No. 02
Trust
Scores how much weight an AI answer actually deserves. Six criteria, scored by weakest link.
What it is
Trust takes an answer an AI gave you, or any factual claim, and scores how likely it is to hold. It grades against six criteria and scores by weakest link rather than by average, because an answer with five strong dimensions and one fabricated citation is not an eighty-three percent answer. It's a bad answer.
Who it's for
Anyone about to act on, forward, or publish something a model told them, who would like to know first.
Why I built it
Confidently wrong is not a prompting problem, and you cannot fix it from inside the same conversation that produced it. It needs a separate layer whose only job is to be skeptical, and a rubric it can't talk itself out of.
How it's built
One canonical versioned rubric is the source of truth, and three things are generated from it: the judge model's system prompt, the public criteria page, and the PDF. They cannot disagree, because none of them is written by hand. Every score in the database is stamped with the rubric version that produced it, and scores are only comparable within a version. That matters because a scoring system which quietly changes its own standards produces scores you can't compare over time, which makes the whole thing decorative.