A Reject Is A Lesson With A File Path

The Sal y Miel asset tree carries 4 folders: 01_Images with 80 files at 34,349,472 bytes, 02_Video with 63 files at 381,987,968 bytes, 03_Finals with 4 files at 251,043,461 bytes and 04_Archive with 42 files at 218,170,938 bytes. Add the 4 counts and you get 189. Add the 4 byte totals and you get 885,551,839. Both sums land exactly on the number the manifest declares for this project, which is 189 files and 845 MB. Now open the last folder. Of its 42 files, 28 carry a _rej_ marker in the filename, 1 is named _split_duplicate, and 13 are survivors weighing 67,038,456 bytes. So 29 of 42 files, or 69.0 per cent of the archive, are the record of what failed. That is the design, and it is worth stating plainly: the folder is not a graveyard that somebody forgot to empty. It is the part of the archive that gets read.

Walk the 7 projects the manifest declares and weigh them. Salvador holds 10 files at 621,044,597 bytes against 592 MB declared. Sal y Miel holds 189 files at 885,551,839 bytes against 845 MB. Apocalypse holds 15 files at 117,934,404 bytes against 113 MB. The H3 dialogue tests hold 29 files at 91,988,334 bytes against 88 MB. Nolan One-Shot holds 2 files at 4,649,866 bytes against 4.4 MB. Songs holds 8 files at 21,935,358 bytes against 21 MB. All 7 projects together hold 268 files and 1,748,443,821 bytes. Six of those 7 counts reconcile. The seventh is Lola: the manifest declares 9 files and 3.3 MB, and the disk holds 15 files at 5,339,423 bytes. That is 6 files more than declared and 2,039,423 bytes more on a decimal reading of MB, or 1,878,815 bytes more on a binary one. One stale line in a manifest is not a crisis. It is the reason the walk gets repeated instead of quoted.

The manifest tells you where to look, and 6 of its names are wrong. It points at finals/, clips/, broll/, stills/, characters/ and archive_rejects/. None of those 6 folders exists under Sal y Miel. The tree that does exist is numbered: 01_Images, 02_Video, 03_Finals, 04_Archive. Every conclusion the manifest draws still holds and every path it prints does not, which is the same defect class as a stale shot count in a status report. The rejects are in 04_Archive. The identity anchors are in 01_Images, which holds 12 numbered reference frames running from 00_hero to 11_violet_glamour, including the 01_direct_gaze frame the manifest names. Read the numbering carefully and 09 is absent, while 01 is used twice: once for 01_body_closeup and once for 01_direct_gaze. Twelve frames, 11 numbers, 1 gap. A naming scheme that skips a number is a naming scheme nobody can audit, which is exactly why the count gets measured from the disk and not read off the index.

Sort the rejects by the b-roll number they belong to and the folder gets sharper. Ten b-numbers survive somewhere: b01, b02, b03, b04, b05, b06, b07, b09, b10 and b13. Three b-numbers appear only as rejects: b08 with 3 rejects and no survivor, b12 with 2 and no survivor, b20 with 2 and no survivor. The 7 numbers that carry both a survivor and its rejects are where the archive earns its keep: b01 holds 3 rejects plus 1 file named _split_duplicate at 6,686,550 bytes, b04 holds 4 rejects, b07 holds 4, b02 holds 3, b05 holds 3, b06 holds 3 and b03 holds 1. Add those up and you are back at the 28 rejects, on 10 b-numbers, inside 1 folder. The 3 numbers with nothing but rejects are the honest part of the record. b08, b12 and b20 were not fixed. They were abandoned, and the folder says so without anybody having to remember it.

42 files in the archive. 29 of them are labelled failures.

The archive is upstream of the work, not downstream of it, and 1 number proves it. Songs holds 8 files at 21,935,358 bytes, and inside that folder sits songs_sal_y_miel.0V1.0.mp3 at exactly 4,524,124 bytes. The cut that shipped this week declares a music bed of 4,524,124 bytes running 206.160978 seconds and trimmed to 83.0 seconds. It is the same file, byte for byte, pulled out of the archive and put back into an edit. That is what archiving is for: the inputs and the finals are kept, and the transient folders of numbered silent clips are not, because the assemble scripts regenerate them from the parts. The same folder holds 2 images from the retired Grok renderer, kept after Grok was withdrawn as a renderer and left on the vision gate only. Files do not get deleted for being out of fashion. They get kept until the reason they were retired stops being obvious, and then they are the reason.

Run the two documents side by side. The failure document carries 25 numbered failures, each with a root cause and a fix: the face guard that rejects a readable human face on the first frame while a body-only still passes, code 97 as a hard stop that kills the whole remaining batch and code 96 as a 30 second wait, the production that paused at 20 of 32 shots because that distinction was ignored, the two-tone skin under white fabric, the 0.72 duplicate and 0.78 mirror correlation thresholds that get checked before any paid vision call, the manifest that said 33 shots while 39 sat on disk, and the $750 cascade at $450 per hour across 25 rounds that came from shipping paths nobody had verified existed. The disk holds 29 labelled rejects. Twenty five named failures against 29 files is not a rounding error, it is 4 rejects with no rule attached yet. On the people side the same discipline reads the same way. Entrade built 200 wood-gasification machines across the UK and Europe, and 200 is a count somebody can walk and check. A EUR 75 million industrial group was restructured, and a counterparty can recount it from its own records. The book published on 8 April 2026 runs 393 pages, and anybody holding it can confirm that in 1 minute. The probability that a reject gets deleted because nobody could name what it taught is not 0, and the 1 check that stops it is a filename. So name the failure, keep the file and write the path. Keep more of yourself, for longer.