Don't trust the AI because it says it worked

An AI agent that reports success after running a command is telling you what it did, not what happened. Those are different claims, and conflating them is how incidents get closed before they're actually fixed.

When Changefy restored a broken Sentinel log pipeline for WRK-2041, the execution log showed every step succeeding: the IAM role was granted, the sink was redeployed, the forwarder restarted cleanly. By that log alone, the job was done.

Verification asked a different question: is Sentinel actually receiving events again? It checked the connector directly, independent of the execution trace, and found a live event 42 seconds after restart with a 0% error rate. That's the claim that matters, and it's the one we show first.

This is why verification is a distinct stage in Changefy's workflow, not a summary generated from the execution log. It re-queries the target system using the same criteria defined in the plan, and the work item isn't marked verified until that check passes.