"name":"Laurie Ford","is_ca":true},{"name":"Pamela Imm","is_ca":false},{"name":"Matthew Chinman","is_ca":false},{"name":"Jeffrey Sheldon","is_ca":false},{"name":"Arlene Bowers Andrews","is_ca":false},{"name":"Cindy A. Crusto","is_ca":false},{"name":"Joy S. Kaufman","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2119641125322008,"gpt":0.5393907107995318,"spread":0.327426598267331,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.03113796,0.0002239515,0.000229503,0.001798123,0.003319719,0.006922999,0.001059531,0.0008154989,0.004190096],"category_scores_gemma":[0.02944086,0.000358882,0.000271639,0.001473563,0.001813096,0.004899811,0.005998421,0.003086574,0.0003578113],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.007069761,"about_ca_system_score_gemma":0.02101115,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01049043,"about_ca_topic_score_gemma":0.03572825,"domain_scores_codex":[0.9866915,0.008635495,0.0004469853,0.0004961949,0.002744606,0.0009853212],"domain_scores_gemma":[0.9852087,0.00784455,0.0005445246,0.0008814234,0.003723733,0.001797046],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"observational","study_design_scores_codex":[0.0001857648,0.00117985,0.0212877,0.0002587174,0.00002497241,0.0001232495,0.01171548,0.003238954,0.002310762,0.1224087,0.04477871,0.7924871],"study_design_scores_gemma":[0.0005224952,0.003396914,0.1662688,0.002226332,0.0001646532,0.0004969281,0.07111351,0.04354369,0.02020607,0.08959783,0.6021252,0.000337573],"study_design_candidate":"observational","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.4171363,0.002186124,0.2319594,0.06332153,0.0006521919,0.004629376,0.0005350341,0.002099107,0.277481],"genre_scores_gemma":[0.7874777,0.001078818,0.1812191,0.004048491,0.0001169615,0.00186352,0.0003974837,0.0001190358,0.02367895],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.03113796,"threshold_uncertainty_score":0.1646752,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1541185587","doi":"10.1002/ev.20112","title":"Credentialed Evaluator Designation Program, the Canadian Experience","year":2015,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Value (mathematics); Professional development; Service (business); Administration (probate law); Political science; Public relations; Management; Engineering ethics; Sociology; Pedagogy; Computer science; Engineering; Law; Business; Marketing; Economics","authors":[{"name":"Keiko Kuji‐Shikatani","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4863368872817085,"gpt":0.5792195495990002,"spread":0.09288266231729175,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.03725708,0.00030941,0.000321254,0.001955503,0.0201888,0.007376547,0.002619594,0.002152307,0.008462339],"category_scores_gemma":[0.04595565,0.0004769019,0.0003194904,0.004317825,0.01034964,0.003870797,0.005922341,0.004094499,0.0005946788],"about_ca_system_candidate":true,"about_ca_system_consensus":true,"about_ca_system_score_codex":0.1481593,"about_ca_system_score_gemma":0.3722418,"about_ca_topic_candidate":true,"about_ca_topic_consensus":true,"about_ca_topic_score_codex":0.964298,"about_ca_topic_score_gemma":0.9854418,"domain_scores_codex":[0.9746241,0.009173535,0.00102672,0.001293943,0.009312254,0.004569395],"domain_scores_gemma":[0.9292789,0.01196008,0.001501552,0.002013017,0.03223557,0.02301086],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.0002249375,0.0007695442,0.02185222,0.000551784,0.00001786672,0.001156062,0.1360699,0.001228611,0.0008138543,0.2445859,0.2926295,0.3000998],"study_design_scores_gemma":[0.00006953974,0.0001480793,0.01604529,0.000441889,0.000012693,0.0003639553,0.04568359,0.0007466643,0.0006807318,0.005745899,0.9299415,0.0001202269],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.1941952,0.01460163,0.01663042,0.3933213,0.004395438,0.001121215,0.0007624858,0.0003938825,0.3745784],"genre_scores_gemma":[0.8433982,0.009041484,0.01698564,0.01754273,0.0002040386,0.0003318883,0.0002998012,0.000198369,0.1119979],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.9627429,"threshold_uncertainty_score":0.9880145,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1505207325","doi":"10.1002/ev.20084","title":"Credible Judgment: Combining Truth, Beauty, and Justice","year":2014,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":5,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Université du Québec à Montréal","funders":"","keywords":"Beauty; Argumentation theory; Economic Justice; Process (computing); Variety (cybernetics); Psychology; Social psychology; Public relations; Computer science; Law; Epistemology; Political science","authors":[{"name":"Marthe Hurteau","is_ca":true},{"name":"David D. Williams","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.260603245733291,"gpt":0.5132600924299058,"spread":0.2526568466966148,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1189736,0.0007265395,0.00130439,0.008952597,0.007388193,0.02302246,0.002538121,0.004100107,0.004495417],"category_scores_gemma":[0.2210006,0.0007243973,0.0008584923,0.004382047,0.0783371,0.02508212,0.01331009,0.005729364,0.0003369363],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01092107,"about_ca_system_score_gemma":0.01215063,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004262019,"about_ca_topic_score_gemma":0.003921961,"domain_scores_codex":[0.7759113,0.1790424,0.00599333,0.005239694,0.02934019,0.00447292],"domain_scores_gemma":[0.6437319,0.2928107,0.0207436,0.01406476,0.02432401,0.004325105],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.00009596281,0.00009165511,0.004904528,0.0004253691,0.000061307,0.0002552707,0.04351084,0.001406487,0.0002490439,0.9062387,0.002382814,0.0403781],"study_design_scores_gemma":[0.00003025347,0.00009619659,0.003013087,0.001081767,0.00005169944,0.0002008428,0.02414166,0.003951899,0.0007675989,0.9486142,0.0179804,0.00007034934],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"other","genre_gemma":"methods","genre_scores_codex":[0.193831,0.006128235,0.2472817,0.05574869,0.000607064,0.0008945283,0.0001147745,0.0002170994,0.495177],"genre_scores_gemma":[0.9811183,0.0004208016,0.0165526,0.0007280466,0.00008407191,0.0000983205,0.00001624026,0.00002801632,0.0009535404],"genre_candidate":"methods","genre_consensus":null,"teacher_disagreement_score":0.8810264,"threshold_uncertainty_score":0.6291998,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2512463988","doi":"10.1002/ev.20199","title":"Building Evaluation Capacity Through CLIPs: Communities of Learning, Inquiry, and Practice","year":2016,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"University of British Columbia","funders":"","keywords":"CLIPS; Context (archaeology); Work (physics); Community of practice; Medical education; Professional development; Higher education; Pedagogy; Knowledge management; Computer science; Sociology; Public relations; Psychology; Political science; Medicine; Engineering","authors":[{"name":"Beverly L. Parsons","is_ca":false},{"name":"Chris Y. Lovato","is_ca":true},{"name":"Kylie Hutchinson","is_ca":false},{"name":"Derek Wilson","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5953867694134086,"gpt":0.5760182223031797,"spread":0.01936854711022895,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.08931211,0.000736178,0.0007385316,0.005123136,0.006891178,0.01451728,0.003440233,0.002640431,0.007440667],"category_scores_gemma":[0.104299,0.0006946943,0.0008111704,0.002216123,0.02695804,0.01635122,0.01934928,0.003219095,0.0008213714],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.009246946,"about_ca_system_score_gemma":0.02342612,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004063401,"about_ca_topic_score_gemma":0.004599255,"domain_scores_codex":[0.8850632,0.09899224,0.001802306,0.002573375,0.008637002,0.002931939],"domain_scores_gemma":[0.8364749,0.1236131,0.004166293,0.01469523,0.01308276,0.007967575],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"not_applicable","study_design_scores_codex":[0.0002058944,0.001171238,0.009489201,0.0008384563,0.00008535632,0.0004120213,0.08412867,0.006040373,0.00124859,0.5821263,0.01482787,0.2994261],"study_design_scores_gemma":[0.000324463,0.0009390807,0.004748099,0.002484974,0.00008651006,0.0004525713,0.1067404,0.03103945,0.003636634,0.6922767,0.1570551,0.0002160155],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.2011511,0.00148648,0.4665452,0.0511254,0.000438787,0.004961348,0.0001340554,0.001271195,0.2728865],"genre_scores_gemma":[0.8337615,0.0004760404,0.157529,0.001170332,0.00007376228,0.001693991,0.0000744069,0.0001069273,0.00511397],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.08931211,"threshold_uncertainty_score":0.4723332,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3204738611","doi":"10.1002/ev.20471","title":"Contribution analysis: A promising method for assessing advocacy's impact","year":2021,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Impact","funders":"","keywords":"Context (archaeology); Policy analysis; Outcome (game theory); Political science; Public relations; Management science; Public administration; Economics","authors":[{"name":"Robin Kane","is_ca":false},{"name":"Carlisle Levine","is_ca":false},{"name":"Carlyn E. Orians","is_ca":true},{"name":"Claire Reinelt","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2401609389579082,"gpt":0.6095421407491777,"spread":0.3693812017912694,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1586432,0.002599075,0.002062961,0.02199066,0.003402268,0.008886466,0.003380887,0.002248912,0.01707628],"category_scores_gemma":[0.3858356,0.001012165,0.003355291,0.0140815,0.006767841,0.01119844,0.007023143,0.003968138,0.00183706],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004120034,"about_ca_system_score_gemma":0.00595845,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003145752,"about_ca_topic_score_gemma":0.002746035,"domain_scores_codex":[0.8005486,0.155636,0.007905859,0.006075598,0.0285726,0.001261403],"domain_scores_gemma":[0.427568,0.5019444,0.01801471,0.02329269,0.02751149,0.001668761],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.002060363,0.0007273215,0.03767585,0.004621612,0.001673442,0.0002639436,0.01732364,0.007871129,0.003836127,0.3095118,0.01603694,0.5983978],"study_design_scores_gemma":[0.0009098262,0.003015792,0.03956637,0.004547396,0.001589526,0.0009388039,0.01849456,0.1118024,0.01836095,0.7010918,0.09889362,0.0007889001],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.03363288,0.001031509,0.8915009,0.002741284,0.0005780857,0.004795033,0.001891789,0.002154769,0.06167375],"genre_scores_gemma":[0.3553033,0.0005197154,0.6301973,0.0005357421,0.0001742131,0.008461528,0.0005099965,0.0004521463,0.003846021],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.1586432,"threshold_uncertainty_score":0.8389953,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4242601229","doi":"10.1002/ev.20107","title":"Editors' Notes","year":2015,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Accreditation; Charter; Certification; Context (archaeology); Credentialing; Conversation; Public relations; Professional association; Scholarship; Task (project management); Computer science; Political science; Engineering ethics; Management; Psychology; Law; History","authors":[{"name":"James W. Altschuld","is_ca":false},{"name":"Molly Engle","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4199570485042934,"gpt":0.5640301890940794,"spread":0.144073140589786,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007956757,0.00140042,0.001525543,0.00468767,0.003588811,0.008597897,0.005706202,0.005931261,0.4067391],"category_scores_gemma":[0.06994286,0.0007954877,0.001515003,0.004095411,0.00140072,0.005555568,0.003515859,0.006889331,0.3229369],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003689009,"about_ca_system_score_gemma":0.009225461,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003783897,"about_ca_topic_score_gemma":0.006605145,"domain_scores_codex":[0.9904056,0.0008678977,0.001090499,0.001381676,0.005446929,0.0008073403],"domain_scores_gemma":[0.9544216,0.007843032,0.002457337,0.003882087,0.02712804,0.00426794],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000007503414,0.000006384879,0.00001472699,0.00008431599,0.000001002445,0.00002176397,0.00002565019,0.000009565534,0.00003676168,0.0009063911,0.9874381,0.01144773],"study_design_scores_gemma":[0.000004531754,0.000003147195,0.00003969228,0.000109639,0.00000144154,0.00002410546,0.00002562247,0.000008948015,0.00004258287,0.0003372842,0.9993993,0.000003642756],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.0002297917,0.008953522,0.001555132,0.1101802,0.7011521,0.0004115448,0.003940918,0.00103781,0.1725389],"genre_scores_gemma":[0.002813769,0.009594838,0.003288877,0.07299873,0.1504685,0.0006257473,0.003380296,0.001134184,0.7556951],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.4067391,"threshold_uncertainty_score":0.8462147,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4403250772","doi":"10.1002/ev.20615","title":"Using contribution analysis to assess evaluation capacity outcomes in community organizations","year":2024,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":4,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Université de Montréal; University of Ottawa","funders":"","keywords":"Evaluation methods; Capacity building; Program evaluation; Process management; Business; Political science; Public administration; Economic growth; Economics; Engineering","authors":[{"name":"Isabelle Bourgeois","is_ca":true},{"name":"David Buetti","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.538010031454944,"gpt":0.5850516251312451,"spread":0.04704159367630112,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1331756,0.0008381415,0.0009471989,0.01574988,0.002683747,0.00612035,0.001894073,0.0007618454,0.004338236],"category_scores_gemma":[0.2660719,0.0004206979,0.001299594,0.009617411,0.002748136,0.005954213,0.007400456,0.001352476,0.0004482051],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.008764398,"about_ca_system_score_gemma":0.009895675,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005603034,"about_ca_topic_score_gemma":0.006007421,"domain_scores_codex":[0.8738123,0.1008664,0.005151742,0.002397954,0.01522503,0.002546562],"domain_scores_gemma":[0.602225,0.2909961,0.03155084,0.01260599,0.05851694,0.004105104],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.002267945,0.002885784,0.3183703,0.00551858,0.001084576,0.0002499357,0.04496251,0.01251718,0.001767238,0.04773562,0.006859386,0.555781],"study_design_scores_gemma":[0.0009297307,0.01025687,0.6057053,0.007188572,0.002010221,0.0003805595,0.1328061,0.07195381,0.02307383,0.08735606,0.05762709,0.0007119321],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.8480289,0.001187414,0.06908111,0.001886221,0.0001464547,0.008467887,0.001562017,0.0003906419,0.06924925],"genre_scores_gemma":[0.9714293,0.0002005192,0.0225943,0.00008135007,0.00001604232,0.004445211,0.00035037,0.00002873676,0.0008540196],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.1331756,"threshold_uncertainty_score":0.7043085,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2017680079","doi":"10.1002/ev.344","title":"Strategy evaluation: Experience at the International Development Research Centre","year":2010,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"International Development Research Centre","funders":"","keywords":"Corporation; Program evaluation; Strategic planning; Modalities; International development; Sociology; Management; Business; Public relations; Political science; Economics; Public administration; Social science","authors":[{"name":"Tricia Wind","is_ca":true},{"name":"Fred Carden","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5427282247833871,"gpt":0.6215806160734704,"spread":0.07885239129008326,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["metaresearch"],"consensus_categories":[],"category_scores_codex":[0.1244159,0.000796193,0.001010135,0.003101812,0.0114931,0.01753287,0.003466094,0.003569349,0.00911707],"category_scores_gemma":[0.06483799,0.0006409856,0.0006686727,0.006110033,0.007687597,0.005926758,0.009161646,0.009480849,0.002093116],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02863347,"about_ca_system_score_gemma":0.04759675,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.03693017,"about_ca_topic_score_gemma":0.03172826,"domain_scores_codex":[0.8678347,0.09384147,0.0033221,0.003379717,0.02103828,0.01058365],"domain_scores_gemma":[0.9057304,0.03804176,0.002352838,0.005983589,0.02872035,0.01917114],"domain_codex":null,"domain_gemma":"evaluation","domain_candidate":"evaluation","domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.00072254,0.005232947,0.01058699,0.001163367,0.00008096587,0.001962651,0.2383221,0.002845929,0.001688871,0.1271673,0.1356118,0.4746145],"study_design_scores_gemma":[0.0001484837,0.001335245,0.005594723,0.001117816,0.0000442267,0.0005222415,0.08343735,0.001776641,0.003321381,0.004864594,0.8976632,0.0001741485],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.2523656,0.02227874,0.02286301,0.06835711,0.001559808,0.001796767,0.0004240707,0.0007904526,0.6295645],"genre_scores_gemma":[0.8869728,0.01156714,0.02390143,0.00710009,0.0003892267,0.0008856975,0.0004801019,0.0006744718,0.06802896],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.8755841,"threshold_uncertainty_score":0.657982,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W3011218257","doi":"10.1002/ev.20394","title":"Advancing the Measurement of Collective Community Capacity and the Evaluation of Community Capacity‐Building Models","year":2020,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Community Health and Development","field":"Health Professions","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Capacity building; Work (physics); Collective efficacy; Community resilience; Adaptation (eye); Public relations; Psychological resilience; Political science; Sociology; Psychology; Computer science; Engineering; Social psychology; Social science","authors":[{"name":"Margaret B. Hargreaves","is_ca":false},{"name":"Brandon Coffee‐Borden","is_ca":false},{"name":"Natalya Verbitsky‐Savitz","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.4752172519940911,"gpt":0.4749833008706206,"spread":0.0002339511234704661,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1376068,0.0009717617,0.001136704,0.005243787,0.002923781,0.009706336,0.002732884,0.001492267,0.005868495],"category_scores_gemma":[0.2800628,0.000583032,0.001259025,0.005236973,0.008505627,0.01192675,0.01000681,0.003946962,0.0005219663],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.01042149,"about_ca_system_score_gemma":0.01961646,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.01951471,"about_ca_topic_score_gemma":0.02479189,"domain_scores_codex":[0.8376846,0.1372168,0.003044364,0.003036474,0.01728367,0.001734153],"domain_scores_gemma":[0.6190163,0.2841763,0.01909508,0.02769468,0.04434738,0.005670193],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.0002100839,0.0008424,0.1181603,0.002640906,0.0006469039,0.00006324764,0.02216185,0.01434408,0.0004534848,0.3639075,0.01567079,0.4608985],"study_design_scores_gemma":[0.0002335856,0.001427105,0.1279257,0.01618261,0.000471135,0.000146841,0.0748893,0.08866965,0.003898321,0.5966482,0.08895989,0.0005476354],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.3193657,0.009765209,0.4256566,0.03960966,0.001477014,0.004532781,0.001462687,0.0007376433,0.1973926],"genre_scores_gemma":[0.8738543,0.00214547,0.1185338,0.000839557,0.0001021408,0.002599682,0.0002756779,0.0000714631,0.001577787],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.1376068,"threshold_uncertainty_score":0.7277431,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4378223570","doi":"10.1002/ev.20540","title":"What we can learn from the international program for development evaluation training (IPDET)","year":2023,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"Program evaluation; Training (meteorology); Monitoring and evaluation; Diversity (politics); Medical education; Corporate governance; Political science; Psychology; Business; Medicine; Geography; Public administration; Finance","authors":[{"name":"Linda Morra Imas","is_ca":false},{"name":"Ray C. Rist","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.5304740430548962,"gpt":0.569693978952799,"spread":0.03921993589790285,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1413985,0.00103758,0.001367711,0.003699779,0.004764757,0.02289083,0.004002737,0.007470985,0.01708294],"category_scores_gemma":[0.222598,0.0006817569,0.001587196,0.004299414,0.008882074,0.03056965,0.0125243,0.01624361,0.006731378],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.009910089,"about_ca_system_score_gemma":0.03367612,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004560879,"about_ca_topic_score_gemma":0.007751058,"domain_scores_codex":[0.8960354,0.07798661,0.004390252,0.003171425,0.01313428,0.005282092],"domain_scores_gemma":[0.7560911,0.1307371,0.007023345,0.01951616,0.05611706,0.03051524],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00005264946,0.0003061058,0.002182538,0.002567788,0.00004831966,0.0002015537,0.006515589,0.0003167312,0.0001537326,0.06889334,0.5485016,0.3702601],"study_design_scores_gemma":[0.00004515281,0.0001638448,0.002310506,0.01272282,0.00003124126,0.0002229681,0.007388561,0.0003002708,0.0003074848,0.05119148,0.9252431,0.00007255779],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.002257787,0.04262037,0.01642799,0.8713218,0.01397382,0.0003458132,0.0003891783,0.0005690255,0.05209426],"genre_scores_gemma":[0.140936,0.1689,0.1686292,0.446188,0.0221651,0.003876936,0.00232141,0.001406552,0.04557685],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.1413985,"threshold_uncertainty_score":0.7477956,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W2626410734","doi":"10.1002/ev.20240","title":"Conceptual Indicators Framework for Strengthening the Chinese Health System","year":2017,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Healthcare Systems and Reforms","field":"Economics, Econometrics and Finance","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"","keywords":"China; Government (linguistics); Plan (archaeology); Conceptual framework; Process (computing); Political science; Equity (law); Monitoring and evaluation; Process management; Public relations; Public administration; Economic growth; Business; Sociology; Computer science; Economics; Geography","authors":[{"name":"Yue Xiao","is_ca":false},{"name":"Kun Zhao","is_ca":false},{"name":"Sanjeev Sridharan","is_ca":false},{"name":"Xiaohong Cao","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.08227829248233864,"gpt":0.3703221274091979,"spread":0.2880438349268593,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02617886,0.001423247,0.0007762956,0.0156336,0.003362134,0.009469719,0.003331364,0.00202973,0.008259454],"category_scores_gemma":[0.01780089,0.0004507431,0.001077461,0.01484485,0.01430298,0.01244404,0.007386644,0.002918381,0.0004625114],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.03317743,"about_ca_system_score_gemma":0.03300628,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.04810024,"about_ca_topic_score_gemma":0.02278261,"domain_scores_codex":[0.9857526,0.008632015,0.000945089,0.001017265,0.002628854,0.001024062],"domain_scores_gemma":[0.9908183,0.002854735,0.001210396,0.000562613,0.003615306,0.0009386886],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"theoretical_or_conceptual","study_design_scores_codex":[0.000004363117,0.00001069932,0.0009051201,0.0000679259,0.000007932611,0.00002398612,0.0007482956,0.002133134,0.00002705923,0.9908275,0.001334519,0.003909415],"study_design_scores_gemma":[0.0000578229,0.0000861784,0.006263881,0.0006850866,0.00009423145,0.00007048852,0.006646208,0.0422355,0.000359012,0.8622165,0.08120938,0.00007569976],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":"theoretical_or_conceptual","genre_codex":"methods","genre_gemma":"methods","genre_scores_codex":[0.06576832,0.005181246,0.5430288,0.07253626,0.0009342831,0.003606776,0.002742717,0.0006491116,0.3055525],"genre_scores_gemma":[0.8500119,0.00206339,0.1359244,0.001210169,0.0002834756,0.002959594,0.0009908257,0.00006581612,0.006490342],"genre_candidate":"methods","genre_consensus":"methods","teacher_disagreement_score":0.04810024,"threshold_uncertainty_score":0.2407203,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4249389693","doi":"10.1002/ev.20282","title":"Issue Information","year":2018,"lang":"en","type":"paratext","venue":"New Directions for Evaluation","topic":"Diverse Scientific and Economic Studies","field":"Economics, Econometrics and Finance","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":false,"ca_fund":true,"ca_venue":false,"about_ca":true},"ca_institutions":"","funders":"University of California, Los Angeles; National Science Foundation; Claremont Graduate University; Duquesne University; Weizmann Institute of Science; Michigan State University; Sierra Health Foundation; Northwestern University; Gordon and Betty Moore Foundation; University of Connecticut; College of Engineering, Michigan State University; École nationale d'administration publique; Syracuse University; University of Southern California; John D. and Catherine T. MacArthur Foundation","keywords":"Computer science; Citation; World Wide Web; Information retrieval; Library science","authors":[{"name":"Leslie Fierro","is_ca":false},{"name":"Todd Franke","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.0711548467400742,"gpt":0.2947114174565125,"spread":0.2235565707164383,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.001418259,0.001593009,0.001682266,0.003759841,0.001889416,0.009196392,0.002046429,0.002387124,0.9296991],"category_scores_gemma":[0.009938553,0.0005818434,0.000942528,0.004508239,0.0005631555,0.005467356,0.002271472,0.00227435,0.8993487],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.001717687,"about_ca_system_score_gemma":0.002830875,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.003166578,"about_ca_topic_score_gemma":0.005158059,"domain_scores_codex":[0.9986148,0.0001543578,0.0001146467,0.00021595,0.0007363455,0.0001639513],"domain_scores_gemma":[0.9932482,0.0008472952,0.0002880054,0.000680989,0.003300357,0.001635091],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001121291,0.000009218323,0.00002443147,0.00009423848,8.649879e-7,0.000006395224,0.000004886648,0.00001383545,0.00003400375,0.0005426423,0.9867054,0.01255274],"study_design_scores_gemma":[0.000008586674,0.000007635449,0.0001353385,0.00009273872,9.849172e-7,0.00001377723,0.00001690212,0.00002084939,0.00003663393,0.000466785,0.9991967,0.000003116725],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"other","genre_gemma":"other","genre_scores_codex":[0.0001924309,0.001134926,0.0008048309,0.004303252,0.01634563,0.0005194983,0.04407755,0.002483122,0.9301388],"genre_scores_gemma":[0.0008803232,0.0009915077,0.0003962023,0.0018458,0.002799865,0.0001674598,0.01899484,0.0009745884,0.9729495],"genre_candidate":"other","genre_consensus":"other","teacher_disagreement_score":0.9296991,"threshold_uncertainty_score":0.1002757,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4378223319","doi":"10.1002/ev.20539","title":"Meeting the challenges of educating internal evaluators","year":2023,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":3,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"Institute for Clinical Evaluative Sciences; Queen's University","funders":"","keywords":"Front line; Public relations; Geopolitics; Coronavirus disease 2019 (COVID-19); Pandemic; Program evaluation; Political science; Business; Medicine; Public administration","authors":[{"name":"Chi Yan Lam","is_ca":true},{"name":"Keiko Kuji‐Shikatani","is_ca":true},{"name":"Arnold J. Love","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3888663697506948,"gpt":0.5659171151344419,"spread":0.1770507453837471,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1775569,0.0008783411,0.001064681,0.002188819,0.008922302,0.02367754,0.002829886,0.005399364,0.007904977],"category_scores_gemma":[0.1826908,0.0007156722,0.0007141907,0.0009225427,0.009942832,0.01261199,0.01500298,0.01348913,0.002817972],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.008452885,"about_ca_system_score_gemma":0.02942002,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00271299,"about_ca_topic_score_gemma":0.005311442,"domain_scores_codex":[0.8542297,0.1151098,0.003889264,0.003615327,0.01607137,0.007084548],"domain_scores_gemma":[0.598514,0.244903,0.01793634,0.01517175,0.08644867,0.03702622],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000210006,0.001474737,0.01817996,0.001724253,0.0001480799,0.0004887077,0.07187869,0.00216578,0.00172426,0.07008024,0.2616135,0.5703118],"study_design_scores_gemma":[0.0002894585,0.0008883237,0.01341982,0.008354655,0.0001039612,0.0009286759,0.1598318,0.006124446,0.004542664,0.1417493,0.6634736,0.0002933007],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"commentary","genre_gemma":"empirical","genre_scores_codex":[0.08380216,0.01468474,0.07172196,0.7635204,0.004373027,0.000584449,0.00006500858,0.0007913565,0.06045701],"genre_scores_gemma":[0.7817616,0.01076962,0.08398671,0.09212858,0.003077033,0.001050103,0.0001225845,0.0004581846,0.02664566],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.1775569,"threshold_uncertainty_score":0.9390219,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4307053258","doi":"10.1002/ev.20514","title":"Identity as a compass when navigating uncharted equitable spaces: Our queer evaluation practices","year":2022,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Saskatchewan; Saskatchewan Health","funders":"","keywords":"Queer; Dialogic; Situated; Sociology; Normative; Identity (music); Epistemology; Norm (philosophy); Social psychology; Psychology; Computer science; Aesthetics; Gender studies; Pedagogy","authors":[{"name":"Andrew Hartman","is_ca":true},{"name":"Brian Hoessler","is_ca":true},{"name":"Vincent Tom","is_ca":true},{"name":"Carolyn Camman","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3794672207651364,"gpt":0.5847148136120114,"spread":0.2052475928468749,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.1227283,0.0005840787,0.000990733,0.003203225,0.03451759,0.03040018,0.002806788,0.005050606,0.006036903],"category_scores_gemma":[0.09185512,0.0006062889,0.00070257,0.002038312,0.09518934,0.02160408,0.02098529,0.009117198,0.0008051072],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.02690808,"about_ca_system_score_gemma":0.020761,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.03077497,"about_ca_topic_score_gemma":0.03665822,"domain_scores_codex":[0.7739963,0.2035161,0.002787008,0.004466782,0.008900447,0.006333375],"domain_scores_gemma":[0.9092591,0.06736936,0.004076526,0.005910384,0.008814386,0.004570271],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.00003401524,0.00004966318,0.001069091,0.00007399299,0.00001027287,0.0003799134,0.569746,0.0003108393,0.000383382,0.4056766,0.00392617,0.01833996],"study_design_scores_gemma":[0.0000201081,0.00008138898,0.001028784,0.0006955076,0.00002783949,0.0003548729,0.6105668,0.001300701,0.001819246,0.1724401,0.2115526,0.0001121709],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.2940722,0.006818529,0.144814,0.1917138,0.0009188841,0.0005824048,0.00005619932,0.0003736409,0.3606505],"genre_scores_gemma":[0.9713679,0.0006625893,0.007781298,0.003478233,0.0000564849,0.0001379423,0.000008960072,0.0001093694,0.01639719],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.1227283,"threshold_uncertainty_score":0.649057,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4236772578","doi":"10.1002/ev.239","title":"Editor's notes","year":2007,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Ottawa","funders":"","keywords":"Citation; Computer science; Library science; Sociology","authors":[{"name":"J. Bradley Cousins","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.2516694502734112,"gpt":0.5602664481446863,"spread":0.3085969978712751,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":["insufficient_payload"],"consensus_categories":[],"category_scores_codex":[0.007299764,0.001334438,0.00173987,0.00257294,0.002503113,0.004756823,0.004471782,0.01216091,0.09518548],"category_scores_gemma":[0.05275503,0.0007844376,0.001817204,0.001609247,0.00130531,0.002491295,0.002005491,0.01109201,0.0516341],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.002865157,"about_ca_system_score_gemma":0.005060075,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.004730962,"about_ca_topic_score_gemma":0.01182339,"domain_scores_codex":[0.9934375,0.0008682173,0.0006638069,0.0007997226,0.003605376,0.0006253698],"domain_scores_gemma":[0.9732171,0.006282153,0.002170503,0.001823579,0.01374008,0.002766578],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.00001247262,0.000005658274,0.0000203563,0.00004951006,0.000003276451,0.00006340627,0.000007546873,0.000008066747,0.00003533172,0.0002663721,0.9958527,0.00367534],"study_design_scores_gemma":[0.00002466971,0.000009483757,0.0002033189,0.0001408327,0.00001398624,0.000135481,0.00003411492,0.0000461028,0.0001328876,0.0004853852,0.9987624,0.00001123045],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.00009934859,0.00301481,0.0002168947,0.151602,0.8315672,0.00005961765,0.0003002913,0.00008976536,0.01305009],"genre_scores_gemma":[0.00284713,0.004210832,0.001152536,0.3117812,0.46873,0.0001603275,0.000337114,0.0001603916,0.2106205],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.9048145,"threshold_uncertainty_score":0.3184272,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4307053074","doi":"10.1002/ev.20511","title":"Empowering the search for pleasure, health and well‐being outside heteronormative definitions: The role of evaluation in shaping structurally sensitive programming for 2SGBTQ+ men who <i>Party and Play</i> in Ontario","year":2022,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"HIV, Drug Use, Sexual Risk","field":"Medicine","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Casey House","funders":"","keywords":"Phenomenon; Agency (philosophy); Transformative learning; Pleasure; Sociology; Public health; Public relations; Health care; Qualitative research; Gender studies; Social psychology; Psychology; Medicine; Nursing; Political science; Epistemology; Social science","authors":[{"name":"Yasser Ismail","is_ca":true},{"name":"Dane Griffiths","is_ca":false},{"name":"Jordan Bond‐Gorr","is_ca":false}],"retraction":null,"screen_n_in":null,"score":{"opus":0.104289379332773,"gpt":0.3925731318662367,"spread":0.2882837525334636,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.01941075,0.0003985815,0.000408364,0.0006940564,0.01696042,0.009319324,0.001938682,0.001204374,0.004571783],"category_scores_gemma":[0.01491612,0.0003350344,0.0004265523,0.0006286351,0.0386342,0.002545618,0.01182676,0.002785288,0.0002204373],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.06756195,"about_ca_system_score_gemma":0.130473,"about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.4781729,"about_ca_topic_score_gemma":0.7452886,"domain_scores_codex":[0.9784475,0.01579825,0.0002639778,0.0005879525,0.002075133,0.002827189],"domain_scores_gemma":[0.9891299,0.00543409,0.0005294266,0.0005074881,0.001693142,0.00270599],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"qualitative","study_design_scores_codex":[0.0001176709,0.0003394546,0.01030859,0.000401082,0.00002164227,0.0005011174,0.7444818,0.0008961026,0.001992709,0.1499403,0.008107147,0.08289243],"study_design_scores_gemma":[0.00009053445,0.000309437,0.02171131,0.001362472,0.00005641309,0.0001636669,0.626711,0.001691379,0.002435984,0.06215195,0.2832126,0.0001032576],"study_design_candidate":"qualitative","study_design_consensus":"qualitative","genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.586773,0.001707055,0.02696964,0.06612931,0.000269723,0.001736003,0.00009795317,0.0001388068,0.3161785],"genre_scores_gemma":[0.9843032,0.0004658522,0.005080451,0.001287323,0.00001319274,0.0003050004,0.00001520332,0.00003725526,0.008492541],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.5218271,"threshold_uncertainty_score":0.9507793,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W1978303279","doi":"10.1002/ev.382","title":"Integrating a new evaluation unit with an old institution: See no evil; hear no evil; speak no evil","year":2011,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Guelph","funders":"","keywords":"Unit (ring theory); CLARITY; Institution; Evaluation methods; Association (psychology); Psychology; Sociology; Computer science; Mathematics education; Social science; Engineering; Psychotherapist","authors":[{"name":"Claire E. Baxter","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.3909805053050648,"gpt":0.5034779046174458,"spread":0.112497399312381,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.05355232,0.0004400616,0.000763208,0.001161552,0.005182517,0.01292186,0.002172179,0.003794493,0.01159493],"category_scores_gemma":[0.1000628,0.000387555,0.0004521878,0.0005923864,0.006101722,0.007667326,0.01047828,0.006419727,0.003574162],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.004407026,"about_ca_system_score_gemma":0.01248381,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.00146666,"about_ca_topic_score_gemma":0.00450404,"domain_scores_codex":[0.928192,0.04689859,0.004059466,0.001315328,0.01694284,0.002591753],"domain_scores_gemma":[0.9073618,0.0389775,0.008430242,0.005974495,0.01841391,0.02084206],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"design_other","study_design_gemma":"qualitative","study_design_scores_codex":[0.00102616,0.003586912,0.02937808,0.001249094,0.00009222425,0.001345037,0.023826,0.00206715,0.004903742,0.1458927,0.1621587,0.6244743],"study_design_scores_gemma":[0.0006942237,0.003734293,0.04828169,0.00412796,0.0002530359,0.002044874,0.06518012,0.02023985,0.02510989,0.08010741,0.7491906,0.001036082],"study_design_candidate":"qualitative","study_design_consensus":null,"genre_codex":"other","genre_gemma":"empirical","genre_scores_codex":[0.2279489,0.002245094,0.1431567,0.1937663,0.009341985,0.003202911,0.0001123331,0.003326198,0.4168997],"genre_scores_gemma":[0.8273485,0.0007777683,0.09592788,0.02241347,0.001265115,0.001782231,0.00006285364,0.0005186198,0.04990361],"genre_candidate":"empirical","genre_consensus":null,"teacher_disagreement_score":0.05355232,"threshold_uncertainty_score":0.2832151,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4378223534","doi":"10.1002/ev.20536","title":"Learning by linking the Canadian Evaluation Society's student case competition within a graduate evaluation course","year":2023,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Evaluation and Performance Assessment","field":"Decision Sciences","cited_by":2,"is_retracted":false,"has_abstract":true,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":true},"ca_institutions":"Queen's University","funders":"","keywords":"Experiential learning; Dialogic; Pedagogy; Psychology; Nature versus nurture; Situated; Competition (biology); Professional development; Assessment for learning; Sociology; Formative assessment; Computer science","authors":[{"name":"Paisley Worthington","is_ca":true},{"name":"Rebecca Stroud Stasel","is_ca":true},{"name":"Katrina Carbone","is_ca":true},{"name":"Jennifer Hughes","is_ca":true},{"name":"Michelle Searle","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.342939436925835,"gpt":0.5475533572533501,"spread":0.2046139203275151,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.02170863,0.0006018087,0.0004985437,0.002112902,0.01945194,0.01359042,0.002550939,0.002764064,0.009603619],"category_scores_gemma":[0.03130071,0.000341153,0.0005389174,0.001500794,0.007964598,0.002652253,0.009988921,0.004721761,0.001342785],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.04149574,"about_ca_system_score_gemma":0.04598787,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.3172786,"about_ca_topic_score_gemma":0.6664461,"domain_scores_codex":[0.9739604,0.01111359,0.0004891384,0.00169995,0.007248088,0.005488792],"domain_scores_gemma":[0.9720063,0.007489794,0.0009049813,0.001411137,0.008212745,0.009974995],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"qualitative","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000268575,0.001276959,0.02512312,0.0001607722,0.00002649449,0.00405554,0.5868398,0.002017821,0.005327994,0.06632254,0.1365421,0.1720383],"study_design_scores_gemma":[0.00002839736,0.000395936,0.01641287,0.0002391196,0.00001416852,0.0008184853,0.5330135,0.002754007,0.002600942,0.006024853,0.4375003,0.0001974413],"study_design_candidate":"not_applicable","study_design_consensus":null,"genre_codex":"empirical","genre_gemma":"empirical","genre_scores_codex":[0.7353532,0.000318347,0.01124226,0.03011095,0.001257572,0.0009359054,0.0002737737,0.0002675707,0.2202405],"genre_scores_gemma":[0.9365143,0.0001397979,0.006224395,0.001820167,0.0000820241,0.0002240966,0.0001463114,0.0001255455,0.05472329],"genre_candidate":"empirical","genre_consensus":"empirical","teacher_disagreement_score":0.3172786,"threshold_uncertainty_score":0.6308636,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null},{"id":"W4252169279","doi":"10.1002/ev.309","title":"Editors' notes","year":2009,"lang":"en","type":"article","venue":"New Directions for Evaluation","topic":"Community Health and Development","field":"Health Professions","cited_by":1,"is_retracted":false,"has_abstract":false,"routes":{"ca_aff":true,"ca_fund":false,"ca_venue":false,"about_ca":false},"ca_institutions":"University of Calgary","funders":"","keywords":"Citation; Library science; Population; Intervention (counseling); State (computer science); Sociology; Medicine; Computer science; Nursing","authors":[{"name":"Judith M. Ottoson","is_ca":false},{"name":"Penelope Hawe","is_ca":true}],"retraction":null,"screen_n_in":null,"score":{"opus":0.1565093687785537,"gpt":0.5236589003624142,"spread":0.3671495315838605,"validation_status":"score_only:v0-immature-baseline"},"prediction":{"model_version":"metacan-v3-hybrid-931329e0061c","candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.013162,0.001179881,0.00181735,0.002314435,0.003217611,0.006470282,0.006322058,0.02065766,0.08014844],"category_scores_gemma":[0.102911,0.0008690734,0.002214249,0.001742312,0.001870061,0.00380384,0.003082533,0.01885095,0.05411454],"about_ca_system_candidate":false,"about_ca_system_consensus":false,"about_ca_system_score_codex":0.003510769,"about_ca_system_score_gemma":0.007849287,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_topic_score_codex":0.005399612,"about_ca_topic_score_gemma":0.01412082,"domain_scores_codex":[0.9873896,0.001737235,0.00132434,0.001532623,0.00668472,0.001331511],"domain_scores_gemma":[0.9436357,0.01591905,0.004394293,0.003249107,0.02689587,0.005906004],"domain_codex":null,"domain_gemma":null,"domain_candidate":null,"domain_consensus":null,"study_design_codex":"not_applicable","study_design_gemma":"not_applicable","study_design_scores_codex":[0.000009081058,0.00000536287,0.00002946037,0.00004393388,0.000002411773,0.000059808,0.00001829761,0.000006814546,0.00002429624,0.0003704508,0.9951813,0.004248771],"study_design_scores_gemma":[0.00002036546,0.000008434516,0.0001569195,0.0001832477,0.00001058177,0.0001226567,0.00006165315,0.00002796389,0.0000793127,0.0004944667,0.9988237,0.00001068566],"study_design_candidate":"not_applicable","study_design_consensus":"not_applicable","genre_codex":"editorial","genre_gemma":"editorial","genre_scores_codex":[0.0001259257,0.003850508,0.0003254184,0.3130656,0.6668798,0.00008102233,0.0002615411,0.0001244071,0.01528581],"genre_scores_gemma":[0.001789658,0.002852859,0.001007245,0.5299438,0.3203251,0.000244627,0.0002063682,0.000150091,0.1434802],"genre_candidate":"editorial","genre_consensus":"editorial","teacher_disagreement_score":0.08014844,"threshold_uncertainty_score":0.2681232,"prediction_status":"machine_predicted_unvalidated"},"labels":[],"label_agreement":null}]}