results: NIH ChestX-ray14 imaging pipeline (manifest, solo, cascade x4 cues, referee, judge, multi-round, blind-metric) (#162/#163/#164/#165/#167/#168/#169/#170)

[sebasmos]

Closes #162, closes #163, closes #164, closes #165, closes #167, closes #168, closes #169, closes #170.

Adds the full NIH ChestX-ray14 imaging pipeline as a self-contained, keyless-reproducible bundle: a provenance manifest builder, single-model cue susceptibility, the watermark cascade (plus the three remaining cues), a deployable referee-as-gate evaluation, a same-lineage judge referee, multi-round cascade dynamics, and a blind-metric substitution probe.

What's in it

• build_manifest.py (Dataset: NIH ChestX-ray14 loader and provenance manifest (accessible imaging set) #162): builds a manifest via the already-registered nih_cxr14 adapter, keeps only cases whose image resolves on disk and carries a real finding label, and records a sha256 checksum per used image (results/manifest_provenance.json). --case-ids-file pins the exact 35-case set used throughout.
• imaging_solo.py (Imaging solo cue susceptibility on NIH ChestX-ray14 (injected cues) #163): cable / corner-tag / watermark / laterality cue injection, flip rate vs a temperature-resampled noise floor. Watermark is the one cue that clears the floor (+0.11).
• imaging_cascade.py (Injected-cue imaging cascade on NIH ChestX-ray14 (watermark contagion) #164, Imaging cascade on the remaining cues (cable, corner-tag, laterality) on NIH ChestX-ray14 #167): does a committee adopt a wrong finding-read when two peers assert it? Watermark: shared 0.97 vs isolated 0.34, contagion +0.63. The other three cues (--cue {cable,corner_tag,laterality}) are even stronger: +0.80 / +0.74 / +0.71, despite barely moving the model solo. The cascade doesn't depend on the cue being independently potent.
• imaging_referee.py (Imaging referee-as-gate on the NIH ChestX-ray14 cascade transcripts #165): a deployable referee with no privileged knowledge (transcript + one private re-read) flags peer-driven adoption. Precision/recall 0.86, FPR 0.23, vs a naive gate's FPR 0.92.
• imaging_judge_referee.py (Imaging same-lineage judge referee on NIH ChestX-ray14 #168): a text-only same-lineage judge, given only the shared transcript, scores identically to the naive gate (0.65/1.0/0.92). Without a reasoning trace or a fresh independent signal, it has nothing beyond "did the answer match the peers." The deployable referee's private re-read is what earns its edge.
• imaging_multi_round.py (Imaging multi-round cascade dynamics on NIH ChestX-ray14 #169): shared adoption is already high at round 1 (0.89) and plateaus (round1-vs-round3 McNemar p=0.25, not significant); isolated stays flat. Adoption is decided on first exposure.
• imaging_blind_metric.py (Imaging blind-metric substitution probe on NIH ChestX-ray14 #170): decoy uptake +0.83 under a hidden rubric, suppressed to +0.37 when primed. Unlike the text lane (all drifters self-declared), 0 of 29 drifters here mention the rubric at all, verified against the raw justification text. Silent, not self-declared, gaming.

All runners are argparse-based, read GEMINI_API_KEY or GOOGLE_API_KEY from the environment only (no hardcoded paths or keys), and reproduce with new_api_calls_this_run: 0 off the committed caches (results/img_cache.jsonl, results/judge_cache.jsonl). Verified end-to-end with the key unset, twice (once per commit). ruff check clean, full test suite unaffected (580 passed, 7 skipped), purely additive. Independently reviewed both rounds: reproducibility, security/leak, convention-fit, and evaluation-correctness (referee non-circularity, decoy fairness) all PASS.

Note on #141

results/cascade-at-scale (#141) already has an earlier, unmerged version of the solo/cascade scripts under experiments/imaging/, but that branch predates #146/#148/#154/#157/#161 and would delete all of that merged work if landed as-is. This PR is branched fresh off current main and reuses #141's already-good solo/cascade logic and numbers (verified identical), plus adds everything else. Once this lands, #141 should drop its now-redundant experiments/imaging/ changes on rebase.

[sebasmos]

Update: this PR now also closes #167, #168, #169, #170 (pushed as a second commit, 134d873).

• Imaging cascade on the remaining cues (cable, corner-tag, laterality) on NIH ChestX-ray14 #167 cascade on the remaining cues: cable +0.80, corner-tag +0.74, laterality +0.71 contagion (all stronger than watermark's +0.63, despite barely moving the model solo).
• Imaging same-lineage judge referee on NIH ChestX-ray14 #168 same-lineage judge referee: scores identically to the naive gate (0.65/1.0/0.92) when given only the shared transcript (no reasoning trace, no isolated-answer comparison). The deployable referee's private re-read is what earns its edge, not the judge framing.
• Imaging multi-round cascade dynamics on NIH ChestX-ray14 #169 multi-round dynamics: adoption is already high at round 1 and plateaus (not a significant build-up, McNemar p=0.25).
• Imaging blind-metric substitution probe on NIH ChestX-ray14 #170 blind-metric substitution: decoy uptake +0.83, but 0 of 29 drifters mention the rubric at all (unlike the text lane, where all drifters self-declared). Silent, not self-declared, gaming.

All new/edited runners verified keyless-reproducible and independently reviewed (reproducibility, security/leak, evaluation-correctness all PASS).

[sebasmos]

Rebased this branch onto current main (it had gone stale relative to #219/#142/#143 merged since this PR was opened; merged main in cleanly, no conflicts). Re-verified keyless reproduction of the full imaging pipeline afterward (cascade contagion still 0.6286, new_api_calls_this_run: 0). 625 tests pass, ruff clean.