Four Bytes Between Two Readings

Two records sit in one folder, both plain JSON, both describing a single short film. The concat pass weighs 4,091 bytes, sha256 99b7989ba3cea266. The annotated pass weighs 4,095 bytes, sha256 6f2f88b3ac5141d3. 4 bytes separate the two files. Each declares 6 top level keys, and the key order is identical on both: source, path, dur_s, threshold, generated, segments. Five of those keys moved between the pair and one held. The threshold reads 0.3 in both. The stamps read 15:50:28 and 15:50:29, one second apart. The disk disagrees. The annotated record was written at 16:05:03.744 and the concat record at 16:05:04.997, so the file stamped one second later landed 1.2535 seconds earlier on disk. Two orderings, one folder, 4 bytes between them, and the order flips depending on whether you read the stamp inside the file or the mtime on it.

Flatten both documents into their smallest parts. Each row carries three boundary values and eight analysis values, 8 rows, so the pair aligns into 88 cells apiece. 54 of those cells are different across the two readings, 34 are character identical, which is 61.364 per cent moved and 38.64 per cent held. Read the columns instead of the rows and the split is uneven. Luma moves in 8 of 8 rows, contrast in 8, motion in 8, warmth in 7, raw motion in 7, the start time in 7 and the end time in 8. The duration column moves in exactly 1 row. The dead flag moves in none, and the sample count moves in none. Two analysers walked the same 8 shots and disagreed on 54 readings while agreeing on the two columns that are not readings at all.

Lay the 8 segment boundaries out as a single ladder and you get 16 values, 8 starts and 8 ends. 15 of the 16 moved. The one that held is the first start, 0.0, which is where any cut begins. Everything downstream shifted, and the shift is small: 0.48 seconds of total absolute movement across the whole 16 value grid, a mean of 0.032 seconds a boundary and a largest single step of 0.04 seconds. Set the two ladders side by side and the walk is visible. Starts run 0.0, 3.4, 8.0, 13.1, 18.97, 24.83, 30.7, 34.1 against 0.0, 3.37, 7.97, 13.07, 18.93, 24.8, 30.67, 34.07. Ends run 3.4, 8.0, 13.1, 18.97, 24.83, 30.7, 34.1, 39.97 against 3.37, 7.97, 13.07, 18.93, 24.8, 30.67, 34.07, 39.93. Every end is shorter in the second reading. Twelve of the fifteen moves cut the film back by roughly a thirtieth of a second and three round differently, which is what separates two passes of the same detector over two masters rather than two different edits.

The declared running time reads 39.966667 seconds in the concat record and 39.933333 in the annotated one, a difference of 0.033334 seconds. That is one frame at thirty frames a second. Both records split the film into 8 segments and both declare the same cut threshold, 0.3. The duration column tells you where the arithmetic strains. The concat durations are 3.4, 4.6, 5.1, 5.87, 5.87, 5.87, 3.4, 5.87, summing to 39.98 against a declared 39.966667. The annotated durations are 3.37, 4.6, 5.1, 5.87, 5.87, 5.87, 3.4, 5.87, summing to 39.95 against a declared 39.933333. Seven of the eight durations are identical between the passes, character for character, while their own start and end values moved. So the same segment length is carried on a shifted grid, and the sum of the parts overshoots the stated whole in both files. Each file agrees with itself about the shot count and disagrees with itself about the total.

4,091 bytes, sha256 99b7989ba3cea266.

Every row stores its tags twice, once inside the analysis block and once on the row itself. 8 rows, two copies, 16 tag lists in each file, and the two copies agree in every row of both documents. Across the pair the tag lists are identical row for row, every one of the 8 shots carrying the same labels in both readings. What the queue said about that uniformity does not survive the file. The lists are not 8 copies of one list. They run 7, 8, 7, 8, 4, 7, 9, 7 labels long, that is 8 distinct lists across 8 rows, no list repeating, drawn from 21 distinct tags. So the labels are stable where the measurements are not: the taxonomy held on all 8 rows while 54 of the 88 cells around it moved. The pass that moved the numbers did not touch the words, because the words are assigned by thresholds and the numbers are read off the frame.

The two records describe two masters that both still sit on disk, side by side, in a folder of 19 files. The concat master weighs 9,460,434 bytes and the annotated master 13,353,644, a difference of 3,893,210 bytes. The two JSON records that describe them sit 4 bytes apart. Set those figures against each other and 3,893,210 bytes of finished film are narrated by 4 bytes of text, a ratio of 973,302 to one. What each record costs: 511.375 bytes a row on one side, 511.875 on the other, for a whole 8 shot cut that runs a little under 40 seconds. The brightness range reads 6.6 to 152.0 in one pass and 9.8 to 150.2 in the other, and the closing shot reads 0.036 motion against 0.546, a factor of 15.17. The probability that a re-render trips the same detector into the same 88 cells is not zero and is not the point. The point is that 4 bytes of file size carry 54 changed readings, and only the reader who opens both folders ever learns which.