Case Corpus v2: ingest 55 attested cases (GIH — Gastroenterology & Hepatology)

[kenleefreo]

Summary

10th tranche-1 source folder — Gastroenterology & Hepatology (55 SOAP notes, GIH-001…055) — transformed to v2 telehealth-reprojected casebundles and ingested. Net-new coverage (not a base-303 refresh); corpus grows 453 → 508. Predominantly GI, with HEPAT (opaque id slot) for liver/biliary.

Bidirectional GI/hepatology domain authored so every condition presents via an acute complication at an out-of-hours home assessment — where the clinician excludes a surgical/vascular mimic by an abdominal exam, bloods, endoscopy or imaging a telehealth AI cannot perform. Tiers match the complication, not the chronic label.

Calibration — honest, no re-tune

Batch-1 sanity check + per-batch source reads across all 4 batches confirmed the high-acuity weighting is honest (forcing lower would under-triage genuine emergencies, weighted 3×).

• Final tiers T2×7 / T3×16 / T4×17 / T5×15.
• Downward discrimination confirmed: IBS×2, functional bloating, lactose, NAFLD/Non-AFLD, anal fissure → T2 (several min-viable T1).
• T4/T5-ceiling guardrail held — the can't-examine-abdomen uplift ceilings at T3 unless the patient can self-report a time-critical feature. GORD nocturnal-chest-pain → T3 is over-triage resistance (can't ECG/troponin remotely), not a defect.
• Label-red-herrings tiered to true acuity: "Diverticular disease"→lower-GI haemorrhage T5, "H. pylori"→bleeding peptic ulcer T5, "Barrett's"→food-bolus impaction T4, "PSC"→septic ascending cholangitis T5, "Inguinal hernia"→incarcerated/strangulating T5, "Hep C"→decompensated cirrhosis T4; two PUD notes discriminate (050 bland flare→T3 vs 051 bleeding→T5).

Safety invariants

No AI-minted analgesic/antiemetic/antispasmodic/PPI/antibiotic/laxative dose (clinician answers sealed in node 12); no paediatric dose; no fabricated lab/lipase/LFT/endoscopy/imaging values (patient home readings ship as verbatim strings, never lab_result). All 55 carry firewall_assertion; dry-run leak-scan clean; refined de-anchor scan 0 genuine dx-label leaks.

Attestation + gates

All 55 attested (KL, bulk_clinician_attestation, 2026-07-18). Recorded borderline: GIH-007 Celiac→T4 (completed syncope) vs GIH-012 Coeliac→T3 (near-syncope only).

• cases:ingest --reseq → 55 dirs (8 reseq'd for collisions vs 453)
• cases:verify-codes → 488 codes receipted, 0 failures
• eval:cases → PASS (506 attested conforming; 2 CARD reference cases 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