Post-flight override triggers

Like every chart in this project, this one is generated by make docs-assets and regenerating it produces byte-identical output, so it cannot quietly drift from the code it describes.

What this shows

Post-flight runs after the draft is written, and it can only make a decision more restrictive — never less. The chart counts, by trigger, how often each of the three checks moved a draft off the publish path in a fixed demonstration battery:

• Critical value — a retrieved lab in a cited critical band forces escalation regardless of the question. This bar is drawn from real data: the one injected potassium of 6.9 mmol/L in the committed fixture (canonical case 4).
• Uncited claim — a clinical statement with no openable source. With retries exhausted the draft is held for review rather than published.
• Diagnosis and Medication change — the two scope-judge triggers that a patient-facing assistant must never do on its own. Each blocks the draft and cites the offending span.

Why the other three bars are a demonstration, not a rate

Only the critical-value bar is a measured property of the population. The other three are run through the real post-flight node against a small, committed battery of crafted drafts — enough to prove each trigger fires and routes correctly, but deliberately not presented as a base rate. A trigger rate over real traffic is a Layer 2 question (evaluation), not populated yet — a count is never shown without the sample it came from.

The honest read of this chart is therefore mechanism, not frequency: every override path exists, is reachable, and lands on the right disposition.