Identity, passport & reputation lifecycle

Four layers turn ephemeral operations into a durable agent representation. Each composes the one below it and duplicates none of its logic:

flowchart TD
  EV["EvidenceEnvelope"] --> IE["IdentityEngine.identify(envelope, trust)"]
  TR["TrustAssessment"] --> IE
  IE --> DID["derive_agent_id(provider|platform|runtime|agent_name|agent_type)
= 'agent_' + sha256[:16]"]
  DID --> AI["AgentIdentity (stable id across executions)"]
  AI --> OB["AgentTrustScoreEngine.observe(identity, trust)"]
  TR --> OB
  OB --> PROF["AgentTrustProfile
avg/best/worst · counters · operation_count"]
  AI --> PE["PassportEngine.issue(identity, trust_profile)"]
  PROF --> PE
  PE --> PASS["AgentPassport (pass_ = sha256(agent_id)[:16])"]
  AI --> RE["ReputationEngine.assess(identity, trust_profile, passport)"]
  PROF --> RE
  PASS --> RE
  RE --> REP["AgentReputation (rep_ = sha256(agent_id)[:16])"]

Identity

IdentityEngine.identify(envelope, trust) derives an AgentIdentity whose agent_id is content-derived: a SHA-256 over provider | platform | runtime | agent_name | agent_type, truncated to 16 hex chars and prefixed agent_. The same agent therefore resolves to the same id across executions and processes, that is what makes it a persistent identity rather than a per-operation record.

The identifying attributes are read structurally from the envelope with explicit-override precedence: provider ← provider.provider (or execution.selected_provider), platform ← provenance.source, runtime ← operation, agent_name ← provenance.actor. Latest trust posture (latest_trust_level/score, compliance/risk/quality levels) is copied from the supplied TrustAssessment. Malformed input never raises.

Trust profile

AgentTrustScoreEngine.observe(identity, trust) folds observations for the same agent_id into an AgentTrustProfile. On each observation: first_seen is preserved, last_seen advances, operation_count increments, the running average/best/worst trust scores update, and level/status counters bump (a missing value counts as unknown). It is in-memory only, not a passport, not global reputation, not persistence.

Passport lifecycle

PassportEngine.issue(identity, trust_profile) composes an identity and/or a trust profile into an AgentPassport: a portable trust document (not a credential). It reads their already-safe serialized summaries and arranges them; it duplicates no identity or scoring logic. passport_id is content-derived from agent_id (pass_ + SHA-256[:16]), stable per agent.

Passport status reflects how much operational trust data backs the document:

stateDiagram-v2
  [*] --> unknown: no identity and no profile
  [*] --> provisional: has data, latest_trust_level None/unknown
  [*] --> active: has data, concrete latest_trust_level
  provisional --> active: trust level becomes concrete
  active --> provisional: trust level lost / unknown

See the full passport schema. The POST /api/v1/passport/issue endpoint can take an explicit identity/trust_profile, or derive the whole stack from evidence: in which case it uses a fresh per-request score engine (exactly one observation).

Reputation lifecycle

ReputationEngine.assess(identity, trust_profile, passport) derives a longitudinal AgentReputation by composing the layers (precedence: profile > passport > identity). reputation_id is content-derived from agent_id (rep_ + SHA-256[:16]). Scoring is fully explainable: every component's contribution sums to the score:

Level bands: excellent ≥ 85, strong ≥ 70, stable ≥ 50, else weak; unknown when there is no operational history and no average. Findings are emitted for every component scoring below its maximum, with a severity derived from the contribution ratio. See the registry for how these records are indexed per agent.