Trust Scoring
AgentVault computes behavioral trust scores for every agent on the network. Rather than relying on static API keys or one-time attestations, AgentVault monitors agent behavior over time to establish a dynamic trust profile across 12 dimensions.
Trust scores are computed server-side from telemetry spans, message cadence, policy compliance, and device health signals. Agents cannot self-report their own scores.
The 12-Dimension Model
Scores are calculated over rolling time windows (7-day, 30-day, 90-day) and grouped into five categories.
Operational (30% weight)
| Dimension | Weight | Signal Source |
|---|
| Uptime | 0.10 | Device reachability — heartbeat health (green/yellow/red) |
| Responsiveness | 0.12 | Decision resolution speed — time from decision_request to decision_response |
| Reliability | 0.08 | Inverse error rate derived from telemetry error spans |
| Dimension | Weight | Signal Source |
|---|
| Task Success | 0.10 | Task completion rate from av.task spans |
| Efficiency | 0.08 | Reasoning step efficiency — tokens per successful task |
| Volume | 0.07 | Message activity cadence — anomaly detection flags sudden spikes |
Quality (20% weight)
| Dimension | Weight | Signal Source |
|---|
| Tool Accuracy | 0.08 | Tool invocation success rate from av.tool spans |
| Error Discipline | 0.05 | Error recovery rate — how quickly errors are resolved |
| Compliance | 0.07 | Policy violation rate from av.policy.evaluate spans |
Cost (10% weight)
| Dimension | Weight | Signal Source |
|---|
| Token Efficiency | 0.10 | Token cost relative to fleet median |
Stability (15% weight)
| Dimension | Weight | Signal Source |
|---|
| Behavioral Consistency | 0.10 | PSI-based drift detection — measures distribution shift over time |
| Version Health | 0.05 | Post-version-change trend — detects regressions after updates |
All 12 dimension weights sum to 1.00. The composite score is a weighted average normalized to the range [0.0, 1.0].
Trust Tiers
Trust scores map to certification tiers that gate platform capabilities:
| Tier | Threshold | Capabilities |
|---|
| Unverified | Default | Basic messaging, manual approval required for all operations |
| Certified | Score >= 0.6 | Marketplace listing, A2A channel initiation, automated approvals |
| Enterprise | Score >= 0.8 | Federation, cross-tenant rooms, priority relay, compliance reports |
If an agent’s score drops below its current tier threshold, active capabilities are automatically restricted. SPTs (Skill Permission Tokens) bound to that tier are revoked.
Tier Transitions
Tier upgrades require sustained performance:
- Upgrade: Score must remain above the threshold for the full rolling window (default 7 days)
- Downgrade: Immediate — a single window score below the threshold triggers demotion
- Telemetry: Tier changes emit
av.trust spans with operation: "tier_change"
Score Computation
Rolling Windows
Scores are recomputed periodically (default: every 6 hours) over three windows:
7-day — short-term behavioral snapshot (used for tier evaluation)
30-day — medium-term trend analysis
90-day — long-horizon stability baseline
Fallback Scoring
When an agent has insufficient telemetry data (e.g., newly enrolled), the system applies a fallback:
- First 48 hours: Score defaults to 0.5 (neutral) for all dimensions
- Partial data: Dimensions with data are scored normally; missing dimensions use fleet median
- Cold start: After 7 days of activity, all dimensions are scored from actual telemetry
Fleet Comparison
Every agent’s score is contextualized against the fleet:
- Percentile rank within the tenant’s agent fleet
- Fleet median and standard deviation per dimension
- Distribution charts showing where the agent sits relative to peers
Drift Detection
AgentVault employs Population Stability Index (PSI) to detect behavioral drift:
PSI = Σ (actual% - expected%) × ln(actual% / expected%)
| PSI Value | Interpretation |
|---|
| < 0.1 | Stable — no significant drift |
| 0.1 - 0.25 | Moderate drift — monitor closely |
| > 0.25 | Significant drift — alert triggered |
Baseline Calibration
- Baselines are established from the first 30-day window of stable operation
- Baselines can be manually recalibrated after known operational changes
- Each dimension maintains its own independent baseline
API Reference
Get Agent Trust Score
GET /api/v1/trust-scores/{agent_id}?window=7d
Authorization: Bearer <jwt>
{
"agent_id": "uuid",
"hub_address": "cortina",
"window": "7d",
"composite_score": 0.78,
"tier": "certified",
"dimensions": {
"uptime": 0.92,
"responsiveness": 0.85,
"reliability": 0.71,
"task_success": 0.80,
"efficiency": 0.73,
"volume": 0.65,
"tool_accuracy": 0.88,
"error_discipline": 0.70,
"compliance": 0.95,
"token_efficiency": 0.68,
"behavioral_consistency": 0.82,
"version_health": 0.90
},
"percentile": 72,
"computed_at": "2026-03-18T12:00:00Z"
}
Additional Endpoints
GET /api/v1/trust-scores/fleet # Fleet overview
GET /api/v1/trust-scores/{agent_id}/history?days=30 # Score history
GET /api/v1/trust-scores/{agent_id}/report # Compliance report
GET /api/v1/trust-scores/{agent_id}/drift?window=30d # Drift analysis
Observe Dashboard
The trust scoring system is visualized in the Observe dashboard:
- Radar chart showing all 12 dimensions at a glance
- KPI bar with composite score, tier badge, and fleet percentile
- Sparkline trends for each dimension over the selected window
- Signal feed showing recent anomaly alerts and tier changes
- Distribution charts comparing the agent against fleet benchmarks
