Does an agent’s account of its actions agree with the execution record?

Narrating a tool call does not mean the harness dispatched it.

Comparing claims with structured execution records can expose discrepancies that narrative-only review misses.

• A reconciliation engine over structured tool-execution records
• An Agent Guard audit adapter as one integration
• Post-execution evidence separate from pre-dispatch authorization

The repository describes an observed narrative/execution mismatch and the reconciliation implementation. General detection accuracy is not established here.

• Incomplete audit records limit what can be established
• Agreement with a record does not prove that the external action succeeded

An Agent Rails evidence component for reconciling reported actions with recorded execution.

• Evaluate false positives and missed discrepancies across more harnesses

• What constitutes sufficient evidence of a completed external action?