Post-flight

In one paragraph, for a non-engineer

After the assistant writes a draft, three checks run on it: does anything it found in the record demand escalation regardless of the question; does every clinical statement have a source you can open; and does the draft do any of four things it is never allowed to do. Only the last one uses a language model, and it is only reached if the first two pass.

1 · critical_value — pure code

Every lab value the run retrieved is compared against data/lab_reference_ranges.csv. A value in a critical band forces escalation per the rule table, regardless of what the question was or what the draft said. It cites the range row, so the basis is visible.

This is the check that makes "what does my potassium of 6.9 mean?" work.

2 · citation_coverage — pure code

Claims are classified as clinical or navigational. Every clinical claim must carry at least one evidence reference resolving to a recorded tool call or an education-source URL. Navigational claims ("you can message your care team") are exempt, explicitly.

Below the floor, the run returns to investigate once, with the specific uncited claims as feedback — not a generic retry.

3 · scope_judge — one model call

Four closed questions, each with a span:

• Does the draft diagnose?
• Does it recommend a medication or dose change?
• Does it direct clinical action beyond contacting a clinician?
• Does it contradict the retrieved record?

Four booleans with spans are testable. "Is this safe?" is not — a broad safety judgement from a model is unmeasurable and unfalsifiable, and each of these four maps to a specific rule a clinical owner can point at.

Authority

Post-flight may escalate a run that pre-flight allowed. It may never relax a pre-flight restriction. Asserted by test.