MIMIC-CXR text (reports) standard-battery replication (tracking)

[sebasmos]

Core battery (apples-to-apples, do these first; each maps to a cross-dataset table cell)

• Solo shortcut susceptibility + noise floor (MIMIC-CXR text) #316 Solo susceptibility + noise floor — PR feat: MIMIC-CXR text adapter + solo susceptibility real-API results (n=35, n=600) #336 (approved by @sebasmos, pending merge)
• Cascade contagion: bare vs case-anchored confident wrong seed (MIMIC-CXR text) #317 Cascade contagion — PR feat: MIMIC-CXR text cascade contagion, bare vs case-anchored seed (#317) #348 (pending review); anchored-seed contagion resolved via Anchored-seed shared-vs-isolated contagion missing from #317 (MIMIC-CXR text) #355 (closed)
• Break-it A: contaminated shared context (system pre-screen flag) (MIMIC-CXR text) #318 Break-it A (system flag) — PR feat: MIMIC-CXR text Break-it A, contaminated system pre-screen flag (#318) #353 (pending review)
• Break-it C: case-anchored plausible seed vs generic (MIMIC-CXR text) #319 Break-it C (plausible seed) — closed, satisfied by PR feat: MIMIC-CXR text cascade contagion, bare vs case-anchored seed (#317) #348's push_c.py dose-response
• Break-it D: blind-metric decoy (a rewarded spurious cue) (MIMIC-CXR text) #320 Break-it D (blind-metric) — PR feat: MIMIC-CXR text blind-metric substitution probe (#320) #356 (pending review)
• Referee detection: naive gate vs targeted vs deployable (MIMIC-CXR text) #321 Referee detection — PR feat: MIMIC-CXR text referee detection, naive gate vs deployable vs judge (#321) #357 (pending review); also closes Anchored-seed shared-vs-isolated contagion missing from #317 (MIMIC-CXR text) #355

---

Tracking issue: replicate the finished MedQA/NIH standard battery on MIMIC-CXR report text so results are apples-to-apples across datasets. Real data + real API only; write results under experiments/mimic_cxr_text/results/ in the same JSON/JSONL format as experiments/medqa/results/. Mirror of the MedMCQA tracker #291.

Foundational (run first)

• Solo shortcut susceptibility + noise floor
• Cascade baseline: bare vs case-anchored confident wrong seed (contagion = shared minus isolated) — bare seed done; anchored-seed contagion gap in Anchored-seed shared-vs-isolated contagion missing from #317 (MIMIC-CXR text) #355

Seed shape

• Plausible vs implausible planted distractor — covered by Break-it C: case-anchored plausible seed vs generic (MIMIC-CXR text) #319 (closed)
• Rationale validity: bare vs valid-but-wrong vs named-fallacy
• Seed confidence (hedged vs confident) + attributed-tier identity
• Cue dose-response + text cue types (primacy/position, negation/qualifier)

Social pressure / committee structure

• Majority-pressure (Asch) + committee-size sweep
• Dissenter / unanimity break
• Deliberation framing crossed with the anchored seed
• Super-additivity 2x2 (system flag AND anchored peer)

Hierarchy / orchestrator

• Hierarchy dominance (order-permutation) + wrong orchestrator vs wrong peer
• Authority gradient ladder + leader-as-auditor remediation

Controls & robustness

• True-peer negative control (plant the CORRECT answer)
• Test-awareness + temperature/run-to-run reliability
• Onset + contagion/deference battery over transcripts

Break-it arms

• Break-it A (system pre-screen flag) — PR feat: MIMIC-CXR text Break-it A, contaminated system pre-screen flag (#318) #353
• Break-it C (case-anchored plausible seed vs generic) — closed via Break-it C: case-anchored plausible seed vs generic (MIMIC-CXR text) #319/PR feat: MIMIC-CXR text cascade contagion, bare vs case-anchored seed (#317) #348
• Break-it D (blind-metric decoy) — PR feat: MIMIC-CXR text blind-metric substitution probe (#320) #356

Referee

• Referee detection: naive gate vs targeted vs deployable — PR feat: MIMIC-CXR text referee detection, naive gate vs deployable vs judge (#321) #357

Adapter: shares the MIMIC-CXR loader (#45) with the imaging half; coordinate with @amarzullo24 on the image/text split. Split any item into its own issue if you prefer; this tracker is the home.