The Same Seventeen Kilobytes Five Times

Six cut lists were written on the afternoon of 2026-09-11, and the whole set of them covers 31 minutes 31 seconds of wall clock. The first, sal_y_miel_movie_sequence_01, was written at 2026-09-11T15:54:16 and weighs 15,838 bytes. It declares 83.0 s of finished film drawn from 7 distinct sources, all of them from a single project, and its target durations add to exactly the length it claims. Two minutes and 26 seconds later the same generator wrote sequence 02 at 2026-09-11T15:56:42: a 78.86 s cut from 11 sources across 5 projects. Between 2026-09-11T15:56:42 and 2026-09-11T16:25:47 four more of these files appeared, and the sixth closes at 2026-09-11T16:25:47. That is the entire wall clock cost of six cut lists for six versions of one film. Nothing new was filmed for any of them, and no human edit decision is recorded anywhere in the set.

Sequences 03, 04 and 05 sit in the archive folder and sequence 06 sits in the working folder inside the same refinery. Their cut lists are 17,013 bytes, 17,013 bytes, 17,013 bytes and 17,013 bytes. That is one number written four times. The four sha256 digests are d5e329f77185eb8bbb6c41b4, ac1be2cc2fb0a0c7eb47dc3e, 4d5335f114cdc74ae236d091 and 9a4099d5ad108cda37a62161, and no two of them match, so the platform is looking at four different files that happen to weigh the same to the byte. Diff the four and the answer is short: 6 keys differ, and all of them belong to the record itself. The generated timestamp. The slug. The four output paths for the final render, the EDL, the FCPXML and the review document. Not one key that describes a shot, a beat, a duration, a source or a score moves between those four files. Four names and four timestamps are the entire difference, and the timestamps run 2026-09-11T16:05:07, 2026-09-11T16:18:24, 2026-09-11T16:24:07 and 2026-09-11T16:25:47.

Read the declared brief in the five files from sequence 02 onward and they are the same document. total_dur reads 78.86 in every one of them. beats reads 12. picks reads 12. The reuse block reads 11 distinct sources drawn from 5 projects and 0 assets used before. gaps is an empty list. The target column sums to 83.0 s in all five. Each file carries 528 keys, and between sequences 03 and 06 the only 6 of them that move are the address and the timestamp. Sequence 02 is the one that really differs, and it differs by 76 keys: it takes conflict_1 from segment 02 of the apocalypse reel instead of segment 03, it takes conflict_2 from segment 02 of the final reel instead of segment 30, and it writes the last two payoff shots in the opposite order. Five files declaring one brief, and the machine still produced five separate cuts of it.

The target column sums to 83.0 s and the cut sums to 78.86 s. The difference is 4.14 s, and it is the one number in the file that no longer agrees with the brief it was generated from. Exactly one row carries it. In sequences 03 through 06 the trimmed row is the final pick, result_2, asked for 10.0 s and cut to 5.86 s. In sequence 02 the same trim of the same shot, the annotated Map Is Dead reel out of TikTok Short-Film, has moved up one place in the order and is labelled result_1 there. The shot did not change. Its position and its label did. The music bed underneath all of this is 206.160978 s of source audio against a 78.86 s cut, which is 2.614 times the runtime, and it is declared identically in all five. The edit decision list at the end of sequence 02 weighs 3,781 bytes while the EDLs for sequences 03 through 06 weigh 3,777. Sequence 01, the cut that trims nothing and matches its targets to the digit, weighs 3,473.

Six cut lists. One window. 31 minutes 31 seconds.

Six files, one film, and the files on disk disagree about how heavy it is. Sequence 01, the 83 second cut, finished at 28,033,437 bytes. Sequence 02 finished at 21,178,047. Sequence 03 at 21,194,430. Sequence 04 at 22,050,360. Sequence 05 at 50,793,258. Sequence 06 at 49,302,140. Sequences 05 and 06 were cut from cut lists that differ in 6 keys, they are the same edit by every measure the files record, and they came off the renderer 1,491,118 bytes apart. Those two are also 2.4 and 2.33 times the weight of sequences 03 and 02 for the identical 78.86 s of output at 1920 by 1080 at 24 fps and -16.0 LUFS. That is not a difference in the film. That is a difference in the encoder, and the cut list cannot see it: each file records 12 shots, 12 durations and one output path, and the byte count of the thing it describes is written down nowhere inside it.

The text block in all six files is 1 line long. hook reads 34 characters and cta reads the same 34, and both of them read Keep more of yourself, for longer. The renderer then made 2 cards for that one line: card_hook.png at 21,812 bytes and card_cta.png at 21,858 bytes, 46 bytes apart, the same sentence framed twice. Now add the paper trail up. Six cut lists weighing 100,808 bytes describe 192,551,672 bytes of finished film, a ratio of 1,910.1 to 1, and 5 of the 6 live in the archive folder while the newest sits in the working folder with its render parts beside it. The probability that five files of identical length hold five different decisions is not zero. It is also not what these files show: they show one decision, four timestamps and two encoders, and a brief that was never edited between them. 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. A cut list that agrees with itself and not with its own output is not signed off. Read the brief, then weigh the file. Keep more of yourself, for longer.