Case Corpus v2: ingest 50 attested cases (GPO — General Practice Other)

[kenleefreo]

Summary

11th tranche-1 source folder — General Practice Other (50 SOAP notes, GPO-001…050) — transformed to v2 telehealth-reprojected casebundles and ingested. Net-new coverage, additive: corpus 508 → 558.

A bidirectional label-red-herring folder — inverted acuity

GPO is administrative / procedural / preventive-screening / medication / sexual-health / post-exposure GP work, so the centre of gravity is lower than the acute organ-system folders — but the folder deliberately authors a benign/administrative title that frequently masks a genuine acute entity in the note's own findings, mixed with genuinely benign encounters. Cases are tiered to the real entity, not the title. The central telehealth trap: the AI cannot perform the hands-on procedure or physical verification the encounter is built around → recognise + arrange the in-person pathway + safety-net, never mint the procedure/dose/vaccine/regimen/certificate/result.

Calibration — hypothesis corrected mid-folder, honest, no output re-tune

The batch-1 source read falsified an initial "T1–T2 dominant" read; the workflow framing was hardened to symmetric bidirectional (decision rule unchanged → no discontinuity) and every tier held across 14 source reads spanning all 4 batches.

Tier	Count	Representative
T1	2	well-child exanthem, fit-to-fly (resolved bronchitis)
T2	13	benign admin/preventive/screening (bowel/cervical/BreastScreen panic, Depo, med-review)
T3	31	telehealth "can't examine/perform/verify → urgent same-day in-person" ceiling
T4	3	failing-SABA asthma · warfarin bleed INR 7.5 · active suicidal ideation
T5	1	hypovolaemic shock (frail 85yo, BP 85/50)

• Masked-acute caught (under-triage weighted 3×): asthma "action plan"→failing-SABA exacerbation SpO2 94% (T4) · "INR monitoring"→active warfarin epistaxis INR 7.5 (T4) · "WorkCover form"→active SI (T4+T5/000) · "IV therapy"→hypovolaemic shock (T5) · plus polypharmacy toxicity / iatrogenic hypoglycaemia BGL 3.1 / migraine-with-aura-on-pill / unexplained syncope / decreased-fetal-movements (T3).
• Over-triage resisted both ways: panic/deadline/catastrophic framing kept genuinely benign encounters at T1–T2; Stage-III colon-ca referral renewal → T3 admin (not escalated), febrile-neutropenia 000-edge preserved.
• Fine window discrimination: open nPEP/PrEP/EC/AED-bridging → T3; closed (21-day) nPEP STI-screen → T2.

Safety invariants held

No antibiotic/analgesic/vaccine/contraceptive/PrEP/nPEP regimen/iron/B12/any drug dose minted as AI output, and no procedure instructed as an AI act — all sealed in node 12 (should_NOT_recommend/not_indicated_here). No paediatric/weight-based dose (2 paediatric cases flagged for in-person). No fabricated vital/BGL/INR/serology/β-hCG/FHR/imaging value — every confirmatory value (BGL 3.1, INR 7.5, FHR 145) sealed in 10/11; patient-obtainable readings ship as verbatim STRINGS in objective_data_offered (never parsed, never lab_result). Firewall intact — all 50 carry firewall_assertion (9 backfilled pre-ingest); ingest leak-scan clean; refined de-anchor scan over patient-facing 01/02 found 0 genuine dx-label leaks.

Gates

• cases:ingest --reseq → 50 dirs (1 reseq'd: SPEC-MH-03-00032→00970)
• cases:verify-codes → 345 codes receipted, 0 failures
• eval:cases → PASS (556 attested conforming; 2 CARD reference cases informational; 0 failures)
• 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