The claim
A human in the loop is not the same as human oversight. Oversight is a measurable practice: escalation that fires, reviewers with authority, decisions that can be overridden, and a record that an auditor can read.
When those practices are not measured, organizations report oversight they do not have. The model passes its governance review. The humans above it are assumed to be working. The assumption is the risk.
Judgment decay
Under sustained automation, the rate at which a reviewer exercises independent judgment falls. Approvals become reflexive. The loop stays nominally human while the human contribution trends toward zero. The framework calls this judgment decay.
The curve is the secondary motif of the system. A reviewer that started well above the threshold crosses it without any single event to flag. Point in time attestation does not catch this. Continuous measurement does.
What COHESION measures
The specification turns oversight into seven weighted dimensions and scores them continuously. The output is a composite on a published 0 to 100 range with evidence attached, not a self assessment.
- Escalation that demonstrably fires, not escalation on paper.
- Reviewers with the authority and competence to act.
- Override paths that are exercised, not theoretical.
- A record an auditor can verify against the spec.
Why this matters now
US and EU AI oversight law both require that high-risk AI decisions involving a real person are made with meaningful human involvement -- not rubber-stamping. First enforcement dates land in 2027. The question regulators will ask is not whether a human was present. It is whether oversight was real, and whether it can be shown.