Nothing looked broken
I sent my AI assistant a shortlink to a Reddit post. It followed the redirect to a real OpenClaw page, read it, and gave me a polished take. The page was relevant. The prose was coherent. It was not the post I had sent.
That was the useful part of the failure. Every individual step looked reasonable while the conclusion rested on the wrong source. The system had completed the task it thought it had. I had no reason to trust the result.
The mistake happened before the analysis
A better prompt for the final explanation would not have repaired the source. By the time I was judging the prose, the important error was already buried upstream.
This is why output review has a limit. A reviewer can catch an unsupported claim or an odd conclusion. They may not recognize that a plausible summary came from an adjacent policy, an earlier version, a different jurisdiction, or the wrong employee record.
The page can be real, relevant, and wrong for the job at the same time.
I changed what verified means
The workflow no longer treats a successful redirect as verification. Before it interprets the page, it compares the canonical destination with a signal from what I actually shared: the title, author, quoted text, or another detail specific enough to identify the source.
If the destination and the signal do not agree, the workflow stops. It does not guess which page I meant or continue because the substitute looks close enough.
- Resolve the shortlink and record the canonical destination.
- Capture a distinctive signal supplied with the request.
- Compare that signal with the destination content.
- Stop and ask for clarification when they do not match.
- Keep the source identity with the analysis and final receipt.
The HR version of this failure is expensive
In HR work, a wrong policy version, jurisdiction, employee record, or nearby legal update can still produce a very convincing answer. The fluency of the output does not reduce the cost of starting from the wrong thing.
I would rather have the workflow say it could not verify the source. That creates a small delay people can see. Quietly analyzing the wrong source creates rework and a decision trail nobody should have to reverse-engineer later.