Slopcheck

It is a tool that reports which AI tells a passage contains, with the line quoted and a probability, and it runs as a Claude Code Stop hook so it checks Claude's own output after every turn. It never claims to answer whether AI wrote a text.

One Jev call asks fifteen noul questions over the whole passage in parallel, and the questions cannot see each other's answers. Jev reports which tells are present and with what probability; the model is never asked whether a human should care, because that policy lives in code.

Document mode makes one call about the whole document and reports no line numbers. Paragraph mode makes one call per block, all at once, and returns line numbers because the position of each block is already known. The author measured nine drafts and says neither mode is a superset of the other.

Code owns every threshold. A noul between 0.40 and 0.60 means the model is not answering, so it is suppressed rather than read as a weak yes, and you can change a threshold in slopcheck/tells.yaml without re-running anything.

Findings are listed with layer, tell, probability and quoted line, followed by a suggested fix. The README's example shows 4 findings in 61 words: a banned word, importance puffery at p=0.94, a binary contrast at p=0.97 and a fake-profound kicker at p=0.90, in 1233 ms with 2 Jev calls for $0.000120.

The regex layer is free and takes about 6 ms. The Jev judge is listed at $0.105 per 1,000 checks and roughly 575 ms, and the README's example cost $0.000120 for two calls. Document mode medians at 604 ms across nine drafts, paragraph mode at 735 ms.

The README visible in the digest cuts off inside its tells section, so installation and the full tell list are not documented here. The mode comparison and medians come from nine unnamed drafts measured by the author, and four tells with scope document are asked about the whole piece in both modes.