Meta-Verification
Semantic judges, monitors, and calibration revisions can themselves be wrong. v0.0.8 adds a meta-verification layer that treats them as fallible evidence sources rather than roots of trust.
What a monitor's output carries
Every retained monitor output carries its full identity and provenance — model/endpoint, prompt version, calibration revision, and when it was called — and is represented as ordinary derived evidence, not as an authoritative override. A disabled, stale, or out-of-scope monitor's output is never silently counted as either supporting or contradicting evidence.
The one rule that never bends
A confident monitor cannot force a verdict past a real, objective
contradiction. This is tested directly: an overconfident, factually
wrong monitor claims a test suite passed, while a real host-observed exit
code says it failed. The fused verdict never reaches Verified — it stays
at Review, and the contradiction remains visible. The monitor's
confidence never overwrites it.
Same-family monitors don't manufacture independence
If two monitors observe the same underlying agent turn, they are grouped as one dependent source, not counted as two independent confirmations — reusing the same source-independence machinery used across the rest of the trust kernel.
Honest performance reporting
A monitor's contextual performance is only reported when there's enough
sample support — below that threshold, the report says so explicitly
(InsufficientSupport) rather than emitting a number. No global
"monitor X is N% trustworthy" claim exists anywhere in this release;
there is no real human-adjudicated dataset yet to support one.