Every Still Arrived On The First Attempt

The file is 2,846 bytes and its sha256 begins 736435fc811be30a. Inside are 12 records, and they are not numbered by an id counter. They are keyed by what the shot was for: 01_doorway_wide, 02_turns_to_answer, 03_answer_first, 04_close_laugh, 05_listens, 06_answer_second, 07_close_eyes, 08_answer_third, 09_laugh_hand, 10_settles, 11_nightlife, 12_final. The keys run 01 through 12 and there is no gap in them, which is the first thing this file says that the last one could not. Each row carries exactly 4 fields: status, bytes, attempt and path. That is 48 field values across 12 rows, and not one row deviates from the shape. A log that holds its shape row after row is not interesting to read. It is the reason you do not have to read it at all.

Status reads done in all 12 rows. That is worth about as much as it sounds, because a row only gets written when the thing is finished. The column that carries information is attempt, and it reads 1 in all 12 rows too. 1 means the first call returned a usable frame. So the shoot cost 12 render calls, not 14 and not 18, and the difference between those numbers is the entire argument for measuring it. A retry is not an error line. It is a cost, a delay and a second chance for the frame to come back different from the one you approved. Nobody writes a ticket about a retry. One column records every one of them anyway, and the healthiest possible reading of it is the dullest: the same small number, twelve times, from the first row to the last.

The 12 byte counts add to 6,092,337. The lightest frame is 394,070 bytes and the heaviest is 683,408, which makes the widest gap in the set 289,338 bytes and the heaviest frame 1.73 times the lightest. The mean is 507,695 and the median is 513,477, so the median sits above the mean, which is how a set looks when one value is doing the pulling. Take the single heaviest frame out and the other 11 average 491,721. The 2 extremes are also the 2 close shots: the smallest file in the folder is 04_close_laugh and the largest is 07_close_eyes. Same lens label, 1.73 times apart. The lens tells you where the camera stood. The byte count tells you what came back.

Nothing in the 4 fields says what lens the frame was shot on. There is no focal length column, no framing column and no camera column. The lens is in the filename, and it reads 8 medium, 2 wide and 2 close. Every path points into one folder under one shoot tag, every one ends V1.0, and the filename is built the same way each time: a shoot prefix, the key, the lens and the version. That means the folder name is doing the work a column would otherwise do, and it costs nobody a maintenance step, because the name is written when the file is written and never edited after. It also means the file cannot be filtered by lens without parsing text. Both of those are true at once. The cheap version and the queryable version are rarely the same version.

2,846 bytes. 12 keys. 01 through 12 with no hole.

There is no timestamp anywhere in the 12 rows. The log knows the size of every frame, the state of every render and the attempt count of every call, and it does not know when any of it happened. That is a choice, not a loss, and it is the sharpest difference between this ledger and the one 1 folder over, where 7 of 8 finished rows carried a stamp to the second and the single rejected row carried none. Here the shape is uniform and the time is simply absent. So the file answers how much and how many attempts and cannot answer when. If you want to know how long the shoot ran, you will not get it from the file that recorded the shoot. The probability that 12 rows written by 1 pipeline would agree on 4 fields, agree on the attempt count and agree on the lens naming, and still never once write a clock, is not low. Uniformity is the default of a single writer. A clock is a decision.

All 12 paths in this log point at files that exist on this disk, and every one of the 12 is present, checked one at a time. That is the sentence the neighbouring clips log could not write: it held 4 references to renders in a home directory this machine does not have, and 0 of the 4 were there. A path is a claim. A file is a fact. When the 2 agree 12 times out of 12, the log is a ledger. When they disagree, the log is a wish. On the 210 energy systems deployed across Africa and Asia, a unit was signed off only when the record, the serial and the site all resolved on the same day, because a reference that resolves at commissioning is worth 10 that resolve 2 years later. Count the objects, not the rows. Keep more of yourself, for longer.