{"meta":{"query_hash":"9f297251ca0a","filters":{"venue":"Educational Evaluation and Policy Analysis"},"cohort_total":14,"direct_labels_cover":0,"predictions_cover":14,"exported":14,"export_cap":100000,"truncated":false,"label_status":"direct model label, unvalidated","prediction_status":"machine_predicted_unvalidated (Codex and Gemma teacher distillation)","score_status":"score_only:v0-immature-baseline","snapshot":{"source":"OpenAlex, pinned release, all 482 partitions","release":"2026-06-24","frame_built":"2026-07-12"},"permalink":"https://metacan.xera.ac/q/9f297251ca0a","api":"https://metacan.xera.ac/api/v1/cohort?venue=Educational+Evaluation+and+Policy+Analysis"},"results":[{"id":"W1977534634","doi":"10.3102/0162373711398125","title":"Who Benefits From Kindergarten? Evidence From the Introduction of State Subsidization","year":2011,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":39,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Socioeconomic status; Subsidy; Immigration; Demographic economics; Revenue; Psychology; State (computer science); Mathematics education; Economic growth; Economics; Demography; Political science; Sociology; Population; Computer science","score_opus":0.08998701385613952,"score_gpt":0.37710806615415127,"score_spread":0.28712105229801177,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1977534634","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.72532874,0.11260956,0.0008172876,0.0708279,0.0007911638,0.00022904646,0.005320992,0.000036855658,0.08403845],"genre_scores_gemma":[0.95897305,0.030315625,0.00027508134,0.0071019013,0.0003768954,0.00007643498,0.0010600785,0.000014192549,0.0018067901],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.98950666,0.0051003164,0.00076267554,0.0010707435,0.0018807938,0.0016787145],"domain_scores_gemma":[0.8920079,0.07274365,0.024123231,0.0027072264,0.005074448,0.0033436497],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011950052,0.0002637073,0.0007717939,0.0018475569,0.0009231494,0.002726009,0.0017413152,0.0017505927,0.015007853],"category_scores_gemma":[0.06080215,0.00043674803,0.001110341,0.003230509,0.0023475282,0.0020580862,0.003267763,0.0033371323,0.00055373757],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0030898154,0.0010366719,0.7960799,0.0035391008,0.0020108556,0.00048037502,0.0029903825,0.0005402138,0.00015531182,0.01835962,0.021410761,0.15030698],"study_design_scores_gemma":[0.00047278032,0.00041135532,0.94770855,0.0063311113,0.0021965129,0.00025167933,0.006007126,0.00027811909,0.00029108018,0.0027247618,0.0332888,0.000038050406],"about_ca_topic_score_codex":0.059954096,"about_ca_topic_score_gemma":0.07820722,"teacher_disagreement_score":0.059954096,"about_ca_system_score_codex":0.002725863,"about_ca_system_score_gemma":0.005168403,"threshold_uncertainty_score":0.11921024},"labels":[],"label_agreement":null},{"id":"W1980838979","doi":"10.3102/01623737027001001","title":"Success for All: First-Year Results From the National Randomized Field Trial","year":2005,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":84,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Randomized experiment; Quarter (Canadian coin); Test (biology); Psychology; Randomized controlled trial; Multilevel model; Control (management); Treatment and control groups; Reading (process); Mathematics education; Statistics; Medicine; Mathematics; Computer science; Political science; Geography","score_opus":0.09148252164579562,"score_gpt":0.47779666160513007,"score_spread":0.38631413995933445,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1980838979","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9411751,0.007828031,0.0076196142,0.0034946038,0.0018400083,0.024276659,0.004145139,0.00049008685,0.0091308495],"genre_scores_gemma":[0.9711731,0.0021140631,0.0061052064,0.0014187769,0.0005451731,0.011777615,0.0017513159,0.00008865775,0.0050260024],"study_design_codex":"randomized_trial","study_design_gemma":"randomized_trial","domain_scores_codex":[0.9221931,0.067499414,0.002829418,0.0024353513,0.0026946343,0.0023481776],"domain_scores_gemma":[0.9149394,0.058013238,0.011385648,0.0064920625,0.0045743478,0.0045952867],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.060382117,0.0016164867,0.0052392446,0.0007287365,0.0010357364,0.0020215819,0.0015330993,0.0040552486,0.0072817737],"category_scores_gemma":[0.08964807,0.0006732393,0.0032873221,0.0007445031,0.0021528618,0.003940534,0.0013136663,0.0029970875,0.00087352627],"study_design_candidate":"randomized_trial","study_design_consensus":"randomized_trial","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.89650625,0.024200823,0.0040518735,0.001713895,0.0040391474,0.00012423628,0.0005262081,0.0013632623,0.0008047683,0.0015381677,0.004533382,0.06059806],"study_design_scores_gemma":[0.59427667,0.36802244,0.017251832,0.00064336666,0.0068190927,0.00017070898,0.0004615657,0.0025749076,0.002227514,0.0018578813,0.0055495985,0.00014441107],"about_ca_topic_score_codex":0.004120848,"about_ca_topic_score_gemma":0.005516115,"teacher_disagreement_score":0.060382117,"about_ca_system_score_codex":0.0022661476,"about_ca_system_score_gemma":0.0052759335,"threshold_uncertainty_score":0.31933492},"labels":[],"label_agreement":null},{"id":"W1987197661","doi":"10.3102/0162373711424206","title":"Differential Effects of Literacy Instruction Time and Homogeneous Ability Grouping in Kindergarten Classrooms","year":2011,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":58,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Psychology; Homogeneous; Literacy; Mathematics education; Developmental psychology; Early literacy; Cognition; Cognitive development; Differential effects; Longitudinal study; Early childhood; Pedagogy; Statistics","score_opus":0.025604482175742054,"score_gpt":0.36474640522516194,"score_spread":0.33914192304941987,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W1987197661","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9991196,0.000113223905,0.00003965783,0.000016908318,0.000003960237,0.000006520229,0.000022332883,0.000008048183,0.00066985615],"genre_scores_gemma":[0.99915576,0.00006269543,0.00010649333,0.0000154093,0.0000037080633,0.0000139504855,0.00006130029,0.000006379924,0.0005743616],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9977163,0.00028804474,0.00018480065,0.00056517887,0.00054715376,0.00069854205],"domain_scores_gemma":[0.992806,0.002683521,0.0021717336,0.00045203255,0.00035094316,0.0015358389],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0008270686,0.0006309271,0.0007294081,0.0013171353,0.00065636163,0.001700113,0.0008760785,0.0005137148,0.0038650963],"category_scores_gemma":[0.006161953,0.00037514593,0.0008209047,0.0008580588,0.0010488594,0.00063867075,0.0024546224,0.00087172614,0.0003123051],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0032686365,0.002294525,0.92859894,0.00018186924,0.00038056463,0.0014887521,0.007230959,0.0004400298,0.025205277,0.00065074937,0.0002773561,0.029982353],"study_design_scores_gemma":[0.000009185836,0.00033784762,0.99733055,0.000016805796,0.000043521424,0.000027502698,0.0013596853,0.000057811234,0.0006159807,0.000056736713,0.00013752452,0.000006961694],"about_ca_topic_score_codex":0.01097906,"about_ca_topic_score_gemma":0.022149071,"teacher_disagreement_score":0.01097906,"about_ca_system_score_codex":0.0012744489,"about_ca_system_score_gemma":0.0012087149,"threshold_uncertainty_score":0.02183032},"labels":[],"label_agreement":null},{"id":"W2124978570","doi":"10.3102/0162373707309073","title":"Early-Grade Retention and Children’s Reading and Math Learning in Elementary Years","year":2007,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":99,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Grade retention; Psychology; Reading (process); Retention rate; Mathematics education; Early childhood; Intervention (counseling); Developmental psychology; Academic achievement; Computer science","score_opus":0.026363589418701528,"score_gpt":0.382660152524002,"score_spread":0.35629656310530045,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2124978570","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99836475,0.00023131767,0.00003174589,0.00016237341,0.0000049342916,0.0000071728505,0.00007070316,0.000004929743,0.0011220232],"genre_scores_gemma":[0.9990833,0.00013585093,0.000044736422,0.000028596107,0.0000066503503,0.000009122727,0.000089909634,0.0000017146001,0.00060003897],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9983645,0.00037160865,0.00008318997,0.00012918033,0.00036944926,0.00068212743],"domain_scores_gemma":[0.9928894,0.0013402444,0.00283644,0.00039108712,0.00074571837,0.0017971435],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.002104271,0.00021581932,0.0004943275,0.0009594502,0.0007418306,0.001133839,0.0007185733,0.00061949936,0.002550444],"category_scores_gemma":[0.008703552,0.00016145372,0.00052415475,0.0006735441,0.0005963788,0.000672209,0.001119254,0.0010726685,0.00035707632],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.000679583,0.002064277,0.9480659,0.00011298032,0.00012514443,0.0003200187,0.003141521,0.00045990606,0.0010793152,0.0011848425,0.00074910285,0.042017486],"study_design_scores_gemma":[0.00001412619,0.0005190185,0.99742126,0.00003591808,0.000034633453,0.000039026992,0.00084183615,0.00007873848,0.0003853979,0.00010197839,0.0005221823,0.000005919917],"about_ca_topic_score_codex":0.037329208,"about_ca_topic_score_gemma":0.074559145,"teacher_disagreement_score":0.037329208,"about_ca_system_score_codex":0.0015235353,"about_ca_system_score_gemma":0.001958682,"threshold_uncertainty_score":0.07422382},"labels":[],"label_agreement":null},{"id":"W2130850661","doi":"10.3102/01623737027003205","title":"Effects of Kindergarten Retention Policy on Children’s Cognitive Growth in Reading and Mathematics","year":2005,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Early Childhood Education and Development","field":"Social Sciences","cited_by":277,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Grade retention; Psychology; Propensity score matching; Multilevel model; Limiting; Promotion (chess); Academic achievement; Developmental psychology; Homogeneous; Harm; Reading (process); Mathematics education; Longitudinal study; Cohort; Social psychology; Political science; Mathematics; Statistics","score_opus":0.020192797292193084,"score_gpt":0.3748215988611014,"score_spread":0.3546288015689083,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2130850661","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9951586,0.0005697705,0.0003523669,0.0012934242,0.000033337,0.000031824133,0.00021970822,0.00003657389,0.0023043863],"genre_scores_gemma":[0.99815243,0.00020697348,0.00021296574,0.00017799214,0.000021003054,0.000034920482,0.00011913336,0.000004634525,0.0010698591],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99164,0.0046038767,0.0003837484,0.00062024314,0.0009343594,0.0018178399],"domain_scores_gemma":[0.9527479,0.028623624,0.011379497,0.0024338954,0.0015318141,0.003283339],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.009345721,0.00037838836,0.00053155783,0.0006394537,0.000645481,0.0012717636,0.0011233332,0.0013860671,0.005791155],"category_scores_gemma":[0.034506872,0.00022282687,0.0012644263,0.00057004153,0.0009828751,0.000855493,0.0014505114,0.0017431265,0.00035245676],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.016245477,0.0058621923,0.8548398,0.00026055265,0.001860224,0.00052944303,0.0010078847,0.016341984,0.002112429,0.009576143,0.003534756,0.087829106],"study_design_scores_gemma":[0.00038909382,0.0067008743,0.97763777,0.0001322843,0.0008097734,0.000084152896,0.001346112,0.006247287,0.0026802642,0.0020643983,0.0018714864,0.000036551788],"about_ca_topic_score_codex":0.024370898,"about_ca_topic_score_gemma":0.03079524,"teacher_disagreement_score":0.024370898,"about_ca_system_score_codex":0.0026916412,"about_ca_system_score_gemma":0.002761843,"threshold_uncertainty_score":0.049425483},"labels":[],"label_agreement":null},{"id":"W2144062847","doi":"10.3102/01623737024003201","title":"Do Low-Achieving Students Benefit More from Small Classes? Evidence from the Tennessee Class Size Experiment","year":2002,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":96,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Class size; Differential effects; Class (philosophy); Mathematics education; Academic achievement; Randomized experiment; Psychology; Reading (process); Differential (mechanical device); Student achievement; Scale (ratio); Quarter (Canadian coin); Mathematics; Statistics; Computer science; Physics; Geography; Political science","score_opus":0.10885470082861255,"score_gpt":0.44710158325711225,"score_spread":0.3382468824284997,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2144062847","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99533373,0.00017128226,0.0002445278,0.0010934108,0.000014335597,0.000025172287,0.00016295417,0.00001272367,0.002941941],"genre_scores_gemma":[0.99768245,0.00011947949,0.00021496978,0.00048274838,0.000014497175,0.000052470023,0.00012705736,0.0000054044735,0.0013009197],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.9968713,0.0019965621,0.00009843424,0.00038093375,0.00037086516,0.0002819361],"domain_scores_gemma":[0.9793115,0.009989089,0.0058899755,0.0019194871,0.0006382299,0.0022517373],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0061343927,0.00030734547,0.00058339944,0.0004417527,0.0007762135,0.00091759796,0.00062536483,0.00096319994,0.0066208695],"category_scores_gemma":[0.015173101,0.00020638529,0.00040264198,0.0004569355,0.0013602258,0.00080959656,0.0007983164,0.0010302529,0.00044820792],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.058797866,0.013304778,0.75994885,0.00040892858,0.0016433813,0.0004901119,0.0038521914,0.0012208782,0.027427956,0.010289659,0.008169155,0.114446305],"study_design_scores_gemma":[0.0015612053,0.0036043704,0.9839946,0.000038156682,0.00026030536,0.000045092846,0.00070624606,0.0006714426,0.0021602772,0.0032140405,0.0037059127,0.000038392183],"about_ca_topic_score_codex":0.01684054,"about_ca_topic_score_gemma":0.025359405,"teacher_disagreement_score":0.01684054,"about_ca_system_score_codex":0.00072768243,"about_ca_system_score_gemma":0.0010125923,"threshold_uncertainty_score":0.033485055},"labels":[],"label_agreement":null},{"id":"W2146700645","doi":"10.3102/0162373708328259","title":"Reading Instruction Time and Homogeneous Grouping in Kindergarten: An Application of Marginal Mean Weighting Through Stratification","year":2008,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":58,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Reading (process); Psychology; Homogeneous; Weighting; Mathematics education; Multilevel model; Reading comprehension; Longitudinal study; Set (abstract data type); Computer science; Statistics; Linguistics; Mathematics","score_opus":0.04578867278231784,"score_gpt":0.3879297195275477,"score_spread":0.3421410467452298,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2146700645","genre_codex":"methods","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":null,"domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2670504,0.0007825299,0.7264994,0.0006007562,0.00013291142,0.0005389336,0.00025487848,0.0002891307,0.0038510305],"genre_scores_gemma":[0.7623647,0.00029388143,0.23505214,0.00012367273,0.0000713477,0.0009054888,0.000151753,0.00007927344,0.0009578184],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9641622,0.028172249,0.0010507258,0.0029950722,0.0026093742,0.001010342],"domain_scores_gemma":[0.8547169,0.12140271,0.006912624,0.012931057,0.0031280091,0.00090861856],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.046125226,0.0007398627,0.0018355506,0.0021981595,0.0013464065,0.0014869565,0.0018642109,0.0009084971,0.002909369],"category_scores_gemma":[0.13780461,0.00079866557,0.0027552485,0.0031538324,0.0035593095,0.002657478,0.0039851572,0.0024893167,0.00015543806],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0019490081,0.00087109435,0.22773449,0.0005443429,0.0031326623,0.0004536364,0.008948515,0.03739807,0.0021961592,0.3396659,0.0025074303,0.37459862],"study_design_scores_gemma":[0.00034158266,0.0024515856,0.14595722,0.00031342398,0.0021181223,0.00031720547,0.0027803944,0.39986685,0.004870671,0.4287201,0.011874465,0.00038845005],"about_ca_topic_score_codex":0.015050883,"about_ca_topic_score_gemma":0.010582025,"teacher_disagreement_score":0.046125226,"about_ca_system_score_codex":0.0024789071,"about_ca_system_score_gemma":0.0028992402,"threshold_uncertainty_score":0.24393636},"labels":[],"label_agreement":null},{"id":"W2153308398","doi":"10.3102/01623737023003251","title":"Class Size and Eighth-Grade Math Achievement in the United States and Abroad","year":2001,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":82,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Class size; Mathematics education; Curriculum; Class (philosophy); Student achievement; Academic achievement; Math education; Psychology; Pedagogy; Computer science","score_opus":0.05050335955880637,"score_gpt":0.4362833388378362,"score_spread":0.38577997927902985,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2153308398","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99847645,0.000028583316,0.000019248619,0.000041316183,0.0000029717169,0.000002511747,0.00047993043,0.0000036939375,0.00094530766],"genre_scores_gemma":[0.9986425,0.00003019355,0.000027774888,0.000011057205,0.0000038110738,0.000006104907,0.00072383817,0.0000020940115,0.0005525967],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99975556,0.000064717264,0.00002371436,0.000043327425,0.00004736465,0.000065354616],"domain_scores_gemma":[0.99657696,0.00064620405,0.0014978612,0.00015381265,0.00028331814,0.0008417583],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0004878906,0.00017691172,0.00021317374,0.0011438826,0.000420248,0.0006404073,0.00032739167,0.00025757652,0.0036401905],"category_scores_gemma":[0.0021394265,0.00011263605,0.0002050632,0.0008530233,0.00031281915,0.0003420152,0.00047332005,0.0005027963,0.00046660495],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00006447223,0.00006191058,0.9973151,0.0000050728604,0.000024389423,0.000038536105,0.00031449212,0.000075302676,0.00009721174,0.00012256764,0.00032672362,0.0015542996],"study_design_scores_gemma":[0.0000025715349,0.000029664272,0.99933124,0.000002870917,0.000004869224,0.0000150097285,0.00030387688,0.00007199717,0.000035824964,0.000015973306,0.00018425458,0.0000018922112],"about_ca_topic_score_codex":0.03861517,"about_ca_topic_score_gemma":0.06389808,"teacher_disagreement_score":0.03861517,"about_ca_system_score_codex":0.0004502083,"about_ca_system_score_gemma":0.00027445753,"threshold_uncertainty_score":0.076780796},"labels":[],"label_agreement":null},{"id":"W2156855199","doi":"10.3102/01623737022004357","title":"Power and Politics in the Adoption of School Reform Models","year":2000,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Educational Assessment and Improvement","field":"Decision Sciences","cited_by":174,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"University of Toronto","funders":"","keywords":"Education reform; Power (physics); Perspective (graphical); Politics; Process (computing); Public administration; Sustainability; Set (abstract data type); Political science; Public relations; Economic growth; Economics; Higher education","score_opus":0.15073264855465782,"score_gpt":0.49702515498723526,"score_spread":0.3462925064325775,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2156855199","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.5752639,0.001899393,0.048275173,0.03957317,0.0002063845,0.0003346324,0.00003931175,0.00007198798,0.33433604],"genre_scores_gemma":[0.9970085,0.00013045546,0.001150416,0.00029796155,0.00002820824,0.000052236934,0.0000041660796,0.000011347942,0.0013167716],"study_design_codex":"theoretical_or_conceptual","study_design_gemma":"qualitative","domain_scores_codex":[0.8915767,0.08318547,0.0019690837,0.0045313714,0.0112259,0.0075113727],"domain_scores_gemma":[0.91043234,0.069126986,0.0065374067,0.0072167255,0.0041856742,0.0025008465],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.07002693,0.00037681405,0.00063017796,0.0037872056,0.012626756,0.016548147,0.0014499382,0.004016387,0.0034967975],"category_scores_gemma":[0.098016225,0.0008000367,0.00076829,0.0032428685,0.05412749,0.012025074,0.008708885,0.0056786863,0.00033187354],"study_design_candidate":"qualitative","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0000576151,0.00006388123,0.005142997,0.000038140857,0.000020809175,0.00018512909,0.04398913,0.0019240238,0.00019906905,0.9354656,0.00056979415,0.012343747],"study_design_scores_gemma":[0.0001324753,0.00018268883,0.008328092,0.00027017356,0.000048220794,0.00019604624,0.068649784,0.0075912946,0.00097544736,0.8537781,0.05976505,0.00008261079],"about_ca_topic_score_codex":0.006338988,"about_ca_topic_score_gemma":0.007360427,"teacher_disagreement_score":0.07002693,"about_ca_system_score_codex":0.012540674,"about_ca_system_score_gemma":0.00920659,"threshold_uncertainty_score":0.3703422},"labels":[],"label_agreement":null},{"id":"W2791023047","doi":"10.3102/0162373718760218","title":"School Improvement Grants in Ohio: Effects on Student Achievement and School Administration","year":2018,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":47,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Regression discontinuity design; Quarter (Canadian coin); Psychology; Demographic economics; Turnover; Regression analysis; Mathematics education; Economics; Statistics; Geography; Mathematics; Management","score_opus":0.032870814993456954,"score_gpt":0.4489629340119663,"score_spread":0.41609211901850934,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2791023047","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.9983695,0.00007002821,0.000095669304,0.00027432272,0.000014328767,0.000012209928,0.00022474217,0.000011625333,0.00092757],"genre_scores_gemma":[0.9990338,0.000043019932,0.00005925517,0.000037216483,0.000009165719,0.00002249051,0.00019547298,0.0000026243004,0.0005968575],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99347043,0.003414367,0.00022964276,0.00049241516,0.0009855529,0.0014076027],"domain_scores_gemma":[0.97767925,0.008843729,0.006787289,0.0011762863,0.0017310573,0.0037823743],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0060174894,0.00024984422,0.000493577,0.001673142,0.0012733472,0.0022546817,0.0009258673,0.0007634811,0.0036747414],"category_scores_gemma":[0.016134892,0.00028442385,0.0008921276,0.002213042,0.0011137961,0.0010175558,0.0023188177,0.0017682128,0.0002252335],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00088910246,0.0023115785,0.96973956,0.000034964596,0.00024552815,0.0000959259,0.00075372995,0.001884785,0.00027874863,0.0013126787,0.0019558743,0.020497508],"study_design_scores_gemma":[0.000051293373,0.0005226852,0.99506915,0.000015808973,0.0001018508,0.000018185852,0.0013944223,0.0013526949,0.00023738384,0.00022336247,0.0010007691,0.000012380331],"about_ca_topic_score_codex":0.08012046,"about_ca_topic_score_gemma":0.08286521,"teacher_disagreement_score":0.08012046,"about_ca_system_score_codex":0.0028617107,"about_ca_system_score_gemma":0.0033319548,"threshold_uncertainty_score":0.1593082},"labels":[],"label_agreement":null},{"id":"W2804370236","doi":"10.3102/0162373718782648","title":"Course Choice and Achievement Effects of a System-Wide Vocational Education and Training Voucher Scheme for Young People","year":2018,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Education Systems and Policy","field":"Social Sciences","cited_by":8,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"","funders":"Queen's University; National Centre for Vocational Education Research; Queen's University Belfast","keywords":"Voucher; Vocational education; Equity (law); School choice; Academic achievement; Training (meteorology); Quality (philosophy); Public economics; Psychology; Mathematics education; Economic growth; Medical education; Pedagogy; Political science; Business; Economics; Accounting","score_opus":0.044016428965278694,"score_gpt":0.4417211995699809,"score_spread":0.3977047706047022,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W2804370236","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.991552,0.00010802485,0.0003149084,0.00059607264,0.00005543403,0.00015828384,0.00016737042,0.000015676036,0.007032339],"genre_scores_gemma":[0.9956762,0.000037458114,0.00016784633,0.00010551626,0.000016755692,0.000057468784,0.00005317661,0.0000031610455,0.0038825278],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.99361974,0.0037333972,0.00021128332,0.00045611334,0.0005906073,0.001388778],"domain_scores_gemma":[0.9822416,0.0094602415,0.00305995,0.0010054721,0.0009950943,0.0032376663],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.008601577,0.0003587284,0.0007699824,0.0006985059,0.0009866933,0.00182392,0.0012565413,0.001465479,0.015670458],"category_scores_gemma":[0.022145152,0.00026670893,0.001075005,0.0008884163,0.0019454753,0.0013117291,0.002213142,0.001995343,0.0006787724],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.1408151,0.0953505,0.3578343,0.00076831225,0.0017214525,0.0007289766,0.00358071,0.05555213,0.006905379,0.09512216,0.0081897825,0.23343119],"study_design_scores_gemma":[0.013196803,0.067521624,0.8232177,0.0002218784,0.0019485985,0.00009540669,0.0044570677,0.05169461,0.007045531,0.018652197,0.011785285,0.00016329378],"about_ca_topic_score_codex":0.032858003,"about_ca_topic_score_gemma":0.021739904,"teacher_disagreement_score":0.032858003,"about_ca_system_score_codex":0.006102841,"about_ca_system_score_gemma":0.0036230052,"threshold_uncertainty_score":0.065333545},"labels":[],"label_agreement":null},{"id":"W3008309614","doi":"10.3102/0162373720906217","title":"Teacher Coaching in a Simulated Environment","year":2020,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Teacher Education and Leadership Studies","field":"Social Sciences","cited_by":181,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":false,"ca_institutions":"Impact","funders":"National Academy of Education; University of Virginia; Spencer Foundation","keywords":"Coaching; Psychology; Perception; Mathematics education; Medical education; Teacher education; Pedagogy; Applied psychology; Medicine","score_opus":0.21363410786154396,"score_gpt":0.4796332351777997,"score_spread":0.26599912731625575,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W3008309614","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.99266535,0.000022894681,0.0049521774,0.000067150686,0.000011710601,0.00020029624,0.000030199659,0.00005703485,0.0019930694],"genre_scores_gemma":[0.99495643,0.000043860335,0.0036923983,0.00002742635,0.000005850538,0.0003593938,0.000052238753,0.0000062602717,0.00085612544],"study_design_codex":"design_other","study_design_gemma":"observational","domain_scores_codex":[0.9965289,0.0024919242,0.00009684606,0.00027178563,0.00035074665,0.00025979834],"domain_scores_gemma":[0.9956779,0.0023997664,0.0005061206,0.00059366156,0.00024098722,0.000581652],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.0014396021,0.00030525928,0.00033867758,0.0002902538,0.0006915922,0.00059381255,0.00052485516,0.0006708934,0.0025996729],"category_scores_gemma":[0.007190526,0.00026800946,0.00030885293,0.0001920681,0.00055201753,0.00038344768,0.0010138105,0.00056705973,0.0004769219],"study_design_candidate":"observational","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.03720787,0.09673452,0.16698128,0.0009961756,0.0004525252,0.0010453784,0.014783909,0.18573694,0.06300106,0.004773986,0.0032878157,0.42499852],"study_design_scores_gemma":[0.01178811,0.2813003,0.3161496,0.00065798283,0.00057818316,0.0019554638,0.013511571,0.22203787,0.09712928,0.0058290856,0.048670247,0.00039226116],"about_ca_topic_score_codex":0.0014330177,"about_ca_topic_score_gemma":0.0023310038,"teacher_disagreement_score":0.0025996729,"about_ca_system_score_codex":0.0005211629,"about_ca_system_score_gemma":0.0014431739,"threshold_uncertainty_score":0.0086967945},"labels":[],"label_agreement":null},{"id":"W4390877660","doi":"10.3102/01623737231218735","title":"Beyond Prescriptive Reforms: An Examination of North Carolina’s Flexible School Restart Program","year":2024,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"School Choice and Performance","field":"Social Sciences","cited_by":1,"is_retracted":false,"has_abstract":true,"route_ca_aff":false,"route_ca_fund":false,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"","funders":"","keywords":"Autonomy; Flexibility (engineering); Principal (computer security); Quarter (Canadian coin); School choice; Mathematics education; South carolina; Psychology; Pedagogy; Public administration; Sociology; Political science; Economics; Management; Computer science; Law; History","score_opus":0.056100336068339335,"score_gpt":0.4434867261868464,"score_spread":0.38738639011850706,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4390877660","genre_codex":"empirical","genre_gemma":"empirical","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"empirical","genre_consensus":"empirical","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.96796197,0.00052746676,0.0005720648,0.008815016,0.000042163894,0.0002072737,0.00042956058,0.000047263176,0.021397406],"genre_scores_gemma":[0.9920284,0.00027692702,0.00053472957,0.0022117186,0.000039917075,0.00019986003,0.0004477242,0.000020287258,0.0042402796],"study_design_codex":"observational","study_design_gemma":"observational","domain_scores_codex":[0.986878,0.00492812,0.00029339077,0.00083976897,0.0040723295,0.0029884218],"domain_scores_gemma":[0.9437372,0.027117075,0.009185608,0.0028011368,0.010028061,0.0071309535],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.011344181,0.00028163023,0.00064877875,0.002091363,0.0040685027,0.003838946,0.0033784986,0.0013695244,0.003902109],"category_scores_gemma":[0.032813717,0.00027829615,0.0007339935,0.0034720562,0.0021781796,0.0021154375,0.003779605,0.0032352682,0.0002791827],"study_design_candidate":"observational","study_design_consensus":"observational","about_ca_topic_candidate":true,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.0012414808,0.0038810226,0.69438434,0.0005581112,0.0005290398,0.0015325928,0.03591042,0.01645435,0.0024655852,0.039973177,0.037879504,0.1651903],"study_design_scores_gemma":[0.00007322142,0.0007475404,0.9359057,0.00028956428,0.000133063,0.00005679467,0.028288975,0.0050489306,0.0006293174,0.0017411959,0.027008114,0.00007746404],"about_ca_topic_score_codex":0.5157947,"about_ca_topic_score_gemma":0.69338703,"teacher_disagreement_score":0.5157947,"about_ca_system_score_codex":0.025341531,"about_ca_system_score_gemma":0.034760166,"threshold_uncertainty_score":0.97411376},"labels":[],"label_agreement":null},{"id":"W4407384694","doi":"10.3102/01623737241311537","title":"Measuring Grading Standards at High Schools: A Methodological Contribution, an Example, and Some Policy Implications","year":2025,"lang":"en","type":"article","venue":"Educational Evaluation and Policy Analysis","topic":"Higher Education Learning Practices","field":"Social Sciences","cited_by":0,"is_retracted":false,"has_abstract":true,"route_ca_aff":true,"route_ca_fund":true,"route_ca_venue":false,"route_about_ca":true,"ca_institutions":"Wilfrid Laurier University","funders":"Canadian Institutes of Health Research; University of Alberta","keywords":"Grading (engineering); Academic standards; Mathematics education; Policy analysis; Higher education; Econometrics; Psychology; Political science; Economics; Economic growth; Public administration; Engineering","score_opus":0.2785140261361238,"score_gpt":0.5528497145082416,"score_spread":0.2743356883721178,"validation_status":"score_only:v0-immature-baseline","prediction":{"id":"W4407384694","genre_codex":"methods","genre_gemma":"methods","domain_codex":null,"domain_gemma":null,"model_version":"metacan-v3-hybrid-931329e0061c","genre_candidate":"methods","genre_consensus":"methods","domain_candidate":null,"domain_consensus":null,"prediction_status":"machine_predicted_unvalidated","genre_scores_codex":[0.2619008,0.01260074,0.5255494,0.13083737,0.0028425388,0.008584947,0.0028459835,0.00023647328,0.054601718],"genre_scores_gemma":[0.6648045,0.0027485902,0.310485,0.008142317,0.00081605575,0.0072797276,0.0007280841,0.00009947094,0.0048961905],"study_design_codex":"observational","study_design_gemma":"theoretical_or_conceptual","domain_scores_codex":[0.6873401,0.26238468,0.012887649,0.0080396775,0.027128393,0.0022195478],"domain_scores_gemma":[0.5959744,0.26404157,0.0376239,0.03943378,0.06116933,0.0017570473],"candidate_categories":[],"consensus_categories":[],"category_scores_codex":[0.17960678,0.0013222679,0.0009950617,0.005620823,0.006129848,0.005798927,0.005487874,0.0036828613,0.0013692146],"category_scores_gemma":[0.33281493,0.0008652242,0.0010200969,0.016951753,0.009230321,0.0038346015,0.004735205,0.0031233933,0.0003538526],"study_design_candidate":"theoretical_or_conceptual","study_design_consensus":null,"about_ca_topic_candidate":false,"about_ca_topic_consensus":false,"about_ca_system_candidate":false,"about_ca_system_consensus":false,"study_design_scores_codex":[0.00031674062,0.0017177948,0.5035846,0.0024211314,0.000688808,0.0008925075,0.027449924,0.0073432922,0.0020093487,0.23833966,0.02108605,0.19415015],"study_design_scores_gemma":[0.0006088793,0.0013197075,0.43293798,0.006716062,0.0012880735,0.0017641828,0.06831056,0.03182545,0.010475295,0.298472,0.1453036,0.0009781318],"about_ca_topic_score_codex":0.16478448,"about_ca_topic_score_gemma":0.21003503,"teacher_disagreement_score":0.17960678,"about_ca_system_score_codex":0.014156786,"about_ca_system_score_gemma":0.021727279,"threshold_uncertainty_score":0.9498628},"labels":[],"label_agreement":null}]}