Case Corpus v2: ingest 50 attested cases (EMD — Endocrine & Metabolic Disorders)

[kenleefreo]

What this adds

The 9th tranche-1 source folder — Endocrine & Metabolic Disorders (50 SOAP notes, EMD-001…050) — transformed to v2 telehealth-reprojected casebundles and ingested. Net-new coverage, additive: the corpus grows 403 → 453. Predominantly the ENDO specialty bucket.

A bidirectional folder, authored as acute complications

EMD spans the full acuity range, but every condition is authored to present via an acute complication / decompensation, so tiers are matched to the complication, not the chronic label. EMD-specific bidirectional calibration clause (endocrine emergencies → T5+000; mass-effect / unexcludable-mimic → T3–T4 recognise+refer; chronic-with-acute-complication → tier the complication; genuinely-mild metabolic → T1–T2), carrying forward the domain-general T4/T5-ceiling guardrail.

Telehealth lab-dependence is central: most endocrine dx are lab/imaging-confirmed, so the AI recognises the pattern + arranges labs/refers + escalates — never fabricating a lab value. A patient's bedside glucometer/ketone reading ships as a verbatim string in objective_data_offered (never parsed, never lab_result).

Calibration — honest, no re-tune (the DCD pattern)

The batch-1 sanity check + source reads confirmed the folder deliberately authors even benign labels as acute complications (dyslipidaemia→acute pancreatitis, empty-sella→CSF leak with 000 activated, gestational-DM→reduced fetal movements, T2DM→infected foot ulcer), so the high-acuity weighting is honest — forcing lower tiers would under-triage genuine emergencies (weighted 3×).

Downward discrimination confirmed: pre-diabetes (EMD-038) correctly floored at T2 (genuinely mild source: candidal intertrigo, Risk Low, routine 3–5 day GP), and the 12 T3s sit at the telehealth "AI cannot examine / cannot obtain bloods → urgent same-day in-person" floor.

Final tiers: T2×1 / T3×12 / T4×13 / T5×24. Every high tier carries a patient-reportable time-critical feature, so the ceiling guardrail held. Emergencies landed at T5 (DKA, HHS, myxoedema coma, adrenal/Addisonian crisis, thyroid storm, phaeochromocytoma, hyper-/hypocalcaemic crises, CAH neonatal salt-wasting, MEN/gastrinoma); mass-effect at T3–T4 (acromegaly/prolactinoma chiasm compression, thyroid cancer airway compression, hyperparathyroidism→renal colic, haemochromatosis→new AF).

Safety invariants held

• No insulin / thyroxine / hydrocortisone / antihypertensive / oral-hypoglycaemic dose minted as AI output — sealed as the clinician answer in node 12 (should_NOT_recommend).
• No paediatric / weight-based dose (CAH, paediatric diabetes).
• No autonomous diagnosis; no fabricated lab/hormone/electrolyte values.
• Scoring-store firewall intact — all 50 carry firewall_assertion (EMD-013/020/023/031 backfilled pre-ingest); leak-scan + de-anchor spot-check clean (0 dx-label leaks in AI-readable nodes).

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 borderlines: EMD-038 pre-diabetes→T2 (genuine mild), EMD-036 PCOS→T4 (red-herring label masking unexcludable ectopic/torsion/haemorrhage — no bedside β-hCG remotely), EMD-011 empty-sella→T5 (note itself activated 000 for a CSF leak), EMD-012 gestational-DM→T4 (reduced fetal movements the AI can't auscultate).

Gates (all green locally):

• cases:ingest --reseq → 50 dirs written (1 reseq'd vs the existing 403)
• cases:verify-codes → 486 codes receipted, 0 failures
• eval:cases → PASS (451 attested conforming; the 2 CARD reference cases are informational)
• npm run verification → Pass: true
• npm test → exit 0

Data-only change — firewall / hashing / schemas / servers / trunks untouched. Source .txt never entered the repo.

🤖 Generated with Claude Code