How the verifiers work

Four rules every one of them follows

No CAIN code. One file, Python standard library plus cryptography for Ed25519. You can read it in an afternoon. We checked the import lines of all three on 2026-11-02.
No network. They read files you give them and nothing else.
You pin the keys. Each takes a trust file (or anchors file) of public keys. The verdict is only as good as the channel you got those keys from, so get them from more than one place.
Re-derive, never copy. The verifier recomputes the verdict from signed records. The execution-truth verifier, for example, ignores the enforcer's own stated verdict (enforcer_stated) and derives its own from the approver-signed and sensor-signed records.

Verdicts and exit codes

Verdict	Meaning	Exit
VERIFIED	Every check holds against the pinned keys	0
FAILED	Evidence contradicts the claim, is not intact, or claims more than it shows	non-zero
INCOMPLETE	Nothing contradicts the claim, but required evidence is missing	non-zero
UNSUPPORTED	The evidence needs a method this verifier does not have, or an unknown schema	non-zero
UNAVAILABLE	Nothing to judge (for example the action was refused before it ran)	non-zero

Only VERIFIED exits 0. Anything unknown, missing or malformed ends non-zero: UNKNOWN never becomes a pass.

Each verifier says what it did not check

The mission-journal verifier prints checks_performed, checks_not_performed, checks_waived and does_not_establish. The Launchpad verifier labels the producer's paging window "not verified" in its own output. A verifier that hides its blind spots is a dashboard with extra steps.

FIRME AI CORPORATION · CAIN-42 is a Byzantine trust fabric. Numbers marked "our own check" were run by us, not by an independent party. Nothing on these pages is a certification.