Early on we built a confidence score for every generated note. A percentage sitting beside each section telling the physician how certain Diagnose was about what it had written. It tested well in demos. Physicians nodded when we showed it to them.
Then we watched them use it in clinic.
The first thing they did was look at the score before reading the note. If it was 94% they relaxed. If it was 87% they read every word twice even when the note was perfectly accurate. We had introduced a number that changed how physicians trusted their own clinical judgment. That was not a feature. That was noise.
We removed it in week six.
The pattern we didnt expect
That became a recurring theme in year one. Not building more but learning what to take away. We removed a session timer that made physicians feel surveilled. We removed a suggested edits panel that turned note review into a negotiation. We removed a daily summary dashboard that no one opened after day three.
Every removal made Diagnose quieter. Every removal made it more useful. It is the same instinct behind what we changed in the note review screen — a process that took longer than it should have because addition is always easier to justify than subtraction.
What we chose not to build
There is a version of this product that does more. More flags more analytics more surfaces. We chose not to build it. The physicians using Diagnose are already managing more information than any person should in a single day. The job of the tool is to reduce that load not add to it.
This is also why specialty specific models matter more to us than feature count. Depth over breadth. Accuracy over coverage. A tool that does one thing without asking anything of you is harder to build than a tool that does everything badly.
We are still learning what to leave out. We probably always will be.
