Sniff Test: a prose linter where Jev is the judge.

Sniff Test: a prose linter where Jev is the judge. It reads a draft one paragraph at a time and asks ten Boolean questions (is the claim hedged three times, does the closer restate the paragraph, is there a cost figure with no price beside it) and prints one line per flag with the probability. Runs as a CLI, pre-commit hook, GitHub Action and Claude Code skill. Measured on 17 Sept: 182 ms median, 63 of 80 planted faults caught, 1 of 54 clean paragraphs flagged, $0.013 per 100 paragraphs. Same questions on Haiku 4.5 flagged 37 of 54 clean. Full bench and raw replies in the repo. One thing I'd love a TypeSafe eye on: in the bench's own rotation Jev returned a reply with one rule missing on 13 of 166 paragraphs, and the eval path never did. Written up in docs/eval-notes.md. https://github.com/DanRWilloughby/snifftest