Small, calibrated language models for clinical classification. The model reads, the code decides, and anything it isn't sure about goes to a person.
A 0.8B reader with rules code and a pre-specified hand-off automated 82.3% of 10,000 synthetic test charts, from authors whose wording was not used in development, while meeting its pre-specified error bounds. All patient data are synthetic.
The same 0.8B model at three stages (off the shelf, trained, and with a confidence hand-off) playing side by side, so every move can be checked against an expert. No patient data.