The proof engine
Don't trust the number. Reproduce it.
A CAI is a claim, and a claim is only as good as its reproducibility. Because an evidence bundle carries the measured dimensions (and a survey states its rubric version), anyone can recompute the headline and check it against the published number — with no access to the analyzer.
How to verify a published survey
- Get the evidence bundle. A published survey links its evidence and states the rubric version it was scored under (e.g. rubric-2026.08.15).
- Recompute. Fold its dimensions through the open scorer — the
calculator does exactly this, and when the bundle carries a published
headlineScoreit shows the reproduction verdict inline (✓ / ✗). Or run the reference CLI:cai verify survey-evidence.json. - Compare. Same dimensions + same rubric ⇒ the same number. A mismatch is falsifiable proof the published number doesn't follow from the evidence.
Signed, so attribution is checkable too
Beyond reproducing the number, a delivered survey is cryptographically signed by
its issuer — so you can confirm who attested it and that nobody has edited it since. The signature
covers the payload's canonical form; the public keys are published at
/api/registry/keys. Paste a signed package below and check it
here — no account, no tooling, nothing installed.
It doesn't reproduce — now what?
A mismatch isn't the end of the conversation, it's the start of one. What you do next depends on which part you think is wrong — the arithmetic, the open scorer, the measurement of your code, or the rubric itself. Each has a documented, deliberately separate review path, and none of them asks you to trust an authority: a reproduction case is self-contained proof.