Honest when it doesn't know
My contributionObserved outcomeCapability: Recover
If the notes don't cover a section, the clinician must see that before signing.
What was difficult
Models fill gaps confidently, and a confident wrong answer is the fastest way to lose a clinician.
Evidence at the time
Audit of 60 summaries: 1 in 3 had a section copied from an earlier stay. Prototype tests, 2 rounds of 7.
Alternatives considered
- Let the model fill every section: looks complete, but invents content.
Selected direction and trade-off
Empty sections say "Not found in the selected notes" and offer Add text or Choose more notes. Sending to ICIS lists every gap and needs an author tick.
Trade-off. Drafts look less complete, on purpose.
Who did what. To confirm What I did, who set or approved the rule, and what engineering built.
Accessibility
Empty sections carry a label, a line of explanation and two actions on a white box, so they never read as filled fields and status is not colour alone. These are design checks.
Result and status
Observed outcome
72% of pilot summaries were edited before sending (tool logs). An edit shows engagement, not a full review.
Open question: the effect on time to discharge.
The evidence
See the low-fi and final screens side by side, with what they do and do not prove