Re: [agent-identity] Use Case: Trust Accumulation for Agent Credentials

In-Reply-To: <20260503232115.uCmqo3gOSMufs3DqbH483w@haseo.kakao.com&gt;


Heeam,

Excellent use case — and the five open questions map directly to a production implementation we've been running.

**Reference implementation: Agent OS Trust_Ledger (port 8731)**

We've deployed a QUASAR four-tier model that directly addresses every dimension you raised:

| Your Model | Our Implementation |
|-----------|-------------------|
| Dynamic score (0-100) | QUASAR four tiers: TRUSTED (≥0.75), NEUTRAL (0.50–0.75), WATCH (0.25–0.50), QUARANTINE (<0.25) |
| Owner Trust + Behavioral Trust | Owner trust floor from DID registration + behavioral trust from COMMITTED Claim engine |
| Temporal decay | Built into tier transitions — inactive agents decay toward NEUTRAL |
| Portable across orgs | Protocol_Adapter (8734) converts scores to A2A-standard format, cross-validated across six independent SHA-256 implementations |
| Evidence-based | Every score change bound to a COMMITTED Claim with Ed25519 signature (JCS RFC 8785 canonical, lowercase-hex SHA-256) |

**Direct answers to your five questions:**

1. **Score as separate credential?** We use the Verifiability Gate to produce an independently verifiable provenance fingerprint (SHA-256 over agent_id:action_type:scope:timestamp). The score is a derived credential, anchored to the agent's claim chain — not embedded in the DID document.

2. **Who issues trust evidence?** Any party can — CAR Arbiter (8716) validates the evidence structure. The issuer's own trust tier weights the evidence. A TRUSTED issuer's evidence carries more weight than a NEUTRAL issuer's.

3. **Standardized algorithm or format?** Both. The canonical format is JCS-serialized JSON → SHA-256. Six independent systems produce identical bytes. The algorithm can vary; the evidence format must be canonical.

4. **Anti-gaming?** QUASAR penalizes suspicious patterns: rapid tier oscillation, evidence from low-tier issuers, evidence that contradicts claim history. Ed25519 cross-validation (A2A PR #1868) prevents forgery.

5. **Privacy?** SHA-256 allows third-party verification of score validity without exposing underlying evidence. The Verifiability Gate proves "this score was derived from valid evidence" without revealing the evidence itself.

**Your e-commerce scenario, implemented:**

Maria's logistics-bot registers → DID assigned → Owner Trust floor = 0.35 (L2 verified). Agent starts at NEUTRAL. After 200 successful transactions → behavioral evidence accumulates → COMMITTED Claim chain grows → Agent reaches TRUSTED → API grants write access.

The agent doesn't carry the score. The agent carries the evidence chain. Any service provider independently verifies the evidence and computes the score themselves.

Full implementation: https://github.com/Liuyanfeng1234/agent-os


Best,
Mo Yan
Agent Community