The evidence behind the eForm Generator
We tested it the way you’d want a clinical tool tested: against real forms, with every experiment written down — including the ones that failed.
Compiled from the project’s own experiment records · Updated 2026-10-12
- AI page analyses across 10 models
- 1,998
- distinct real forms
- ≈400
- fields in the largest single study
- 14,724
- scored evaluation runs
- 895
- tests run
- 3,938
What we tried
122 documented experiments. Most ideas didn’t make the cut.
| Area | Tried |
|---|---|
| Prompt variants | 4 |
| Text-recognition layers | 5 |
| Geometry clean-up | 32 |
| Second pass over page strips | 9 |
| Field-to-chart linking rules | 37 |
| AI-assisted linking | 15 |
| Proprietary technological testing | 14 |
Outcomes
- adopted32
- rejected44
- stopped early9
- ongoing11
- records & corrections26
Results
Where accuracy stands today
| Metric | Now |
|---|---|
| Detection F1 | Over 84% |
| Checkbox recall | Over 97% |
| Text-field recall | Over 80% |
| Validation sample, F1 | 96.1% |
“The eForm Generator adapts over time, improving its results with each use.”
Time saved
30–60 minutes by hand. About 5 with the Generator.
At $10 a form it pays for itself after about 4 minutes of a physician’s time, or 20 minutes of staff time. Converting 50 forms saves roughly 33 hours.
Manual and end-to-end times are estimates; analysis time is measured at 21–40 seconds per page.