Findings
Findings / Verdict flips
When we changed our mind 5 claim families changed their verdict as more data arrived; here is each mind-change, in sequence.
5 published examples changed verdict.
Steps are shown in source order; 142 runs have no recorded date.
NBA Three in four fatigue First NULL_LOCAL on n=4762 (player_boxscores_2024_25_2025_26), then CONFIRMED_LOCAL on n=7222 (player_boxscores_2024_25_2025_26).
Run 1 -- NULL_LOCAL on player_boxscores_2024_25_2025_26, n=4,762, -0.916 (run date not recorded)Run 2 -- CONFIRMED_LOCAL on player_boxscores_2024_25_2025_26, n=7,222, -1.213 (run date not recorded)MLB Reliever 3in3d fatigue First NULL_LOCAL on n=9054 (bullpen_relief_chains+savant_full__2023-2025), then CONFIRMED_LOCAL on n=8927 (bullpen_relief_chains+savant_full__2023-2025).
Run 1 -- NULL_LOCAL on bullpen_relief_chains+savant_full__2023-2025, n=9,054, 0.01839 (run date not recorded)Run 2 -- NULL_LOCAL on bullpen_relief_chains+savant_full__2023-2025, n=8,897, -0.03538 (run date not recorded)Run 3 -- CONFIRMED_LOCAL on bullpen_relief_chains+savant_full__2023-2025, n=8,927, -0.04412 (run date not recorded)MLB Reliever 3in3d fatigue combined First PROVISIONAL on n=26878 (bullpen_relief_chains+savant_full__2023-2025), then NULL_LOCAL on n=26878 (bullpen_relief_chains+savant_full__2023-2025).
Run 1 -- PROVISIONAL on bullpen_relief_chains+savant_full__2023-2025, n=26,878, effect not published (run date not recorded)Run 2 -- NULL_LOCAL on bullpen_relief_chains+savant_full__2023-2025, n=26,878, effect not published (run date not recorded)MLB Staff dayafter fatigue chain First NULL_LOCAL on n=2087 (savant_full__2023-2025_team_day_pitchcount), then NULL_LOCAL on n=2078 (savant_full__2023-2025_team_day_pitchcount).
Run 1 -- NULL_LOCAL on savant_full__2023-2025_team_day_pitchcount, n=2,087, 0.0789 (2026-07-10T08:23:40Z )Run 2 -- CONFIRMED_LOCAL on savant_full__2023-2025_team_day_pitchcount, n=2,160, 0.3915 (2026-07-10T08:23:40Z )Run 3 -- NULL_LOCAL on savant_full__2023-2025_team_day_pitchcount, n=2,078, 0.3266 (2026-07-10T08:23:40Z )Tennis Break point conversion by set number First NOT_TESTABLE on n=34145 (slam_points_2011_2015), then CONFIRMED_LOCAL on n=34145 (slam_points_2011_2015).
Run 1 -- NOT_TESTABLE on slam_points_2011_2015, n=34,145, -0.017 (run date not recorded)Run 2 -- CONFIRMED_LOCAL on slam_points_2011_2015, n=34,145, -0.017 (run date not recorded)Retracted-claim latency
Claim Sport Days lived What killed it Meta label divergence bucket Q1 MLB undated FAILED_REPLICATION on corpus 'ingame_grade_joined_mlb_synthetic_checkpoints_disjoint' (n=20, effect=0.45) Meta label divergence bucket Q2 MLB undated FAILED_REPLICATION on corpus 'ingame_grade_joined_mlb_synthetic_checkpoints_disjoint' (n=19, effect=0.3158) Meta label divergence bucket Q3 MLB undated FAILED_REPLICATION on corpus 'ingame_grade_joined_mlb_synthetic_checkpoints_disjoint' (n=19, effect=0.3158) Meta label divergence bucket Q4 MLB undated FAILED_REPLICATION on corpus 'ingame_grade_joined_mlb_synthetic_checkpoints_disjoint' (n=18, effect=0.3333) Milb debut freshness flag MLB undated REJECT on corpus 'games_current_season_2026__savant_debut_proxy' (n=39, effect=0.000394) Oaa trailing fielding quality babip conversion MLB undated REJECT on corpus 'savant_playlevel__2025_2026' (n=55063, effect=None) Substitution timing moderates shift replication big4 2015 16 Soccer 0 FAILED_REPLICATION on corpus 'disjoint_competition_group' (n=1691, effect=0.01861)
Scroll horizontally for all columns.
run_ts present on 145 of 287 history rows -- run_ts is present on only some rows; undated flips show sequence without dates.
5 flips is a small, anecdotal set -- this is qualitative case study, not a rate. run_ts is missing on many history rows, so 'days-lived' is a lower bound and some flips are undated. a flip reflects more/re-cleaned data arriving, NOT that the earlier verdict was careless. Snapshot generated 2026-07-25
Findings