The Caption Uses Nine Per Cent Of The Allowance

Two publish specs, 2,657 bytes and 2,798 bytes, written the same minute on 2026-09-11. The first carries sha256 46d1045e56c191c9, the second d7823732089e0ea2. Read their keys and there are 13 in the first and 13 in the second, in the same order, down to the position: slug, title, theme, generated, asset, asset_exists, duration_s, aspect, canvas, sources_used, source_count, posts, rails. The two documents are named agency and craft. That is the only place the difference lives. Both declare 6.0 seconds, 9:16 on a 1080x1920 canvas, 6 sources and the same 6 source paths in the same order, 3 posting rails with the same 3 names. Their generated stamps sit 16:18:18 and 16:24:01, 5 minutes 43 seconds apart. 5,455 bytes between them, and all 141 bytes of the difference are prose.

Now count what the prose is measured against. The 13 keys hold 1,107 characters of caption across 6 posts, and those posts declare 18,800 characters of allowance. 5.89 per cent used. Split it by platform and the numbers separate: the two Instagram captions take 9.61 per cent of their pair of 2,200 character budgets, the two TikTok captions 6.52 per cent, the two YouTube captions 3.97 per cent of a pair of 5,000 character budgets. The heaviest single caption is 233 characters of 2,200, which is 10.59 per cent. The lightest is 122 characters of 5,000, which is 3.54 per cent. The smallest caption by characters is 122. Set the whole thing against the film and the allowance is 3,133.33 characters for every one of its 6.0 seconds, of which 184.5 characters a second get used. 17,693 characters, 94.11 per cent of the allowance, are never spoken.

Open the second spec's Instagram caption and it runs 5 lines. Line two reads Cut from what was already shot. Nothing new was filmed. Line three repeats the claim in the other vocabulary: Cut from material that already existed. Same assertion, two sentences, one caption, four lines apart in a 233 character field. Line three is not new writing either. It is the first spec's Instagram line, carried across verbatim. Check the pair line by line: Instagram shares 2 of its 5 lines, YouTube 2 of 5, TikTok 1 of 4. 5 shared lines across the 14 in each document. Two themes, 2 slugs, 2 assets, and one of the two captions is still quoting the other one.

The same 6 paths appear in both files and all 6 of them resolve on the disk: 469,915,061 bytes of source material named by 5,455 bytes of spec. That is 0.0012 per cent of the weight of the thing it describes. Divide by the runtime and the sheet names 78,319,176.8 bytes of source for every second of finished film. The two finished films exist: 2,352,460 bytes and 2,138,653 bytes, 4,491,113 between them, a spread of 1.1 between the two cuts of the same 6 seconds. The record is 0.1215 per cent of those two files. No column in it holds a resolution, a frame rate, a bitrate, a duration of any individual clip, or the name of any shot. It holds a path list and a caption budget, and it is correct on both.

2,657 bytes and 2,798 bytes.

The rails block is the same 3 entries in both documents, in the same order, with the same statuses. Composio MCP: not wired, OAuth required, the keys rejected. Postiz agent: not installed. Manual: ready now. 2 of the 3 rails are recorded dead in their own words, and the remaining one is a person opening the asset and pasting the caption. So the block that decides whether any of this reaches a feed says this: 469,915,061 bytes of source, 6.0 seconds of film, 6 captions and 18,800 characters of allowance, carried by one ready rail that is a hand on a keyboard. A spec that names two dead rails is not hiding anything. It is telling you the work is manual, and it is right.

I have audited records like this pair in other rooms, and the shape is always the same. In a EUR 75 million industrial group restructuring, every figure had to tie to a document before a lender read the next page, and the discipline was never that the numbers were wrong. Across 210 energy systems deployed across Africa and Asia, the commissioning sheets and the financial sheets counted different things and both were defensible alone. Entrade built 200 wood-gasification machines across the UK and Europe on the same pattern. Nobody in any of those rooms was lying. Every one of them was working from a record nobody reconciled. The probability that these six captions were written to fit their budgets is not zero, and this pair cannot settle it either way. The budgets are generous to the point of being decorative: 3,133.33 characters of allowance for every second of film. What the pair can settle is the arithmetic. 5,455 bytes. 13 keys each. 1,107 characters of caption against 18,800 of allowance. 9.61 per cent on Instagram, 6.52 on TikTok, 3.97 on YouTube. 5 lines shared between two documents that claim two different themes. 2 of 3 rails recorded dead. Count what the record measures, then count what it repeats, then count what it leaves to a person who has to remember. The distance between those three numbers is the audit, and it costs one read of the folder.