An abstention-gate bench: striking unlicensed actor-attributions in language-model clinical transcription

A deterministic abstention gate that reads which agent a clinical note actually licenses for an action (pure regex, no model) and, on that basis, strikes an unlicensed actor-attribution or keeps a licensed one. Built by gating recorded responses from a local model (llama3.1:8b, 240 source calls) across matched note families where the actor is either stated (named, low salience) or absent (deleted, action kept). When the actor is absent the model supplied an unlicensed actor on 69.2% of calls (83/120) and abstained on 25%; when the same actor is stated it read it faithfully (0% unlicensed fill, 120/120 licensed). The fabrication rate tracks how the actor is grammatically buried (100% in a subordinate clause or a passive-with-agent, down to 20% in a parenthetical). The gate struck all 83 unlicensed fills and kept all 120 licensed reads, deterministically. The failure is not misreading what is present, it is completing what is absent, and that is the case the deterministic layer is built for. The gate here is a small purpose-built pure-regex probe; the general (In)Canon engine of record (frame_bind.ofm, OFM-PBL, UK patent application GB2606072.3) is not part of this deposit. See NOTE.md and DEPOSIT_README.md.