Verda drafts clinical notes in real time while a physician speaks with a patient. Getting it into the hands of 600 clinicians across 38 clinics took eight months, one major architectural reversal, and a UX decision that mattered more than any model choice.
Where the models actually differed
We evaluated against a set of 400 anonymised visit transcripts scored by clinicians. Claude consistently produced safer clinical language — it hedged where hedging was correct — while other models were more fluent but overconfident in exactly the places a chart must not be. For medical documentation, calibrated uncertainty is a feature, not a weakness.
Prompts were the wrong abstraction
Our first architecture encoded specialty-specific charting rules in ever-growing prompts. It collapsed under its own weight around specialty number five. The rewrite moved structure into typed templates the model fills section by section, with per-section validation. Smaller prompts, testable pieces, and a chart that fails visibly rather than plausibly.
The decision that doubled adoption
Draft, never send. The assistant proposes; the clinician disposes — every sentence remains editable, and nothing enters the record without explicit sign-off. Adoption in pilot clinics roughly doubled after we made the human authorship visually obvious. Clinicians did not want an autopilot; they wanted a fast first draft they could own.