Two Ledgers One Second Apart

Two JSON files sit in one folder named segments. The first weighs 1,488 bytes and the second 1,531, 3,019 between them, 43 bytes apart, which is 2.809 per cent of the larger file. They carry different hashes and different names, and they are the two lightest files in a folder of 12 that weighs 96,461 bytes, so they are 16.667 per cent of the records and 3.13 per cent of the bytes. Each file stamps the moment it was made: 2026-09-11T15:50:29 for the first, 2026-09-11T15:50:30 for the second, 1.0 second apart on 2026-09-11. The disk says something else. The first file was written at 16:04:05 and the second at 16:04:06, 1.038 seconds apart, and both land 00:13:36 after the timestamp each carries inside it. The whole folder's 12 files fall inside 59.481 seconds of 16:04:05, so the mtimes are one arrival, not 12 separate saves. The only clock either file writes down is the one it disagrees with.

Both files carry the same 6 keys at the top: source, path, dur_s, threshold, generated, segments. Both declare 3 segments, the same threshold of 0.3, and the same length of 30.041667 seconds. Read the boundaries and the agreement is exact to the hundredth of a second in all six rows: 0.0 to 10.01, 10.01 to 20.03, 20.03 to 30.04. Each row stores 5 keys and the analysis block inside it stores 8. The stored length of every segment is 10.01 seconds, written 6 times, and the 3 of them add to 30.03 against a declared 30.041667: 0.011667 seconds missing. A third of 30.041667 is 10.013889, so each boundary is the rounded running total of three equal thirds, and the last one lands 0.001667 seconds short of the length the file itself declares. Neither file carries a hash, a date stamp or a checksum of any kind: the key list of both is 6 fields long and none of them names the bytes of the clip. The tag array is stored twice in every row, once inside the analysis block and once beside it, and the two copies are identical in all 6 rows.

Each file names one clip, and both clips are where the ledgers say they are. The first source weighs 44,714,996 bytes and the second 52,812,176, 8,097,180 more, 18.108 per cent heavier. Both are 1280 by 720 h264 at 24 frames a second, both hold 721 frames, and 721 frames at 24 a second is 30.041667 seconds, which is exactly the number both ledgers declare. The containers do not agree with each other: the first reports 30.041667 seconds and the second reports 30.048000, a gap of 0.006333 seconds, while both video streams read 30.041667. That is 6.333 milliseconds, under a fifth of a frame, sitting in the wrapper of one file and nowhere in its own picture. The clips were written on 2026-08-31 at 19:21:07 and 19:21:13, 11 days before the ledgers that describe them, and in the reverse order: the file logged last on 2026-09-11 was saved first on disk. The folder holding the clips holds four and 116,554,146 bytes; these two are 97,527,172 of it, 83.68 per cent, and no ledger in the segments folder names the other 19,026,974. Put the record beside the thing recorded: 32,304.462 bytes of video for every byte of ledger, and the ledger is 0.003096 per cent of what it describes.

Both files divide the same 30.04 seconds into three, and both label the first third the same way: dim, lowkey, cool, blue, moving, 5 tags, identical in both files. Every number behind those 5 tags differs. The luma reads 57.3 against 53.1, the warmth -32.8 against -34.0, the contrast 41.1 against 38.6, the motion 4.169 against 4.275, so the two files agree on the words and on none of the readings the words are supposed to summarise. After the first third the words separate. The first file keeps 5 tags in every segment and calls the last third dark, night, cool, fast, action. The second file grows to 6 tags in the second and third segments, calls the last third dim, lowkey, cool, blue, fast, action, and never uses dark or night at all. Across the six rows the two files draw on 9 distinct tags to describe what is, by their own field list, a light reading, a colour reading, a contrast reading and a movement reading. Four of those tags belong to one file only: dark, night and moving on one side, dim, lowkey, fast and action on the other. The two ledgers agree completely on exactly 1 of 3 segments.

1,488 and 1,531 bytes.

The light runs in opposite directions. The first file reads 57.3, then 28.7, then 21.3: down 36.0 across the clip, 62.83 per cent darker at the end than at the start, and every step is a fall. The second reads 53.1, then 50.6, then 75.4: up 22.3, 42.0 per cent, and the last third carries almost all of it. Subtract one from the other and the gaps are 4.2, -21.9 and -54.1. In the last third the second file is 3.5399 times brighter than the first. Motion moves the same way: the first file's raw reading runs 26.71, 19.29, 74.59, the second 27.39, 40.07, 40.18. The first file sits under the second by 0.68 and 20.78 in the first two thirds, and over it by 34.41 in the last. The comparison turns over once, in the last third, and only there. Contrast behaves differently again: the first file drops 41.1, 22.6, 21.4, the second stays flat at 38.6, 43.7, 41.5. Every row stores motion twice, once as motion and once as motion_raw, and the second reading is the first multiplied by 6.4059 to 6.4122 across all six rows, a spread of 0.0983 per cent. Two names, one measurement.

Every row records that it sampled 3 frames and that nothing in the segment read dead. A segment of 10.013889 seconds at 24 frames a second is 240.333 frames, so 3 samples is one frame in every 80.111, 1.248 per cent of the picture. That is the whole instrument: 3 frames decide the words for a third of a clip, and the words are then compared between two files as if they described the same thing. Nothing in either file says what 0.3 thresholds, what luma units the readings are in, or which frames were picked. Neither file names the other. Neither carries a hash, so the only tie between a ledger and its clip is the filename, and the filename reformats the version stamp: 0V1.0 on the video becomes 0v1_0 on the ledger. Read the folder by byte weight and the probability a reader lands on either ledger is 3.13 per cent: 3,019 bytes of record inside 96,461, sorting the two smallest files in a folder of 12. Both files are pure ASCII, indented one space, and neither ends with a newline. Whatever happened in those 30 seconds was written down in 3,019 bytes, once, one second apart, and the second file cannot be told from the first without opening it.