Take the title case apart first, because it is the cleanest of the five. A validation rejects article titles longer than 58 characters, and the generation prompt says nothing about a limit [5]. So the model returns 63 characters, the validator rejects it, the loop sleeps, and the same prompt goes back to the same model [5]. Three attempts, three 63-character titles, then the give-up path with no fourth try [5]. Nothing between attempts changed the input, so nothing could change the output. note-autolike produced zero rewrites for three days behind that, with its last success stamped 2026-08-06T09:03:20Z [4].
The moderation ladder in ai-portraits is worse, because it looks like classification. Three stages, each softening the prompt, up to 21 minutes per concept and up to 63 minutes for three [6]. Twenty-one times three is 63, so the duration was specified in advance rather than discovered [2]. And when the target content is itself what trips moderation, stage three lands exactly where stage one landed [7].
The slot lane's failure is basic arithmetic. There are three global slots, a run takes 14 to 62 minutes, and launch attempts arrive at 18 to 34 an hour [8]. At the 14-minute floor three slots can start 3 x (60/14) = 12.9 runs an hour; at the 62-minute ceiling, 2.9 [3]. Demand exceeds the best case before anyone chooses a retry policy. The pre-fix script printed `SKIP: global limit reached` and exited 0 the moment acquire failed [9]. Exit 0 is the success convention, so whatever schedules the job files the skip as a clean run [4]. Across the 9, 11 and 12 o'clock hours, 29 runs never started and the counters stayed green [1][10]. One account hit skip on both of its runs and finished the day with 0 likes and 0 follows [10].
The fourth lane inverts the error. watch-arb recorded only `extract_failed` when a product page would not fetch, which reads on a dashboard as 16 shops with markup you cannot parse [11]. The actual response was HTTP 429, and a prober that hit the same shops slowly, one at a time, fetched them correctly [11]. The same 429 turned up under seven different names across the logs [12]. The operator's own account is that the wrong name kept him fixing in the wrong direction [13].
For any of these counts to transfer, your environment has to look like the one described: a single operator's SNS automation, jobs expected to self-correct overnight, and five separate lanes filing retry bugs on 8 and 9 August 2026 [2][15]. The counts will not transfer. The three questions will: will waiting fix it, when do you cut it off, and what do you change before the next attempt [14]. Answering the first one is fault classification, and it decides whether the other two are worth writing. A `for attempt in range(3)` with `time.sleep(2)` answers none of them, and it is correct only for the case it was written for, which is a network blip or a server that got temporarily slow [3]. That sleep is a bet about the world outside your process: that something is busy and will stop being busy [3].
