Medication safety net

Available onmacOS · on-device drafting pathWindows

Medications are where a scribe error costs the most, and they fail in two distinct ways:

1. Speech recognition garbles the name. “Empagliflozin” is exactly the kind of word a general speech model mangles, and a drafting model can then smooth the garble into a real, plausible, wrong drug that reads as correct.
2. The drafting model “corrects” a dose. Language models will quietly replace a dictated value with the one they saw most in training: a dictated aspirin 300 mg (the Australian dose) comes back as 325 mg (the US dose).

The desktop apps run layered, on-device defences against both. None of this leaves your machine.

The layers

• Fidelity rules on every enhance. Every drafting request carries rules that forbid changing any dictated number, dose, unit, route, or date, and forbid stating a result that was never dictated. They are applied by the app itself, not by the prompt text, so they also cover prompts you write yourself. (Converting notation is still allowed: “one thousand two hundred and forty picograms per mil” may become “1240 pg/mL”.)
• Drug-name matching against a medicines list. Garbled names are matched against an Australian medicines list, so “empagliflozin” dictated through a bad microphone still comes out as empagliflozin.
• Dropped / invented check after every enhance. The drafted note’s medication names are compared against what you actually dictated, and any drug that looks invented or dropped is flagged. It runs with no added delay, entirely on-device.

Platform specifics

Windows

• The post-enhance medication check shipped in 0.3.2 and is on by default; turn it off under Settings → LLM (Quality check) if you need to.
• Fidelity rules ride on every enhance, cloud or on-device, since 0.3.0.

macOS

• The full net (name matching, dropped/invented check, uncertainty highlights) runs on the on-device drafting path, where the small local model needs it most. Where a name stays uncertain, the app highlights it and offers candidate picks rather than guessing.
• Fidelity rules ride on every enhance on every backend, and tell the model to keep unclear dictation marked unclear rather than guess a drug name.

A net, not a substitute

Current cloud models already pass most of these checks on their own; the net matters because it travels with the app rather than with any one model, and because the on-device model is smaller. It reduces risk, it does not remove it: always review medications and doses before a note enters the record.