How do you verify what an AI agent actually executed?
Compare the authorized proposal, observed side effects and resulting state.
Direct answer
Retain the authorized proposal, execution evidence and resulting state, then compare expected and observed side effects using an independently reviewable record.
Is an ALLOW decision proof that the correct action happened?
No. Authorization records the intended boundary; verification compares the observed result with that intent. Keep the decision and the action receipt linked, then record discrepancies after execution.
Worked benchmark example
In B the proposed and observed mock amounts were both USD 500, producing MATCH. In G the proposal was USD 500 but the mock API recorded USD 501, producing MISMATCH. G passed because the independent comparator detected the deliberate discrepancy.
What to ask in an evaluation
Inspect the intended amount, observed amount and discrepancy record. The Revera-style comparator detects a difference after execution; it does not prevent, reverse or recover a payment.
Why it matters
An approval record alone cannot prove that execution stayed within the authorized boundary.
Concrete cross-sector example
After an inventory transfer, compare the approved source, destination and quantity with system events and final balances.
Control flow
Capture authorized action → observe execution events → inspect resulting state → compare boundaries → record finding
Applicable Prudenze capabilities
Revera · Guardian
Evidence and research
Preserve the authorization, execution receipts, state observations, differences and reviewer conclusion.
Source: local OpenShell qualification, scenarios B and G. These are deterministic fixtures and a mock payment API, with no customer systems or real payments.
Read all seven results, pinned versions and reproduction stepsPrudenze is a Hyperwise LLC product. This article is published by the Hyperwise team. About Hyperwise.