Order resolution PoC

Help reviewers find missing or conflicting order information, inspect its evidence, and record a decision. AI proposes; code checks; humans decide.

Synthetic data only. No clinical decisions, messages, or order updates are executed.

Try the demo

Use Docker on Linux, macOS, or Windows. No AWS account or saved run files needed.

docker build -t order-resolution-poc:local .
docker run --rm --name order-review-demo \
  -p 127.0.0.1:3000:3000 --read-only --tmpfs /tmp \
  --mount type=volume,source=order-review-demo-data,target=/app/runs \
  --cap-drop ALL --security-opt no-new-privileges \
  order-resolution-poc:local

Open http://localhost:3000. The demo seeds five prepared cases; reviews persist in the named volume. Reanalyze replays prepared responses and makes no cloud calls.

On Linux with Node 26+, you can also run:

npm ci
npm run build:ui
npm run serve -- --demo

• The UI worker uses Linux process locks; use Docker on macOS or Windows.
• Five-minute walkthrough: missing fields, conflicts, quarantine, drafts, and review history.
• Live synthetic demo: ask the project owner for demo credentials. You do not need AWS access.
• Deployment, sign-in, and recovery.

What it demonstrates

• Exact source quotations, grouped evidence, dates, and quarantined attachments.
• Runtime schemas, candidate-value checks, deterministic status rules, and human approval.
• Edited drafts and decisions saved separately against immutable source snapshots.
• Cognito sign-in, viewer/reviewer roles, and explicit case assignments.
• Durable jobs, isolated workers, cancellation, usage metrics, backups, and restore tooling.
• A TypeScript core with a replaceable model adapter and container deployment.

Evaluation

Dataset	Live AI	Rules baseline	Meaning
Five development cases	5/5	5/5	Development fixtures
Five narrative regression cases, latest repeat	3/5	0/5	Used during prompt development
Two frozen final cases, first pass	2/2	0/2	Small fresh sample

• These results do not establish accuracy at scale or measured business improvement.
• Narrative failures include a missed supplied field and a rejected model response.
• Human assessment of the new outputs and draft usefulness remains pending.
• The baseline handles labeled text; it is not a general narrative parser.
• Saved synthetic evidence includes successes and failures, with portable paths.

npm run verify:records
npm run evaluate -- --playback
npm run evaluate -- --dataset heldout --playback

Verification rechecks saved live outputs with current validation; it makes no model calls. Playback checks prepared responses and the harness, not model quality.

Check the project

npm run check
npm test
npm run build:ui
npx playwright install --with-deps chromium
npm run test:ui
npm audit --audit-level=high

Backend and browser tests cover evidence integrity, roles, review persistence, cancellation, and unsaved edits. CI also builds and scans the container.

Use the CLIs

npm run load-case -- --case fixtures/patients/patient-001
npm run analyze-case -- --case fixtures/patients/patient-001 --proposal fixtures/proposals/complete-case.json
npm run evaluate -- --playback
npm run ui

• Each command supports named flags and --help.
• ui opens a saved evaluation; use serve -- --demo for the self-contained demo.
• To preview one saved result, pass the path printed by analyze-case to npm run review-case -- --result <path>.
• CLI runs use a local lock. After a crash, confirm the worker stopped before removing runs/.case-run.lock.
• Run artifacts are private and Git-ignored; published synthetic records are a separate curated copy.

Live analysis

AWS_PROFILE=your-profile npm run analyze-case -- \
  --case fixtures/patients/patient-001 --model YOUR_MODEL_ID --region YOUR_REGION
npm run evaluate -- --profile YOUR_PROFILE --region YOUR_REGION --model YOUR_MODEL_ID

• Paid calls require your AWS login and an accessible Converse-compatible model.
• One call per case, no automatic retries, a 60-second timeout, and 4,096 output tokens.
• Cancellation stops local startup/work; a remote call already sent may still finish.
• Use --dataset heldout for regression cases or --dataset final for the final set. Final playback responses are not included.
• Cost remains unpriced until a verified pricing basis is configured.

Limits

• Quote/value matching does not establish semantic correctness or source authenticity.
• Human review remains required, including for apparently complete cases.
• One host and one active worker; multiple hosts and another cloud are untested.
• OCR/images, real patient data, external messages, and system updates are outside scope.
• No HIPAA-compliance or production-readiness claim.
• Contracts and rules, analyzer, security, and status.