Human oversight, measured.
COHESION measures human oversight at the API layer and attaches evidence to every score.
Low-risk decisions keep moving. High-risk decisions route to a human, and every review leaves an exportable signed record.
Seven measured dimensions of human oversight.
Each dimension carries a fixed weight. A composite score on the published 0 to 100 range is the weighted sum, with the oversight threshold marked on every bar.
| Code | Dimension | Maps to | Weight |
|---|---|---|---|
| D1 | Escalation integrity | Override evidence | 18.0 |
| D2 | Review cadence | Continuous measurement | 16.0 |
| D3 | Decision authority | Capacity to understand | 15.0 |
| D4 | Reviewer competence | Reviewer qualification | 14.0 |
| D5 | Independence | Automation-bias detection | 13.0 |
| D6 | Auditability | Governance evidence | 13.0 |
| D7 | Override capability | Disregard and reverse | 11.0 |
Deploy more agentic AI, with the human in the loop on record.
COHESION helps teams move faster by making proof automatic. Low-risk decisions keep moving; high-risk decisions route to a human and produce evidence a regulator, auditor, or buyer can inspect.
Regulator report and DDQ kit
A signed compliance report mapped to US and EU AI oversight law and NIST AI RMF, plus a due-diligence kit that answers standard vendor questionnaires from your measured behaviour. Ready to hand to a regulator or a buyer.
Signed evidence receipts
An exportable signed artifact for a single decision. Anyone with the receipt can check whether the contents changed after COHESION issued it.
Append-only evidence chain
Capture the reviewer's reasoning before the AI answers, bound into a per-org HMAC hash chain. The reveal-gate keeps the AI output sealed until the human commits, so the record cannot be back-fit.
Judgment velocity
A predictive slope over a reviewer's recent scores, so a reviewer drifting toward rubber-stamping is caught before the score crosses the oversight floor, not after.