Exact Dx

A reference build on one pattern: criteria + cite + click. Published clinical criteria run as explicit, versioned rules; a model only quotes the notes (every quote is verified) and drafts neutral wording; a named clinician decides every answer.

Apps

• Exact Dx

  Before you sign: checks the patient's own results (kidney function, and other measurements against a library of published definitions) in both directions, on clinician-typed and AI-drafted notes, shows measurements only (no diagnosis named), reports the net, and replays every check. Also: a coder view, the definition library with review sign-off, and the extended ten-condition criteria engine (experimental).

  Kind: worklist · Evaluation · Audit

How decisions are made here

1. Ingest synthetic tables and notes (or external CSVs through a field mapping).
2. Cite: the model returns note facts, each with a quoted source span. Spans that are not literally in the note are shown as "Not found", never guessed.
3. Criteria: plain, versioned rules (with their published sources) decide what the chart supports, with a full trace. Where a source needs judgement, the rule says so.
4. Worklist and query: a ranked queue (both directions: more specific, or not supported) and a compliant, non-leading query checked by a linter.
5. Human click: a CDI specialist sends; the clinician answers, and signs any addendum.
6. Audit: every model call and click is appended to a hash-chained log (verify it).
7. Metrics: observed numbers with confidence intervals; projected numbers are always labelled SIMULATED.

Not medical advice. Not an official product. Synthetic data only.