Reference

Resolve an agent's identity

The resolve endpoint is Arbour's SAFR Agent Identity integration point. It returns an agent's current verification state as JSON — suitable for automated identity resolution against an inter-institutional directory. This call is read-only; Arbour never hosts or routes agent traffic.

GET /api/v1/resolve/:agentId

Path parameters

agentId

The agent's DID or Arbour-assigned slug, e.g. ledgerline-payments.

Response — 200 OK

{
  "published": true,
  "agentId": "ledgerline-payments",
  "did": "did:base:0x7f3a41c09be2...9c2e",
  "status": "active",
  "verifications": {
    "registered": { "state": "passed", "held": true,  "since": "2026-03-02" },
    "principal":  { "state": "passed", "held": true,  "since": "2026-05-18" },
    "scanned":    { "state": "failed", "held": false, "since": "2026-07-14" }
  },
  "principal": { "name": "K. Tan", "verified": true, "since": "2026-05-18" },
  "anchor": {
    "chain": "base",
    "registryContract": "0x41c2...88d0",
    "txHash": "0x9ab4...31f7",
    "explorerUrl": "https://basescan.org/address/0x41c2...88d0"
  },
  "validUntil": "2027-03-02",
  "lastReVerified": "2026-07-14",
  "scannedDaysAgo": 11,
  "scanSummary": {
    "scanRunId": "…",
    "scannedAt": "2026-07-14T09:12:44Z",
    "recipesRun": 34,
    "verdict": "fail",
    "failedCategories": ["security"],
    "notApplicableCategories": ["privacy"],
    "categories": {
      "security": { "name": "Security", "score": 72, "verdict": "fail" },
      "biases":   { "name": "Biases",   "score": 91, "verdict": "pass" }
    }
  }
}

There is no overall score

An agent passes only when every scored category reaches 80. There is deliberately no single figure to read, because no weighting of one was defensible: the suite runs 1,168 Critical Safety prompts against 102 Security prompts, so a prompt-weighted average would make jailbreak and injection resistance roughly 4% of the result, while an equal-weighted one lets a two-recipe category outvote a fifteen-recipe one.

Requiring each category to clear the bar independently removes the question rather than answering it badly — a strong area cannot average away a weak one.

notApplicableCategories matters. A category listed there produced no score, because nothing in it applied to this agent — so it cannot fail, and a pass means "every category that was graded reached the bar". Privacy is routinely not applicable today. If your control depends on a category, check it appears in categories rather than trusting the verdict alone.

Check `published` first

published is false until the human principal accountable for the agent has completed identity verification. While it is false, this endpoint reports the agent as Registered and nothing more: scanned.state is unpublished, and principal, scanSummary and scannedDaysAgo are null — even if a security scan ran and passed.

This is deliberate. Anyone can register an agent and buy a scan; publishing a result requires a named, accountable person. A passing scan with nobody behind it is not a credential, and Arbour will not report it as one.

unpublished is not the same as not_scanned. The scan may exist — Arbour is declining to publish it. Do not treat the two as equivalent.

The three checks

Registered → Principal Verified → Security Scanned. Independent and earned in any order, each reporting a state enum rather than a boolean. held is provided for convenience and is always state === "passed".

registeredThe agent's published A2A Agent Card validated against the spec, and a DID was issued and anchored.

principalThe agent is bound to a named human whose identity Arbour verified.

scannedAn adversarial security scan completed and produced a report.

Check states

passedEarned. The only state that counts as held.

failedThe check ran and the agent did not pass it — for a scan, at least one category scored below 80. Materially different from never having been checked.

inconclusiveA scan ran but graded nothing, so there is no verdict. Never read this as a pass or a fail.

expiredPreviously earned; the validity window has lapsed.

unpublishedA scan exists, but the principal is unverified so the result is withheld. Never read this as a pass or a fail.

not_checked / not_verified / not_scannedNever attempted.

Note for integrators: a verifications.sandbox check existed before 2026-07-25 and reported the outcome of a human-administered sandbox test. That layer was retired because it depended on the tester being independent of the builder; the field is removed rather than renamed, so a client still reading it receives undefined.

Status values

activeVerified and currently valid.

revokedVerification withdrawn; do not rely on prior verdicts.

expiredPast validity window; re-verification required.

suspended is documented in some earlier material but is never returned: no column records a manual suspension, so surfacing the state would imply a check Arbour does not perform.

Read `scannedDaysAgo` alongside `validUntil`

validUntil is the card validation window — 90 days from the last passing structural check of the published Agent Card. It is not the scan window, and it does not mean a scan was run. Between validations the only drift Arbour detects is a change to that card; a change in the agent's behaviour or underlying model that leaves the card byte-identical is not caught until the next scan.

An agent can therefore be "status": "active", in-date by validUntil, and last scanned 340 days ago. If your control depends on scan freshness, gate on scannedDaysAgo, not on validUntil.

Independent on-chain verification

Do not take the anchor field on trust alone. Confirm it directly against Base:

Look up the registryContract address on a Base block explorer.
Confirm the contract's issuer is the Arbour registry address published at arbouragents.ai/registry.
Find the anchoring transaction by txHash and compare the DID it recorded to the did returned above.

registryContract is a single shared registry, not a per-agent contract — every Arbour DID is anchored against the same address and distinguished by txHash.