Score outputs based on their credibility and ability to be externally verified.
Scores how checkable an output is — whether claims can be independently confirmed, whether citations resolve and support, whether the confidence expressed matches the evidence available. Sits above groundedness by assessing verifiability rather than support.
Where a downstream reviewer must decide how much scrutiny to apply, and a triage signal is more useful than a binary pass.
Where every output is reviewed anyway. The score adds nothing if the scrutiny is constant.
The score must be calibrated against actual outcomes periodically, or it drifts into a number people trust without basis.
Uncalibrated confidence. A trust score of 0.9 that corresponds to 60% accuracy in practice is worse than no score, because it licenses reduced review.
A review queue ordered by verifiability score, so a reviewer's limited attention goes to the outputs least able to support themselves.