Problem
A reviewer needs to exercise AI-backed file workflows without paid credentials, model downloads, invented medical fields, or hidden provider behavior.
Contribution
- Separated API, orchestration, domain, and provider adapters; an automated architecture test enforces the boundaries.
- Validated file signatures instead of trusting extensions and preserved raw evidence for every parsed laboratory value.
- Returned explicit no-speech, non-lab, unparsed, and unknown-fixture states rather than hallucinating a successful result.
Case note 01
One contract, replaceable providers
The application layer depends on provider-neutral ports. Offline replay adapters make review deterministic; OpenAI transcription and Azure layout OCR are selected only through validated configuration.
Case note 02
Preserve uncertainty
Numeric values stay numeric, while comparators, intervals, raw OCR tokens, raw lines, and unparsed rows remain explicit. Documents without credible lab evidence return `not_lab_report`.
Case note 03
Known limits
The live providers are contract-tested but not benchmarked on a representative private dataset. Synthetic report fixtures test failure modes, not clinical performance.