Why we make the model say "I don't know"
Bosshh · Jul 25, 2026
A confident, complete-looking assessment generated from a thin transcript is worse than an obviously incomplete one. Here is how we built around that.
The tempting way to build an AI charting tool is to make it produce a finished-looking document every time. Every field populated, every item scored, nothing left blank. It demos beautifully.
It is also the version that gets someone hurt.
A clinical record is read later by people who were not in the room — a reviewer, a coder, the next clinician, sometimes an auditor. Those readers cannot tell the difference between a score the model derived from an explicit observation and a score it produced because the field wanted filling. Both look identical on the page. So a system that fills gaps with plausible content is not producing a slightly imperfect record; it is producing a record whose reliability cannot be assessed at all.
We built the opposite behaviour into the prompt, the schema, and the interface, in that order.
The prompt tells the model to ground every statement in the source material and to record absence as absence. It says plainly that a plausible guess in a clinical record is worse than an acknowledged gap. That instruction does real work, but on its own it is a suggestion.
The schema makes it structural. Scored items are a separate array from gaps, and each scored item requires an evidence field — the phrase from the transcript that supports it — plus a confidence level. There is no way to emit a score without also emitting what it rests on. Vitals are individually nullable, and the field descriptions say to leave them null unless the value was actually stated.
The interface makes it visible. Every score displays its evidence quote underneath. Gaps get their own panel, sorted by consequence, and the ones marked blocking prevent the chart from being signed at all. A clinician reviewing a draft is not being asked to trust it; they are being shown what it heard and where it came up short.
The result is a draft that sometimes looks unfinished. That is the point. An unfinished draft with four flagged gaps tells the clinician exactly where to spend their attention. A finished-looking draft tells them nothing, and quietly asks them to sign for content they have no way to check.