Testing
Every credible test of AI output begins with a rule you could have written yourself. Tests built on a model's opinion inherit that model's blind spots, and no two runs agree on the number.
The work is the evidence that it held.
Deterministic checks are boring and that is their value: the same input produces the same findings today and in three years. When a rule fires, you can read the sentence behind it, decide whether you agree and open the line it flagged.
Model-judged grading cannot do that. It is useful for exploration, but you cannot attach it to a release decision, because a re-run may quietly reach a different conclusion.
Across generated code, copy, documents and media, a small number of checks catch a disproportionate share of the defects.
No automated check knows whether the work is useful, whether the strategy is sound or whether the phrasing will land badly in one market and well in another. A green result means the known defects are absent, nothing more.
Read the score that way and it stays honest. Treat it as approval and it becomes the reason a problem ships.
Credential literals
Keys, tokens and passwords assigned in source, in code or in an example that will be copied.
Injection surfaces
Strings handed to eval, new Function or innerHTML without a parser in between.
Truncated output
Work that stopped mid-structure and cannot be executed or read as written.
Claim density
Paragraphs of assertion with no citation, no measurement and no qualifier.
Lossy media
Re-encoded images and audio whose original signals no longer survive.
AI testing
The automated pass is free, needs no account and uploads nothing. Escalate to a named reviewer only where being wrong is expensive.
