The method
A standard you can read, argue with and re-run.
Most quality signals for AI output are opaque: a percentage, a green tick, a claim of 'AI-detected'. This page is the opposite of that — it is the whole method, in the order it runs.
Check the seam, not the summary of the seam.
Four principles
Offline by architecture
The checker is a bundle of plain functions. There is no model in the loop, so results are deterministic: the same input always produces the same report, and can be re-run years later to prove it.
One rule, one finding
A finding exists because one rule matched. We don't aggregate rules into a vague 'quality' signal you have to trust. Every deduction is traceable to a sentence you can read.
Decay, not cancellation
Deductions are weighted and then decayed, so a single systemic problem cannot bury the score at zero and hide ten smaller ones behind it.
A defined edge
When a judgement needs context, taste or a credential, the automated report stops and says so. That is the boundary the human reviewer is paid to cross.
The severity scale
Every finding lands on one of four levels, and the level sets how much it removes from a perfect hundred. The weights are published because a score you can't reconstruct is a score you shouldn't trust.
| Level | Weight | What it means |
|---|---|---|
| Critical | −22 | Ships a defect that can harm someone: leaked credentials, injection paths, a claim that is factually dangerous. |
| High | −12 | Will embarrass you or break something obvious the first time a real person uses it. |
| Medium | −6 | Degrades quality or trust without breaking anything — unsourced numbers, unescaped HTML, missing alt text. |
| Low | −2.5 | Style and polish. Worth fixing, not worth blocking a release over. |
Where the machine stops
Automated checks are good at rules and bad at meaning. A checker can tell you that a statistic has no source; it cannot tell you whether the statistic is being used fairly. It can tell you an image has no provenance metadata; it cannot tell you whether the subject consented.
So the console stops at that line and hands over. A named reviewer picks up the same file, reads it in context, and writes a sign-off that states what they checked and what they excluded.
Browse the reviewer networkVersioning
The rule set carries a version number. A report states which version produced it, so a finding from six months ago can be reproduced exactly — even after a rule has been tightened, split or retired.
Rule set v1.4
See it run
The fastest way to understand the standard is to trip it.
Load one of the sample files in the console. They are written to fail in specific, instructive ways.