Obaidullahmiakhil/ich-ef03-scores-repaired

EF03 — repaired score dumps

Re-scored val/test outputs for phase2-effnetb3-bigru-ef3-v1 after fixing the image-path defect that Step 4 uncovered: ~20% of val and test slices were scored as blank black frames because the manifest's split column disagreed with the directory the PNG actually sat in.

Nothing was retrained. Same encoder weights, same sequence-head weights, same architecture, same preprocessing, same labels, same train/val/test assignment. The only change is that images are resolved by disk_split instead of split.

Re-score version	step5-repaired-v1
Splits re-scored	val, test
Reload fidelity	PASS (max abs diff 9.52e-04, tol 0.002)
Quarantined blank slices	190

Contents

scores/seq_gate_{val,test}.npz      cascade gate head
scores/seq_sub_{val,test}.npz       cascade subtype heads
scores/seq_winner_{val,test}.npz    RSNA-2019 winner-style head
scores/cnn_slice_{val,test}.npz     2D CNN, no sequence context
tables/damage_report.csv            old vs new AUC, split by was-blank / was-correct
tables/fidelity_check.csv           proof the checkpoint reloaded faithfully
tables/quarantined_blank.csv        slices excluded (blank at source)
figures/                            PDF + PNG, 300 dpi
results.json                        every number, machine-readable

Each .npz holds scores, targets, ids — Step 2's schema exactly.

How these were validated

About 80% of val/test slices were already at the correct path, so Step 2 scored them from real pixels. Re-scoring those same slices reproduces the original values to within 0.002. That is what makes the new scores on the other 20% trustworthy: the encoder demonstrably reloaded as itself.

Next step

Re-run the Step-4 audit notebook with its score directory pointed here. Experiments F (conformal validity) and G (study-level evaluation) then execute unchanged on honest scores.