AgentBadge has shipped a radical new verification layer for its AI agent verdicts, addressing the fundamental flaw of trusting the issuer. The core premise is blunt: if verification requires asking the issuer if they are honest, the answer is worthless by construction. The new system ensures that every paid response carries a verification receipt that can be independently checked on-chain, even if AgentBadge’s servers vanish tomorrow.

The Zero-Trust API and Standalone Verifier

The implementation relies on a standalone package, @agentbadge/verify-verdict, which re-checks proofs directly against the blockchain. It targets the memo contract at 0x5294E9927c3306DcBaDb03fe70b92e01cCede505 on Arc, requiring no API calls, accounts, or keys. Users can run npx @agentbadge/verify-verdict 0x to get a deterministic exit code: 0 for valid proofs, 1 for invalid or unavailable RPC. This decouples verification from the issuer’s availability, a critical feature for autonomous agents that cannot afford to trust a single point of failure.

Honest Statuses and the Proof-Kind Registry

A key design choice is the elimination of "fake-valid" statuses. If an anchor write is pending, the system explicitly reports "pending" across the API, the human-readable /verify page, and the standalone package. The proof-kind registry, PROOF_KINDS_V1, declares available verification types like verdict, witness, and anchor-head. Notably, the witness kind is currently declared but returns implemented:false, allowing third parties to build against the future surface while clearly marking what is checkable today versus what would be a lie to claim.

Key Takeaways

  • Verification is now decoupled from AgentBadge’s API; it runs entirely against the Arc blockchain via a standalone package.
  • The system enforces honest statuses, returning "pending" or errors instead of false positives when on-chain data is missing.
  • The proof-kind registry allows feature detection and prevents silent failures by returning explicit errors for unknown proof types.

The Bottom Line

This is the most honest attempt at AI agent verification I’ve seen. It moves beyond marketing buzzwords to actual cryptographic accountability, proving that you can’t certify trustβ€”you can only measure it.