Tuning a Routine That False-Alarms Weekly

In the case study, a nightly cloud routine reconciles source-table row counts against each export pipeline's CSV output and, on a discrepancy, opens a GitHub issue tagged data-integrity and posts a summary to Slack.

Three months later, the routine has become background noise: it fires a "discrepancy found" alert roughly once a week, and every single time so far it's turned out to be a false alarm — a batch export that was still mid-write when the routine's nightly run happened to check it, not an actual missing-rows bug.

1. Diagnose the likely cause of the false alarms, being specific about what in the routine's design produces this failure mode.
2. Propose a concrete change to the routine's prompt or configuration that would fix it, and explain why your fix doesn't just trade one failure mode (false alarms) for the opposite one (missing real incidents).
3. What would you check in the routine's run history before assuming your fix worked, rather than just deploying the change and hoping?