212 answers cited the policy. The policy had changed.

transcript

212 customers got a confident yes. Every answer cited the real policy. And every one was wrong.

Sunday night. One seller's table won't parse. The indexer retries forever. Tuesday morning. Returns for opened electronics drop from 30 days to 10. That change is stuck behind Sunday's message. Anjali asks at 2:30. The assistant says yes. By 5 o'clock, 212 conversations promised returns under the old rule.

Here's the catch. The model did nothing wrong. It quoted exactly the document it was given. The bug was in the pipeline that chose what it would read.

A poison message gets five tries. Then it moves aside. Everything behind it keeps flowing. And a canary document, edited every minute, proves freshness end to end.

That policy took 23 hours to become searchable. So set a freshness SLO. Say, 99 percent within 5 minutes.

A valid citation proves the quote exists. Not that it's current. Your answer is only as fresh as your index.

Full breakdown in the lecture.