A Trust Score for Every Action
One number a board or a buyer can read at a glance: how much should we trust this action right now? Derived deterministically from signals the decision already computes - never an invented risk.
Illustrative product view
What the score is
The Action Trust Score is the headline primitive of AegisAI - the credit-score analogy for actions. It is a single value from 0 to 100, where higher means more trustworthy, plus a band and the human-readable factors that drove it. It is computed as action_trust_score(risk) = clamp(100 − enriched_risk), so it never invents risk; it just makes the existing decision legible.
Action Trust Score
action_trust_score = clamp(100 − enriched_risk, 0, 100)
Trust is the inverse of enriched runtime risk, clamped to the 0–100 range. Because the risk is deterministic, the same action always produces the same trust score.
Trust bands
| Band | Meaning | Typical verdict |
|---|---|---|
| trusted | Expected actor, expected action, low blast radius | allow |
| monitored | Allowed but recorded for audit | allow_with_logging |
| step_up | Allowed only after step-up verification | require_step_up |
| review | Held for a human decision | human_review_required |
| contained | Sandboxed while reviewed (simulated) | contained |
| blocked | Deny recommended | deny_recommended / blocked |
The factors behind the number
- Non-human actor - a machine or agent lowers trust versus a known person
- Blast radius - high or critical target impact lowers trust
- Sensitive data - touching PII, secrets, or regulated data lowers trust
- Toxic combination - a detected toxic pattern lowers trust
- Policy violations - violations on the action path lower trust
- Active-compromise signals - a live compromise indicator lowers trust
Where the score appears
Surfaces that carry the score
- The PDP response returned to every enforcement point
- The Trust Ledger entry recorded for each decision
- Decision Intelligence, which ranks the riskiest recent actions
- Trust CORTEX, which compares the single highest-risk decision across domains
Frequently asked questions
How is the Action Trust Score calculated?
It is the inverse of the enriched runtime risk, clamped to 0–100: trust = clamp(100 − risk). The risk itself is computed deterministically from the action envelope and enrichment signals, so the trust score never introduces new risk of its own - it makes the decision readable as one number.
What is a trust band?
The band is the label implied by the score and verdict: trusted, monitored, step_up, review, contained, or blocked. The score is the magnitude and the band is the intent - together they tell you how risky the action is and what the platform recommends doing about it.
Is the score explainable?
Yes. Every score ships with its trust_factors - non-human actor, blast radius, sensitive data, toxic combinations, policy violations, and active-compromise signals - each marked as raising, lowering, or neutral. A reviewer never sees a bare number; they see why it landed where it did.
Can I use the score to auto-approve low-risk actions?
That is the intent as you move enforcement from simulate toward enforce. A trusted band on a low-risk, in-scope read is the kind of action you can let flow, while step_up and review bands hold for verification or a human. Because the score is deterministic, those thresholds behave predictably.
Stay in the loop
One practical finance briefing a week - new guides, checklists, and benchmarks.
Put a trust score on every action
Score your agents’ actions 0–100 with the factors that drive each verdict.
Talk to us