MIMIC-CXR imaging results are not independently verifiable: no per-case rows are committed

[sebasmos]

Raising this as a verification gap rather than a bug. Nothing here is wrong, but a third of the imaging evidence currently rests on files nobody outside the run can check, including me.

experiments/mimic_cxr_image/.gitignore excludes:

results/**/*.jsonl
results/**/*.log
manifests/provenance.json
manifests/*.csv

So the lane commits 8 summary JSONs and zero per-case rows. For comparison, on main: MedQA commits 176 per-case files and 25 cache files, NIH 28 and 8. Every MIMIC-CXR imaging rate is therefore unrecomputable from the repo, and the per-image checksums the docs describe are not present either.

This is legitimate under the PhysioNet terms, and I am not suggesting the raw data be committed. The problem is only that there is currently no DUA-safe path to confirm the numbers.

What rests on it:

• the MIMIC contagion rows, n=43 / 87 / 179 / 169
• the MIMIC blind-metric column, 78 drifters, 0 naming the rubric
• the n=834 image solo null
• the referee comparison at n=417, which I have now withdrawn from the paper's referee table for a separate reason: that cohort's own README records 79% of plants coinciding with the true read, so its adoptions largely measure the holdout being moved toward the correct answer, and it cannot be restricted to the genuinely-false subgroup without the per-case rows

Either of these would close it:

1. Commit a DUA-safe derived file. Per-case rows reduced to non-identifying fields only, for example case_index, iso_adopt, shared_adopt, solo_flip, gt with no study or patient identifiers and no report text. That is enough to recompute every published rate and to run the plant-direction restriction.
2. Independent confirmation. A second person with PhysioNet access re-runs the arms and confirms the summaries, recorded on this issue.

@amarzullo24 you own this lane. @armaanvgrewal you have access and offered to run #347's arms, so option 2 may be quickest if travel makes option 1 awkward.

Related: #343 covers the missing runner scripts, which is the other half of making this lane reproducible. #371 is separately blocked on a solo_flip versus iso_adopt polarity question in the same lane, and resolving that likely needs the same per-case data.

[amarzullo24]

Fixed via option 1 -- de-identified per-case rows committed on #372 (c2104b6), under experiments/mimic_cxr_image/results/deid/:

file	n	columns
solo.csv / nih_match_solo.csv	834 / 35	case_index, clean_correct, {cable,corner_tag,watermark,laterality}_flip, noise_flip
cascade.csv / referee_cascade.csv / system_flag.csv	215 / 417 / 834	case_index, clean_correct, iso_adopt, shared_adopt
strength_cascade.csv	834	case_index, clean_correct, op{0.15,0.3,0.45}_{iso_adopt,shared_adopt,solo_flip}
referee.csv	417	case_index, gt, ref_flag, naive_flag
blind_metric.csv	141	case_index, base_is_decoy, blind_is_decoy, aware_is_decoy, named_rubric_when_drifted

Non-identifying only: case_index is assigned by sorting on case_id then dropping it -- no dicom/patient/study id, no report text, no raw reads. DUA-safe. Recomputes every published rate and supports the plant-direction restriction (clean_correct==1 is the genuinely-false subgroup you flagged). Verified to reproduce the committed summaries exactly (cascade contagion 0.1442, system-flag 0.1271, solo watermark flip 0.1763 / noise 0.1727, referee precision 0.4808, blind uptake 0.5319).

These rows are from a complete corrected rerun (#372), cross-checked against every related issue/PR

Before committing I re-ran the whole battery and audited it against the thread, to make sure nothing here reintroduces a known problem:

Checked	Source	Result
Ground-truth plant on all cascade-family arms	#332 #333 #338 #339	planted-truth 0% on every transcript; iso adoption ~0.78-0.85 (the constant-"no" signature; the old flip(clean) plant gives ~0.18)
nih_match sized by image (35), not 50 studies-expanded	#329	manifest+summary n=35; natively re-run behind a manifest-verify guard
Cache pinned inside the MIMIC lane	#343 #347	run_battery pins results/img_cache.jsonl
No --n truncation	#347	n=215/417/834
All arms incl. referee (#315) + nih_match	#315 #343	8 arms (#371 was missing both)
vs_peer_assertion_cascade populated (the b8b4905 regression)	#347	present
manifest = summary = transcript = de-id n	internal	all match, all 8 arms
Same bug class / fix direction as the NIH lane	#350	consistent

Supersedes #371 (partial: missing nih_match + referee, and its nih_match ran on a stale 50-image manifest). On the referee cohort you withdrew: clean_correct in referee_cascade.csv now lets that be restricted to the genuinely-false subgroup without the raw rows, if it helps re-include it. Remaining open item is the tab:imgcascade presentation (full-cohort vs genuinely-false subgroup) -- numbers for both are ready.

[sebasmos]

Closing as part of the pre-release cleanup. The paper is submitted and the repository is being prepared to go public, so the board is being reduced to the single item that still gates a published number: #393, the MIMIC-CXR same-lineage judge cell. Everything else is out of scope for this release. This is not a judgement on the work: unmerged branches were archived to git bundles before deletion, and anything here can be reopened if the lane is picked up again.