Fail-closed verification is a checker built so a broken check can never count as a pass. If the verifier errors, times out, or returns something it cannot read, the safe default is fail, not succeed. A claim is treated as unsupported until it is actually shown to hold.
Why it matters
Most systems fail open: when a check breaks, the thing under it slips through. For a lock or a fact-checker that is exactly backwards. A verifier that passes a claim because it crashed is worse than no verifier, because it looks like a green light.
The failure mode is quiet. A judge returns malformed output, the code cannot parse it, and a naive system shrugs and counts it as fine. The unverified claim ships wearing a checkmark it never earned.
Fail-closed flips the default. Anything the checker cannot positively confirm is treated as not confirmed, so an error can only ever make the gate stricter, never looser.
How Jinn treats it
In Jinn's research verification every claim is judged against the actual source text it cites, and the verdict is parsed strictly. If that verdict is unreadable or the judge errored, the claim is recorded as unsupported with the lowest possible score, never passed by accident.
The report is then gated on how many claims are genuinely supported, and on regulated ground a claim that is not verified is dropped entirely rather than softened and kept. The conservative default runs all the way through.
Related terms
A broken check should fail, not pass.
Anything the verifier cannot confirm is treated as unconfirmed. See how Jinn gates a claim before it ships.