Could Not Confirm Review Outcome#
Context#
A reviewer may make a diligent attempt without obtaining enough evidence to accept, reject, or honestly escalate a candidate. Encoding that non-answer as rejection, verification, or disputed projection changes the meaning of the claim and contaminates confidence calibration with a label the reviewer did not provide.
Decision#
ReviewOutcome.resolution is an additive terminal-resolution field. Existing
outcomes may omit it and retain status-inference behavior. The
could_not_confirm resolution requires a non-empty resolutionReason, may
carry attemptEvidenceIds, and is illegal with verified or rejected status.
It records actor and review time, but preserves the claim's pre-review proposed
or assumed status.
All explicit resolutions are status-checked: accepted requires verified or
assumed, rejected requires rejected, and held permits only non-rejected retained
postures.
Could-not-confirm is terminal for the review round, not the candidate set. It
does not create escalation, product apply actions, or a new Surface posture.
Projection is equivalent to the unreviewed path; attempt evidence and review
timing stay out of the bundle, and { reviewProofs: true } deliberately omits a
Surface integrity anchor for this resolution. Canonical review proof v3 commits the resolution,
reason, and attempt ids while v1 and v2 proofs remain verifiable.
could-not-confirm deliberately bypasses producerPolicy.decisionMode because
that policy constrains candidate-selection actions and this resolution produces
none. Oversight metrics still count these decisions in decisionCount and
reviewer-behavior rates such as overrideRate; those metrics describe reviewer
activity, not correctness calibration.
Calibration excludes these outcomes from both numerator and denominator.
Survey learning projections emit learning.could-not-confirm so producers can
aggregate repeated unconfirmable targets without treating the signal as truth,
rejection, or comfort-zone posture.
Deferred work#
- Aggregation and threshold policy for repeated unconfirmables remains producer-owned.