The Lab · 02 · MIT licensed

trustlint

A deterministic linter for the trust properties of an AI answer. It flags the patterns that make a fluent answer untrustworthy: figures with no source, absolutes no evidence can support, hedging used as decoration, and the expensive one — asserting confidently when there was nothing to assert from.

It cannot tell you whether the answer is true. There is no model here and no retrieval — it reads text with regular expressions. That makes it a linter, not a benchmark: the same relationship ESLint has to correctness. A clean pass means the answer is shaped honestly, and nothing more. A tool that claimed to score truthfulness from regexes would be committing the exact sin it audits, so this one says its limit out loud and returns it in the payload.

Run it

Paste an answer, or try one of these

Everything happens in this tab. Nothing you type is sent anywhere, logged, or stored — the module has no network access by construction.

The answer

Citations are recognised as [1], [cap table p.4], (source: …) or a URL.

The lint pass

Waiting for text.

0Figures
Cited
0Sources

    The rules

    Seven checks, and why each one costs trust

      Where this came from

      The review I used to do by hand

      On the private-equity screening tool, every model output I reviewed got read for the same handful of failures — a number nobody could trace, a superlative with nothing behind it, a confident paragraph built on one thin document. That review was a habit living in my head, which meant it did not survive contact with a backlog.

      This is that habit as a function. It is the same move as trustlayer.js: take a judgement I keep making in review, and give it a failing state so it can run without me in the room.