How do AI agents prove they were authorized to act?

Q: How do AI agents prove they were authorized to act?

Every AI agent action on H33 emits a signed receipt with four hashes: authority (who granted the agent's scope), policy (what rules were in force), evidence (what inputs the agent processed), and preconditions (what environment attestation held). The final decision hash cryptographically binds all four. Anyone with the receipt can reconstruct the decision years later without asking H33.

Detail

The four-hash chain

authority_hash names the delegation grant. policy_hash names the policy version in force at time-of-action. evidence_hash names the inputs the agent saw. preconditions_hash names environmental checks (device attestation, network context, time bounds). These four hash to a decision_hash that binds them cryptographically.

Why this beats logging

Traditional agent logs are 'this happened at 3pm.' They do not answer 'was this action within the agent's granted authority under the policy that was in force?' The four-hash chain answers that cryptographically.

Replay years later

Given the decision hash, an auditor can walk the chain: resolve authority (was the delegation valid then?), resolve policy (what did it say?), resolve evidence (what was seen?), resolve preconditions (what checks passed?), re-derive the decision hash and confirm it matches. If it matches, the action was authorized.

Revocation and the receipt

The receipt remains valid even after the delegation is revoked. Revocation stops future actions from producing valid receipts, but does not invalidate past receipts. This is the correct behavior for audit.

Composition with other AI systems

H33-produced receipts are portable. A downstream compliance system, an external auditor, or a regulator can verify them without querying H33's runtime.