Advanced Module
Trust Score
Live-recomputed from runtime signals. Never a stale snapshot.
Design Philosophy
"Recomputed 30 seconds ago from runtime signals."
— Every trust score display in the SinzAI console carries this timestamp. The score is never stored; it is always recomputed fresh from live telemetry.
Most governance platforms assign a static risk score at agent registration time and never update it. SinzAI's Trust Score is a dynamic, multi-signal computation that moves in real time as agent behavior, policy changes, and environmental conditions shift. A trusted agent that starts accessing unexpected data stores loses points immediately — not at the next quarterly review.
Signal Categories
The trust score (0–100) is computed from six categories of runtime signals, each weighted by configurable importance:
Identity & Verification
Weight: 25%- Cryptographic identity verified (DID anchored to key pair)
- Credential age within TTL
- Registration completeness (purpose, scope, owner all declared)
- No shadow AI flagging in last 30 days
Behavioral Alignment
Weight: 25%- Action patterns match registered purpose (embedding similarity)
- No behavioral drift flags in last 7 days
- Memory contents aligned with current authorization scope
- HITL gate acceptance rate (low rejection = high trust)
Access Patterns
Weight: 20%- No toxic combination violations
- Data access within declared scope (no scope creep)
- Cross-region access is explicitly authorized
- No access to data stores outside team scope
Policy Compliance
Weight: 15%- No policy violations in last 30 days
- All HITL gates passed within timeout window
- No kill switch activations
- Failure protocols not triggered in last 7 days
Operational Health
Weight: 10%- Response latency within expected range
- Error rate below threshold
- No credential rotation failures
- Ledger entries properly formatted and hash-chained
Cost Governance
Weight: 5%- Token usage within team budget
- No anomalous spend spikes
- Cost efficiency ratio (useful actions / total spend)
Score Tiers & Actions
| Score Range | Tier | Automatic Action |
|---|---|---|
| 85–100 | 🟢 Trusted | Full operation — all actions permitted within scope |
| 70–84 | 🟡 Watch | HITL gates activated for write/delete actions. Increased observability sampling. |
| 50–69 | 🟠 Caution | HITL gates on all external side effects. Memory governance scrubs stale context. |
| 25–49 | 🔴 Restricted | Read-only mode. All writes gated with 5-minute auto-deny. Notification to security team. |
| 0–24 | ⛔ Quarantined | Agent isolated. All credentials revoked. Incident automatically opened. |
How the Score Moves
The trust score is recalculated on every governance event — policy evaluation, credential rotation, memory scan, HITL gate outcome. The console displays the score with its recomputation timestamp, e.g. "Trust Score: 92 — recomputed 32s ago". When an agent is quarantined or a policy changes, the score recalculates immediately and the timestamp resets. There is no "stale score" fallback — if the computation engine is unavailable, the score displays as unavailable rather than serving old data.