Trust Score.
Evidence compressed into a decision signal.

What should influence the score

• Task and objective alignment
• Output quality and evidence coverage
• Policy and constraint checks
• Tool/action risk
• Known missing evidence and uncertainty
• Recent evaluation history where appropriate

What it should not be

A Trust Score should not be marketed as a universal probability that an AI system is “safe.” It is an operational routing signal tied to a specific evaluation contract, policy version, action type and evidence set.

Example routing

score >= 90  → allow
75–89        → allow with warning
50–74        → human review
< 50         → block

// thresholds are policy-specific, not universal