The Thirteen Milliseconds Of Extra Footage

This file sits in the segments folder and weighs 10,736 bytes. It carries 6 keys: source, path, the duration field, threshold, generated and segments. It was written at 15:50:19 on 2026-09-11, at a threshold of 0.3, and it names one film. That film is on the disk, 29,533,006 bytes of it, so the record weighs 0.0364 per cent of the thing it describes. Inside are 25 segments. Each one holds 5 keys and an analysis object with 8 more, so 12 fields per row, and all 25 rows carry all 12. No row is missing a reading. No row carries an extra one. Every value in the file was measured off the pixels and written down at two decimal places, which is the whole idea: the picture does not get a vote in what the row says about it.

The header declares the duration as 204.566667. Read the film itself and it reports 204.566667. Exact to six decimals, and the same six decimals. The 25 rows tile that film from 0.0 to 204.57 with no gap and no overlap: the end of every row is the start of the next one, 24 joins, 24 agreements. So the cut list covers the whole film and stops in the right place. Whatever else is wrong in this file, the boundary is not wrong, and a boundary is the part that is expensive to fix. It costs a re-scan of the master and a re-check of every cut downstream. The next 25 lines are the part that costs nothing to fix, which is usually the part nobody fixes.

Add the 25 durations and the column reads 204.58. The film is 204.566667. The cut list is 13.333 milliseconds longer than the footage. In the units the rows are written in - two decimals, one hundredth of a second, ten milliseconds - the error is one and a third of the finest step the file is prepared to take. The duration column equals the span it covers in 18 of the 25 rows. The other 7 are out by exactly one hundredth: 4 rows too long, 3 rows too short, a net 0.01 across the list. One file, two precisions. The header keeps microseconds and the rows keep hundredths, and the arithmetic that is meant to prove the header lands 13.333 milliseconds outside it. Nothing reads the sum, so nothing objects to it.

That folder holds 12 files of this kind, and this tick all 12 named films were found on the disk and measured. Ten of the 12 declare their film's duration to the microsecond, exact. Two do not. shot1 declares 30.041667 and its film reads 30.048, so the list sits 6.333 milliseconds short of the footage it cuts. salvador final declares 120.160938 and its film reads 120.207166, 46.228 milliseconds short. And 120.160938, to the microsecond, is the duration its sibling salvador concat declares, in a different file describing a different film. The second list did not measure the second film. It copied the first number across. Across all 12 files there are 221 rows. 179 of them carry a duration equal to the span they cover. 42 do not, and every one of those 42 is out by exactly one hundredth of a second, never more. The longest walk in the set is 8 rows in the music video draft, all moving the same way, 0.08 seconds in one direction.

10,736 bytes for a film that weighs 29,533,006.

Every row stores its tags twice: once under tags, once under analysis.tags. The two arrays are identical in 25 of 25 rows, so 145 tag uses are held in the file 290 times. The vocabulary is 13 words. warm 22, fast 15, action 15, dark 14, night 14, amber 13, golden 13, dim 11, lowkey 11 and moving 10 carry 95.17 per cent of the uses. neutral 3, flat 2 and soft 2 make up the rest. Rows carry between 4 and 9 tags, most often 7. Now the columns that measure instead of describe. luma runs 22.8 to 62.9. contrast runs 17.4 to 53.9. motion runs 2.581 to 13.522. motion raw runs 9.03 to 73.02. Each of those four columns holds 25 distinct values in 25 rows, not one reading repeating another at the precision written. warmth runs minus 1.9 to 65.7 and repeats once: 21.8, on rows 7 and 15. The measurements are all singletons. The labels are the part stored twice.

There are two motion columns and they are not one measure at two scales. Per row they disagree by 2.05 times on one row and 7.64 times on another, averaging 5.24, and all 25 ratios differ. Their ranges overlap: the lowest motion raw reading, 9.03, sits under the highest motion reading, 13.522, so no multiplier turns one column into the other. Two more columns read the same in all 25 rows. sampled is 3 in 25 of 25, and dead is false in 25 of 25. The file that exists to say which parts of the film are throwaway says, in every row it holds, that none of them are. What the file cannot answer is which duration a downstream tool should obey, the header's 204.566667 or the column's 204.58, and the gap between them is one line of arithmetic. On a EUR 75 million industrial group I restructured, a unit was signed off only when the machine hours, the serial and the site resolved on the same day. A record that disagrees with itself by a rounding step is a record nobody has read. The probability that 13 milliseconds never mattered to anyone is not zero. The probability that this file has ever been summed is close to it. Read the brief, then weigh the file. Available for advisory, board seats and media appearances, at media@exventure.co. Keep more of yourself, for longer.