A 4B clinical scribe plus a fixed verifier keeps three quarters of notes on-device and matches frontier acceptance
Synopsis
Clinical dialogue-to-note generation is high-volume and intolerant of fabrication; a 4B-parameter model inside a fixed four-evaluator verification stack comes close to a frontier model on factual entailment, with its deficit concentrated in inter-section structure. Treating verifier rejection as an escalation signal reaches 99.6% end-to-end acceptance against 98.4% for frontier-only, statistically indistinguishable, while three quarters of encounters need no frontier generation call.
Interpretation
With the verification stack and refinement policy held fixed and only the generator swapped, notes from a 4B-parameter model approach the frontier model on factual grounding, and the remaining gap sits mainly in how sections are arranged. The prior arrangement sent every encounter to a third-party frontier API, and a small model used alone could not separate its failures from its successes. On the synthetic dialogue test split, the small model scores 0.984 on clinical factual accuracy against 0.992 for the frontier model and 0.989 for reference notes; structural fidelity is 0.778 against 0.843 and 0.873.
The verifier turns the capability gap into a per-note admissibility decision, escalating rejected notes to the frontier model while kept notes carry an explicit guarantee. Run alone, the small model is accepted on 75.2% of encounters against 98.4% for the frontier model. End-to-end routing accepts 249 of 250 encounters (99.6%) against 246 (98.4%) for frontier-only, with three discordant encounters all favoring routing and an exact McNemar test not significant; only 62 encounters go to the frontier model.
Refinement is scoped to sections, rewriting only those carrying outstanding feedback while every other section stays byte-identical by construction. Whole-note rewriting let the 4B model introduce structural violations in one section while satisfying another's completeness, including copying Plan content verbatim into Assessment. Without the refinement loop the small model is accepted on 26.8% of encounters against 74.8% for the frontier model; restoring it lifts them to 75.2% and 98.4%, with the frontier model needing 1.28 rounds on average and the small model 2.07.
Perspective
The findings rest on a synthetic dialogue-to-SOAP corpus with a 250-encounter test split; the small model is Gemma-3-4B-IT with a QLoRA adapter generating on one Apple Silicon machine with 64 GB unified memory and no discrete GPU, and the frontier baseline is GPT-4.1. Verification calls go to a hosted judge in both conditions, so transcripts still leave the machine for scoring.
Acceptance certifies conformance to this rubric-based verifier, not clinician-adjudicated correctness; the reference-note calibration and the kept-versus-escalated separation are indirect checks on that verifier. Results come from a synthetic corpus and need replication on real encounters with clinician-authored notes, and whether other small models reach a comparable share of the workload, or how the approach behaves on less structured tasks, remains open.