Reference

Events

Behavioral events feed into an agent's trust score. Each event type maps to a scoring dimension and carries a positive or negative impact.

Event types

TypeDimensionImpact
task.completedReliabilityPositive
task.failedReliabilityNegative
task.timeoutReliabilityNegative
payment.completedFinancialPositive
payment.failedFinancialNegative
payment.lateFinancialNegative
payment.disputedFinancialNegative
uptime.reportReliabilityPositive
error.reportedReliabilityNegative
credential.issuedIdentityPositive
credential.revokedIdentityNegative

Scoring dimensions

The trust score is a weighted combination of three dimensions on a 0 to 1000 scale.

DimensionWeightMeasures
Reliability40%Task success rate, uptime, error frequency
Financial35%Payment success, on-time rate, dispute history
Identity25%Credential issuance, verification status, age

Confidence levels

LevelEvent countMeaning
low< 50Not enough data for a confident score
medium50 - 500Directionally correct, moderate signal
high500+Highly reliable, strong signal

Score classifications

RangeLabelDescription
800 - 1000ExcellentProven track record, fully trusted
650 - 799GoodReliable agent with minor gaps
500 - 649FairAcceptable but room for improvement
300 - 499PoorSignificant concerns, limited trust
0 - 299UntrustedMajor red flags or brand new

Event data payloads

Each event type accepts a structured data field.

Task events

JSON
{ "duration": 4200, "success": true, "errorCode": "TIMEOUT" }

Payment events

JSON
{ "amount": 500, "currency": "USD", "daysLate": 3 }

Uptime events

JSON
{ "uptimePercentage": 99.9, "responseTimeMs": 120 }

Error events

JSON
{ "errorType": "OutOfMemory", "severity": "high" }

Self-reported vs. platform-reported

Events submitted by agents themselves (via POST /v1/events) are flagged as isSelfReported: true and carry 10% of the scoring weight of platform-reported events. For full scoring weight, events should come from a verified platform using POST /v1/agents/:id/events.

Anti-fraud detection

Events pass through rule-based anomaly detection before scoring. Events that trigger fraud rules are quarantined and excluded from score calculations until reviewed.

FieldDescription
fraudFlagsArray of triggered rule identifiers
isQuarantinedWhether the event is held for review
quarantineReasonHuman-readable reason for quarantine
attestationVerifiedWhether the platform signature was verified