An evidence-to-action ER nurse assistant simulation
🔗 Live: https://baymax-bice.vercel.app
Patient said: “My mother’s legs are swollen.” Baymax retrieved the heart-failure precedent nobody mentioned.
Baymax is not impressive because it has more data. It is impressive when another perspective changes what it decides to do.
👁 Patient world 👁 Drug world
"What is similar?" "What is hidden?"
\ /
└──🧠 Brain──┘
👃 Nose ──┤ changes action when perspective changes
↓
🛑 Brakes → ⚡ Nerves → 🤝 Hands → 🛡 Verify
make sync
make audit
make uiOpen http://localhost:8000/ui/. The live surface reads the generated audit
receipt, narrates what Baymax noticed, admits uncertainty, shows what acted or
stopped, and only claims verification when the receipt proves it.
CASE 1 · RETRIEVAL DISCOVERY
"My mother's legs are swollen"
→ raw words miss the target
→ evidence-language query retrieves a CHF precedent
→ Baymax surfaces unmentioned evidence before answering
→ next question: short of breath at rest or when lying flat?
CASE 2 · ATTENTION FLIP
routine medication refill
→ served NOSE assigns ESI-attention tier 4
→ expensive eyes stay closed
→ no action attempted
CASE 3 · DECISION FLIP
same critical chest-pain patient
→ bed available: assign_bed
→ ER gridlock: divert
→ capacity perspective changes the action
CASE 4 · BRAKE SAVE
abdominal pain after Ibuprofen
→ patient-only: discharge_plan
→ drug eye: 16/17 exact-Ibuprofen reports marked serious
→ population signal only; no causality claim
→ brake blocks autonomous discharge
→ clinician-review nerve receives and ACKs the case
That is the recruiter path. Four cases show that Baymax discovers hidden evidence, allocates attention, changes decisions when another perspective matters, and stops itself before an unsafe autonomous action.
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
make test
make auditThe final receipt is outputs/baymax_audit.json. It contains the discovery receipt and action trajectories:
- attention skip: NOSE stops a routine case before expensive evidence work
- cross-domain brake: another perspective changes action into human review
- bounded action: durable state changes and the outcome is independently re-read
It also records exact source commits, organ grades, the solo-engineer A+ ceiling, and the remaining shortest path.
Baymax includes a deliberately thin FDE-adjacent rehearsal package:
make readinessIt verifies five acceptance gates against the live audit receipt, then injects a false-success receipt and proves the release contract detects it and requires rollback. The result is eligible only for simulated shadow, never presented as hospital deployment or adoption.
outputs/baymax_audit.json ends with a honesty_ledger: every organ maps to one
capability lane, may only claim the one sentence its live proof earns, and Baymax
ships at the strength of its weakest load-bearing organ — never above it.
| Organ | Capability lane | Verdict gate |
|---|---|---|
| NOSE | signal-routing | served patient signal; ≥95% serious recall gate |
| Left eye | data-truth | 55,500 rows scanned this run |
| Right eye | evidence-retrieval | 5,000 openFDA reports scanned; population signal, not causality |
| Brain / Brakes / Hands | action-engine | decision change, autonomous block, and durable outcome re-read |
| Nerves | clinical-handoff | receiver ACK proven; cross-service recovery not yet |
| Mouth | explanation | states the FAERS causality boundary |
A live gate that fails drops the organ to ❌, and ❌ leaves are never claimed.
The headline verdict is the minimum verdict across load-bearing organs, pinned in
CI by tests/test_audit.py::test_honesty_ledger_gates_headline_at_weakest_load_bearing_leaf.
| Organ | Public source | What this repo audits |
|---|---|---|
| NOSE | healthcare-genai-engineer + healthcare-signal-platform | Served patient attention signal plus a separate five-signal openFDA batch proof |
| Left eye | healthcare-ai-data-engineer | Full 55,500-row synthetic encounter corpus and reconciliation |
| Right eye | healthcare-da + healthcare-signal-platform | Real openFDA evidence plus governed semantic-layer lineage |
| Brain + hands | healthcare-genai-engineer | Triage, Bed Ops decision, durable action, ACK, and outcome verification |
Baymax is a closed-loop ER simulation, not a deployed clinical system. The action engine changes durable SQLite state representing a Bed Ops disposition; it does not write to a hospital EHR or bed-management platform.
The per-case NOSE is deliberately cheap: a versioned deterministic ESI attention signal decides whether to open both eyes. On 497 labelled synthetic patient cases it preserves 95.39% serious-case recall while skipping 4.83% of expensive paths. The five statistical/ML/retrieval signals are separately batch-evaluated on 5,000 real openFDA reports; this repo does not pretend those models score patient queries online.
baymax/
├── baymax/demo.py ✅ plays the three memorable cases
├── baymax/audit.py ✅ generates deep evidence receipts
├── baymax/served_nose.py ✅ serves and evaluates pre-eye attention
├── deployment-readiness/ ✅ simulated acceptance, rollback, and handoff
├── cases/legendary_cases.json ✅ the public three-case screenplay
├── docs/baymax/constitution/ 📖 versioned architecture health records
├── tests/test_audit.py ✅ pins every movie outcome in CI
├── scripts/sync_sources.sh ✅ fetches the four public sibling sources
├── outputs/ 🟡 audit + simulated rollout receipts
├── .github/workflows/audit.yml ✅ regenerates proof on every change
├── Makefile ✅ one-command audit entry points
├── SPEC.md 📖 scope and proof boundary
└── README.md 📖 recruiter audit map
- retrieval discovery: symptom-language expansion retrieves an unmentioned CHF precedent