| { | |
| "comparison": "Epoch 21 to epoch 22 to partial epoch 23 (79.95% through training epoch)", | |
| "estimator": "Exact per-tag AP on matching image identities and tag support", | |
| "counts": { | |
| "continued_declines": 2007, | |
| "partial_recoveries_high_suspect": 1959, | |
| "new_declines_since_e22": 4543 | |
| }, | |
| "continued_declines_definition": "E21 > E22 > partial E23", | |
| "partial_recoveries_definition": "E22 < partial E23 < E21; treat as high suspect pending review", | |
| "new_declines_definition": "E22 >= E21 and partial E23 < E22", | |
| "interpretation": "Any numerical decrease counts. These are model/tag stability and data-review flags, not proof of wrong labels or statistically significant deterioration." | |
| } |