Case Corpus v2: ingest 50 attested cases (DEI — Diversity, Equity & Inclusion)

[kenleefreo]

What this adds

The 8th tranche-1 source folder — Diversity, Equity & Inclusion (50 SOAP notes, DEI-001…050) — transformed to v2 telehealth-reprojected casebundles and ingested. Net-new coverage (a psychosocial / minority-stress / mental-health domain not in the base 303), so this is additive: the corpus grows 353 → 403. All 50 land in the MH (Psychiatry & Mental Health) bucket.

A different domain — the calibration clause was rewritten

Every DEI note has a psychosocial/mental-health primary diagnosis (adjustment / stress-and-trauma-related disorder, somatic symptom disorder, body-image / gender distress) but presents as an acute somatic/autonomic/panic crisis whose serious physical differentials (ACS, spontaneous pneumothorax, occult trauma, secondary headache / meningitis, acute abdomen, symptomatic severe hypertension, PE) the in-person clinician excluded only by physical examination Breath-Ezy cannot perform remotely.

So this is a telehealth-uplift + dismissive-under-triage ("medical gaslighting") domain, not a "reassure, it's just stress" domain. The scored prime error-of-omission is dismissive under-triage — premature psychosocial closure / failure to exclude the organic mimic — and the AI must validate the symptom as real AND escalate (under-triage weighted 3×). Benzo-as-AI-output and premature reassurance are errors-of-commission.

Calibration — one catch, corrected

The batch-1 sanity check caught a real over-triage: DEI-007 (Code-Switching Exhaustion) came back T5 despite entirely normal vitals + neuro exam, source Risk Level Low, and a routine 7-day GP disposition. Corrected to T3 (urgent in-person to cover the exam the AI can't do) and added a T4/T5-ceiling guardrail to the transformer prompt: "the clinician did an exam to be sure" ≠ a T4/T5 justification — reserve emergency tiers for patient-reportable time-critical features. Batches 2–4 ran under the guardrail with zero further over-triage.

Final tiers: T3×31 / T4×19 (zero T5 baseline, zero T1/T2). Zero T2 is honest — the folder is authored entirely as acute out-of-hours crises with an unexcludable organic mimic; the T3 floor reflects "AI cannot examine → urgent same-day in-person." Genuine emergencies (active SI with plan, IPV with imminent danger / weapon, deterioration) are conditional T5 / 000 escalation edges in node 13, not baseline.

Safety invariants held

• No dose or benzodiazepine synthesised as AI output — in-person diazepam / analgesia / beta-blocker doses sealed in node 12, should_NOT_recommend.
• No autonomous diagnosis.
• Scoring-store firewall intact — all 50 carry the canonical firewall_assertion (DEI-010/020 backfilled pre-ingest); ingest leak-scan passed; a de-anchor spot-check found 0 genuine answer-label leaks in AI-readable nodes.
• Legacy DSM-IV/ICD-9 V-codes not copied — mapped to plausible unverified ICD-10-AM candidates (F43.x / F45.x / F64.x / Z-codes).

Attestation + verification

All 50 attested by the clinical sign-off authority (reviewer KL, bulk_clinician_attestation, 2026-07-18). Statement records the reviewed-and-accepted items: DEI-007 T5→T3 + ceiling guardrail; DEI-028 ("Medical Gaslighting Fatigue" title is a red herring → suspected acute cholecystitis, T4); and the psychosocial-label-masks-physical-emergency cases (DEI-014 dehydration/AKI, DEI-041 symptomatic severe hypertension, DEI-017 medication-rationing decompensation → T4).

Gates (all green locally):

• cases:ingest --reseq → 50 dirs written (2 reseq'd for collisions vs the existing 353)
• cases:verify-codes → 484 codes receipted (mock_verified_pending_live_ncts), 0 failures
• eval:cases → PASS (401 attested conforming, ≥45 required; the 2 CARD reference cases are informational warnings, 0 failures)
• npm run verification → Pass: true
• npm test → exit 0 (176 checks)

Data-only change — firewall / hashing / schemas / servers / trunks untouched. Source .txt never entered the repo; only data/cases/ + the CHANGELOG entry are staged.

🤖 Generated with Claude Code