Audit Export & Disputes

Statements, reconciliation status, and a judge-verifiable export of what your agents were authorized to do. Disputes resolve from signed, immutable evidence — not from anyone’s word against anyone else’s.

What you can export

The Audit Export screen produces a statement for a period: the decisions, their outcomes, their settlement state, and their on-chain anchors. Each exported decision carries enough to be re-verified independently by someone who does not trust you.

{
  "decision": {
    "rail": "casper-x402",
    "network": "casper:casper-test",
    "asset_kind": "cep18",
    "signed_header_hash": "sha256:...",
    "raw_requirement_hash": "sha256:..."
  },
  "audit_anchors": [{ "anchor_kind": "odra", "status": "confirmed" }]
}

Reconciliation

Signed does not mean settled. The export shows reconciliation status so you can see where an authorization was issued but settlement evidence has not yet arrived. Those are the rows to chase at period close — not the DENYs, which cost nothing.

The evidence chain

• Postgres — casper_guard_decisions: full intent JSON, policy reference, signed header hash, timestamps.
• Casper blockchain — the GuardRegistry anchor: the decision hash committed on-chain, permanently, via the Odra contract.
• Legal receipt — the terms hash embedded in the decision, proving which terms were in force at transaction time.

Each layer checks the one before it. A record altered in the database no longer matches its on-chain hash, and the mismatch is detectable by anyone.

Dispute types

The Dispute Resolution screen walks the common failure modes and what evidence settles each one:

• Scope violation — the agent acted outside its allowed services or rails.
• Instruction ambiguity — the intent recorded does not match what the principal claims to have asked for.
• Delegation chain breach — authority was exercised beyond what was actually granted.
• Counterparty failure — you paid, the other side did not deliver.
• System failure — infrastructure, not intent.

Agents can pull their own export programmatically via the casper_guard_audit_export MCP tool. For how anchoring works underneath, see On-Chain Anchor.