Lab reports to named markers, transcribed and never interpreted

The problem

A lab report is the one input where a wrong digit or an invented flag flows silently into the plan and nobody notices until a client does. The default behavior of a capable model is the opposite of what you want here: it summarizes, it interprets, it helpfully flags what it thinks is high.

What we built

The model runs on the largest tier by choice. Volume is a handful of documents per intake, and a misread digit costs more than the call. The prompt frames the job as data entry: transcribe only, never infer a flag, never invent a marker, omit rather than guess, and leave out names and dates of birth.

The schema does the enforcing. Each marker has a name, a value stored as a string so results like "less than 0.1" or "1:40" survive, a unit, the reference range as printed, the panel, and a flag from a closed list of five or nothing at all. The collection date is the one on the report, not the upload date. A test asserts the schema has no field that could carry an interpretation, and the plan stages carry a matching rule: labs are described observationally, only the lab's own flags count, and a flag belongs only to the marker that carries it.

The parser is tolerant about packaging and strict about content: a payload that fails the schema is a failed extraction staff can retry, never a half-filled panel, and a model refusal is recorded as its own status. Extraction runs in the background after upload, the Generate button waits for it, and a failed document does not block the plan. The lab summary the plan stages read is built once, deterministically, and reused byte for byte on every regeneration.