The gate that fired on its own code: check that you were allowed to generate, not the output

The three logged timestamps are worth reading as data rather than as an error log. The first posts landed on 2026-07-28, 2026-08-04 and 2026-08-11 ^[13], which is seven days apart, twice ^[2], and all three fell within two minutes and twenty-one seconds of the same clock time ^[4]. The weekly assumption the author distrusts held perfectly for this season. What was wrong was the thing it got compared against: the stub generator had stamped the current date onto every episode it created ^[14].

Four stubs were made, and only three of them produced a mismatch entry ^[1]. If episode 7's thread followed the same seven-day cadence, its first post was 2026-08-18, five days before the claimed 2026-08-23, and the function's rule of `abs(gapDays) <= 5` returns `ok` ^[3]. So the check cleared one record that was wrong for exactly the same reason as the three it caught. The five-day tolerance that stops time-zone drift and overseas streaming lag from raising false alarms ^[11] is also wide enough to absorb a five-day error in the claim. Nothing got published anyway, because every episode must pass or nothing is written ^[12]. That is the aggregation rule earning its keep in the precise way the author predicted: the passing record was only agreeing by accident ^[12].

What does the work is which side each operand comes from, not the subtraction ^[10]. Calendar arithmetic can nominate candidates but it never observes a broadcast ^[5], so a validator reading only your own store is asking whether a value is well-formed, and 2026-08-23 is well-formed ^[15]. The output-shaped question has no answer here. The authorization-shaped question, has anything outside my system recorded this episode airing, does.

The reason this is cheap rather than architectural is that the evidence lives where the payload already comes from: MyAnimeList's per-episode discussion threads are both the comment source and the proof of broadcast ^[7][8]. The marginal cost is parsing one more field out of a response you were already fetching. Pipelines that scrape a source and then validate against their own metadata are usually one field away from the same gate.

Two limits on the read, and both come from the post itself. This is one operator's account of one automated site ^[1]. And the failure the design exists for, a page for an episode that has not aired ^[2], is not the failure that was logged. Both worked examples of the check are given a first-post timestamp ^[17], so what the comparison returns when no thread exists yet is not demonstrated ^[5]. The day the gate fired, it fired on a bad date in a stub, which is the cheaper cousin of the real risk: a pipeline that decides an empty page looks bad and fills it, putting invented sentences under the names of real viewers ^[3].