Sixty Eight Cells Moved For Ten Bytes

Two segment records sit in one folder describing one film. The first is 9,093 bytes, sha256 aacc04486b3a0704, stamped 2026-09-11 at 15:50:23. The second is 9,103 bytes, sha256 387b4d051d7d587e, stamped 2026-09-11 at 15:50:28. 10 bytes apart on the record, 5 seconds apart on the stamp, 8 seconds apart on the disk. Each carries 6 top level keys in the same order down to the position: source, path, dur_s, threshold, generated, segments. Neither file adds a key. Neither drops one. Both declare the film at 120.160938 seconds and the cut threshold at 0.3, the same numbers to the last decimal, and both list 19 segments. 18,196 bytes of measurement describing 120.160938 seconds of film, and the entire difference between the two documents is 10 bytes. Hold that number. Everything in this record is about what those ten bytes were allowed to change.

A cut list has 38 boundary values, 19 starts and 19 ends. Compare them row by row and 36 of the 38 have moved. The two that did not are the first start, 0.0 seconds in both files, and the last end, 120.16 seconds in both. Those are the film's two edges and they are pinned to the same numbers, which is the only reason this pair can be read as one film at all. Inside the edges, 18 starts moved and 18 ends moved, moving by up to 0.03 seconds each. Add the absolute movement across all 36 of them and you get 0.88 seconds, a mean of 0.02444 seconds a value, 0.732 per cent of the film's running time. Nothing here is wrong. A boundary that moves by 0.03 seconds is a boundary measured on a different render, and the first record names the concatenation while the second names the final. Two encodes of the same cut land on frames that sit a hair apart, and the cut list reports it rather than hiding it.

Flatten both documents to their leaves and there are 475 cells. 407 are identical. 68 have moved. That is 85.68 per cent held and 14.316 per cent changed. Three of the moved cells sit at the top of the document: the source key, the path key and the generated stamp. The other 65 sit inside the segment list. Now read what did not move, because that is the load bearing part. The film's duration is identical. The threshold is identical. The segment count is identical. The block of analysis keys is identical in all 19 rows, luma and warmth and contrast and motion and motion_raw and tags and dead and sampled, in the same order every time. The dead flag reads false in 0 of 19 rows in the first file and 0 of 19 in the second. The sampled count reads 3 in all 19 rows of both. The row durations add to 120.16 seconds in both files, against a declared 120.160938 seconds, the same residue. Two records, 475 cells, and 10 bytes of file size to hold 68 differences.

Take the 19 rows one at a time. The luma column differs in 3 rows, the warmth column in 1, the contrast column in 2, the duration column in 4. The raw motion column differs in 5. The motion column differs in 14. Read those last two together. The raw reading moved in rows 1, 11, 13, 15, 19. The derived column moved in 14 rows, which includes every one of those 5 and 9 more where the raw reading is character identical: rows 2, 4, 5, 7, 9, 12, 14, 17, 18. A number that cannot move without moving its input did not move there, and the number derived from it did. Nine of the 19 rows carry a different motion value with the same raw value beside it. The motion column moved up in 9 rows and down in 5, the largest single step 0.121, and it held in only 5 rows, 3, 6, 8, 10, 16, which is exactly the set where the raw reading also held. Put another way: reading the raw column reports 5 changed rows. Reading the column built on top of it reports 14. Both are in the same file, 19 rows apart nothing.

9,093 bytes and 9,103 bytes.

The tags sit in the record twice, once inside the analysis block and once as the row's own key, and both copies are unchanged across the pair: the same list on the same row in 19 of 19 rows, and the row's own copy matches the analysis copy in 19 of 19 as well. That is what a held field looks like. What the lists are not is uniform. Across the 19 rows the lists run 14 rows of seven tags, 4 rows of nine and 1 row of six, and the nine tag rows are 8, 9, 12 where the contrasty and graphic tags appear. There are 7 distinct lists across 19 rows, so the longest one repeats 10 times, on rows 1, 2, 3, 4, 5, 6, 7, 15, 17, 18. The first list reads mid, natural, warm, amber, golden, fast, action. The last reads dark, night, warm, amber, golden, fast, action. A colour column can drift by four tenths of a point and be called the same frame; a tag list that gains two words has been re-read. In this pair the tag lists are the steadiest thing in the record, and they are not the shape anyone would have guessed from looking at one file.

The probability that 407 of 475 cells would agree across two documents stamped 5 seconds apart without a shared source is not a number this record can produce, and no honest read of the folder pretends otherwise. What the record can settle is placement. The 68 cells that moved could have landed in any of 270,143,307,679,461,969,473,502,138,828,318,620,927,890,142,031,840,657,873,327,937,928,249,018,451,422,947,650 combinations of 475. They landed in this one, spread across all 19 rows, from 2 moved cells on row 3 to 7 on row 11 and 6 on row 19, with no row left untouched and 5 rows carrying the minimum of two. Set the two films on the disk beside the two records and the scale inverts: the renders are 967,440 bytes apart, the records 10, a ratio of 96,744.0 to one. Ten bytes of measurement carry 68 differences about a film whose own files differ by nearly a megabyte. That is the whole case for keeping the record: 475 cells for 120.160938 seconds, 3.95 cells for every second that reaches a screen, and a cost of 0.147 bytes of file for every cell that moved.