{ "schemaVersion": 1, "study": "temporal-selection", "reviewed": "2026-09-20", "evidenceType": "extracted-source-report", "reproducedInThisWebsiteReview": false, "sources": [ { "name": "EXP011g-placebo-gates.md", "sha256": "8986d1842c8084791b3137047669833b67deeae00004f50919b6c0baaaa5371e" } ], "tables": [ { "source": "EXP011g-placebo-gates.md", "rows": [ [ "run", "split", "mode", "pair_acc", "calc_tol", "render_tol", "invalid_order", "group_both", "group_pA*pB", "group_lift" ], [ "---", "---", "---", "---:", "---:", "---:", "---:", "---:", "---:", "---:" ], [ "p44-exp011g-time-selection-s42", "eval", "real", "0.8542", "0.8542", "0.8583", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_ood_primary", "real", "0.9525", "0.9525", "0.9550", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_far_stress", "real", "0.9425", "0.9425", "0.9475", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_same_text_diff_time", "real", "0.8975", "0.8975", "0.8850", "0.0000", "0.8050", "0.8055", "-0.0005" ], [ "p44-exp011g-time-selection-s42", "test_same_time_diff_question", "real", "0.8875", "0.8875", "0.8825", "0.0000", "0.7900", "0.7875", "0.0025" ], [ "p44-exp011g-time-selection-s42", "eval", "shuffle", "0.3208", "0.0292", "0.0292", "0.1167", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_ood_primary", "shuffle", "0.2975", "0.1175", "0.1150", "0.0600", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_far_stress", "shuffle", "0.2275", "0.1175", "0.1175", "0.0675", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_same_text_diff_time", "shuffle", "0.2825", "0.0350", "0.0325", "0.1175", "0.0150", "0.0775", "-0.0625" ], [ "p44-exp011g-time-selection-s42", "test_same_time_diff_question", "shuffle", "0.2800", "0.0425", "0.0400", "0.1625", "0.0750", "0.0782", "-0.0032" ], [ "p44-exp011g-time-selection-s42", "eval", "zero", "0.3292", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_ood_primary", "zero", "0.3300", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_far_stress", "zero", "0.3450", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s42", "test_same_text_diff_time", "zero", "0.3350", "0.0000", "0.0000", "1.0000", "0.0000", "0.1122", "-0.1122" ], [ "p44-exp011g-time-selection-s42", "test_same_time_diff_question", "zero", "0.3200", "0.0000", "0.0000", "1.0000", "0.0900", "0.1023", "-0.0123" ], [ "p44-exp011g-time-selection-s43", "eval", "real", "0.8417", "0.8417", "0.8417", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_ood_primary", "real", "0.9200", "0.9200", "0.9200", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_far_stress", "real", "0.9225", "0.9225", "0.9250", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_same_text_diff_time", "real", "0.8300", "0.8300", "0.8200", "0.0000", "0.6900", "0.6859", "0.0041" ], [ "p44-exp011g-time-selection-s43", "test_same_time_diff_question", "real", "0.8400", "0.8475", "0.8400", "0.0000", "0.7150", "0.7054", "0.0096" ], [ "p44-exp011g-time-selection-s43", "eval", "shuffle", "0.3208", "0.0292", "0.0292", "0.1208", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_ood_primary", "shuffle", "0.2875", "0.1200", "0.1175", "0.0525", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_far_stress", "shuffle", "0.2350", "0.1275", "0.1275", "0.0625", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_same_text_diff_time", "shuffle", "0.3000", "0.0325", "0.0275", "0.1400", "0.0250", "0.0894", "-0.0644" ], [ "p44-exp011g-time-selection-s43", "test_same_time_diff_question", "shuffle", "0.2875", "0.0400", "0.0375", "0.1575", "0.0700", "0.0824", "-0.0123" ], [ "p44-exp011g-time-selection-s43", "eval", "zero", "0.3500", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_ood_primary", "zero", "0.3450", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_far_stress", "zero", "0.3375", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s43", "test_same_text_diff_time", "zero", "0.3375", "0.0000", "0.0000", "1.0000", "0.0000", "0.1134", "-0.1134" ], [ "p44-exp011g-time-selection-s43", "test_same_time_diff_question", "zero", "0.3250", "0.0000", "0.0000", "1.0000", "0.1200", "0.1055", "0.0145" ], [ "p44-exp011g-time-selection-s44", "eval", "real", "0.8625", "0.8625", "0.8625", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_ood_primary", "real", "0.9100", "0.9100", "0.9150", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_far_stress", "real", "0.9350", "0.9375", "0.9400", "0.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_same_text_diff_time", "real", "0.8800", "0.8800", "0.8700", "0.0000", "0.7650", "0.7743", "-0.0093" ], [ "p44-exp011g-time-selection-s44", "test_same_time_diff_question", "real", "0.8750", "0.8775", "0.8650", "0.0000", "0.7700", "0.7655", "0.0045" ], [ "p44-exp011g-time-selection-s44", "eval", "shuffle", "0.3333", "0.0208", "0.0208", "0.1167", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_ood_primary", "shuffle", "0.2975", "0.1225", "0.1200", "0.0550", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_far_stress", "shuffle", "0.2250", "0.1250", "0.1250", "0.0600", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_same_text_diff_time", "shuffle", "0.3025", "0.0325", "0.0325", "0.1250", "0.0300", "0.0905", "-0.0605" ], [ "p44-exp011g-time-selection-s44", "test_same_time_diff_question", "shuffle", "0.2750", "0.0400", "0.0400", "0.1450", "0.0550", "0.0754", "-0.0204" ], [ "p44-exp011g-time-selection-s44", "eval", "zero", "0.3667", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_ood_primary", "zero", "0.3075", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_far_stress", "zero", "0.3600", "0.0000", "0.0000", "1.0000", "-", "-", "-" ], [ "p44-exp011g-time-selection-s44", "test_same_text_diff_time", "zero", "0.3525", "0.0000", "0.0000", "1.0000", "0.0000", "0.1225", "-0.1225" ], [ "p44-exp011g-time-selection-s44", "test_same_time_diff_question", "zero", "0.3375", "0.0000", "0.0000", "1.0000", "0.1150", "0.1125", "0.0025" ] ] } ], "limitation": "Published tables are source-report extracts. Model weights, complete datasets and independent reproduction are not included." }