Four Logs, One Reshoot

The spec was locked on 2026-09-15 and it names 20 plates, from wide cafe at 01 through overhead edge at 20. Framing is the only variable it allows: the register block, the single-plate table layout, the room and the messy-and-real clause are identical in all 20 prompts. Four log files were written against that spec in one afternoon and together they hold 75 rows. Every row in all four reads done. Not one row reads failed, refused or retried. The first plate landed at 19:00:48 and the last at 20:29:12, so the record spans 88 minutes and 23 seconds, and the 75 finished plates on disk weigh 46,832,010 bytes. Every logged byte count matches the file it names, to the byte. The folders agree with the logs as well: 20 files in v3.0, then 19, then 17, then 19.

The first log weighs 5,442 bytes and holds 20 rows across 4 fields: status, file, bytes and ticks. There is no task id in it. There is no url in it. Nothing in those 20 rows reaches back to the call that produced the file, so the plates that came out of the first pass cannot be matched to a job, a response or a bill. The ticks field reads 400,000,000 in all 20 rows, one value written 20 times, which means the cost column cannot tell those plates apart. The byte column can. It runs 381,092 at the lightest to 480,603 at the heaviest, 20 different numbers for 20 rows, and the whole pass closed in 6 minutes and 48 seconds. A 5,442-byte file is the entire trace of 8,409,209 bytes of finished work.

The other three logs were written by a second renderer on a different model, seedream-v5-pro-uncensored, against the same shot list. They carry 5 fields instead of 4, and the two extra fields are the ones that make a plate traceable: a task id and a cdn url. Across those three logs there are 55 rows, 55 task ids and 55 urls, and no value repeats. Every task id is 7 characters long. Every url sits on cdn1.wiro.ai, every url names its own task id and Image0, and every url opens with a path segment of exactly 107 characters that takes 55 values but only 13 distinct openings. So provenance exists for 55 of the 75 plates. What the logs never carry is an attempt count. There is no retry key, no attempt key and no tries key in any of the 75 rows. A plate rendered once and a plate rendered nine times are written identically.

Two of the four logs weigh exactly 8,396 bytes each. They are the second pass and the fourth, and they describe the same 19 shot names in the same order, wide cafe at 01 through overhead hands at 19. Line them up cell by cell and there are 95 values across the 19 rows and the 5 fields. 76 of those values differ, 4 per shot, because the byte count, the task id, the url and the file path all moved. Exactly one field agrees in all 19 rows, and it is status, which reads done in both files. The weights behind that: shot 01 is 707,264 bytes in pass two and 772,643 in pass four. Shot 04 is 849,969 and then 614,018, which is 27.8 per cent lighter for the same framing. 9 shots came out heavier on the fourth pass and 10 came out lighter, and both passes carry the same 19 framings. Two files of identical size, 38 different plates.

One spec. Four logs. 75 rows.

Pass one delivered 20 plates. Pass two delivered 19. Pass three delivered 17. Pass four delivered 19. Three of the four passes are short, and every missing plate is an overhead framing: overhead edge in passes two and four, and overhead plate and overhead hands as well in pass three. Where each pass stopped is not random. Pass two stopped after shot 19, pass three after shot 17, pass four after shot 19. The reason sits outside the logs: the second renderer's own header says the model refuses top-down framing when a seated person is in frame, which is the constraint the shot list ran into. The logs kept the plates that came back and dropped the reason. The same log shows the overhead plates at the bottom of the weight table. In pass two the three overheads are the three lightest files at 450,301, 508,131 and 514,743 bytes. In pass four they are the three lightest again at 321,161, 472,519 and 472,731. In pass one, the only pass that delivered all four, they rank 4, 14, 16, 17.

Add the paper trail up. Four log files, 75 rows, 75 finished plates, 46,832,010 bytes, and not one duplicate byte count anywhere in the 75, so no two plates are the same file under two names. Pass two carries 4,869,728 bytes more than pass one while holding exactly one plate fewer. The heaviest plate of the set is 849,969 bytes and the lightest is 321,161, a spread of 2.65 times for one spec, one afternoon and two models. What is absent from this record is not data about the images. It is the two columns that decide whether the record can be trusted: which call made each plate, and how many times each plate had to be made. On 210 energy systems deployed across Africa and Asia, a unit was signed off only when the machine hours, the serial and the site all resolved on the same day. The probability that a hole in the log is a hole in the shoot is not zero, and the fix costs one field per row. Read the brief, then weigh the plate. Keep more of yourself, for longer.