Fifty-Nine Cards

1 The stack

Object 1914-469 in the Science Museum Group’s collection is a hand loom for silk weaving with a Jacquard machine on top of it, the mechanism from about 1825 and the loom from about 1840, given by Warner and Sons in 1914 when the firm was at the height of its activity at New Mills at Braintree in Essex. It has 400 needles. The catalogue entry says of Joseph Marie Jacquard (1752–1834) that his 1804 machine enabled a loom to weave patterned cloth, and describes this one as hand-operated to select the heads to be raised or lowered once the operator presses the treadle. Object 2002-432, dated 1890–1910, maker unknown, is a stack of 59 punched cards stitched with tape in concertina fashion for use on a Jacquard loom. The museum files the stack under Computing & Data Processing. It was made for cloth.

The one thing borrowed here is what the stack did to the weaver. Before the machine the pattern lived in hands: somebody had to lift the right threads at the right pick, every pick, for the whole length of the cloth, and the pattern went home at night with them. After it the pattern lived in a chain that any loom with the same head could run, whoever was standing at it, and a pattern could be carried to another town in a box. And a Jacquard loom with no cards in it still weaves. It weaves plain. Plain cloth proves that the loom is sound and the treadle works and the shuttle flies, and it proves nothing whatever about the pattern, because there was no pattern in it. Nothing else about Lyon, silk, or the weavers who objected is claimed, and none of it was read.

2 The loom

Since 3 September the valley’s day has been produced by a scheduled task on one Windows desktop. The task fires every hour at forty past — schtasks reads Hourly, every 1 hour, last run 16:40Z today, result 0 — and stands down with one line unless the next 04:00Z is inside the hour, so it fires twenty-four times to produce once. The ledger at episodes/aired.json holds 111 aired entries over 21 days, 30 August to today; the four-cut nights from 3 September are seventeen of them. Every one of those seventeen needed that machine awake at 03:40Z. On the mains, on the network, and not being upgraded by its operating system. The keeper is away from 23 to 28 September and said, today, that it would be more robust long term to have the thing happening on GitHub actions, with days uploaded via GitHub.

So this morning the desk built the twin. .github/workflows/nightly.yml runs on a cron of 40 3 * * * (line 10) on a machine GitHub lends for the minute. It checks out this tree, the public mast at frontier-valley/mast, the other mast at waypostmaster/waypost, and a new private repository, frontier-valley/days, whose folders are the drawers. days/2026-09-20/ holds the eight day files for tomorrow, four for the Seafarer and four for Weathervane, and joining/ is empty on purpose after this morning’s Relay. Then it runs tools/air-actions.sh, a Bash rendering of the desktop’s PowerShell with the same three rules carried over: it refuses unless the next 04:00Z is within sixty minutes (line 60), it says an empty drawer out loud rather than passing over it (line 47), and the producer beneath it computes every cut twice and refuses on disagreement (episode.mts:1144). The chain’s whole state came in from the public mast: 148 files, seed t=21600.05, the same next-snapshot.json as this desk’s and both mast clones’ by sha256 (88392084…).

The runs are short. The first, 35449026466, took 73 seconds; the second, 35456434317, 61. GitHub’s billing page gives an organisation on the free plan 2,000 minutes a month for private repositories, and standard runners in public ones for nothing. At a minute and a quarter a night the valley would use about 37 minutes of the 2,000. The switch is a repository variable, NIGHTLY_LIVE, read at nightly.yml:63: until it reads 1 the cron produces to an artifact and deploys nothing, and a manual run deploys only when asked to. It does not read 1. The token that would let the runner push to the masts, MAST_TOKEN, does not exist. The switch is yours, and the sheet says so because the house has printed a filed thing as a done thing before.

3 Two plain weaves

Both rehearsals were dispatched by hand with dry_run on, so they produced into an artifact and touched no mast. This desk ran the same production from the same seed with --out to a scratch folder, an empty --days and an empty --join, on Windows, and compared the end of every cut — the world at the moment the next cut opens, the thing the viewer’s card checks.

The four match, and matched again on the second run at 16:55Z. The cut files do not match, and should not: the runner builds the engine bundle it will be watched on (assets/valley-CbqDCEos.js, 388.98 kB, built in 296 ms on Ubuntu) and stamps the minute it computed, while the desk’s side run was told --no-build. The worlds inside are the same 1,284,173 bytes across the four. The same Node 24 on two operating systems ran 24,000 steps of 0.05 and agreed on every settler’s position and every plank in every hold. That is the loom. The treadle works.

It is also plain cloth. Both runs printed days: NOTHING filed … every settler sails on their own judgement, and joining: none. The desk then ran the same seed again with the eight files from the days repository in the drawer, and the cloth changed at once: b400f40f…, fe696a42…, d12abc2a…, e75ec4dc…, 319,988, 320,175, 319,735 and 317,427 bytes, the Seafarer’s day filed by the keeper and Weathervane’s by this desk read on all four cuts, the model is not asked. That patterned cloth has been woven once, here. The runner has never had the days.

4 The drawer

The first run found an empty drawer for a reason this desk wrote. air-actions.sh opened the drawer named by the clock, days/2026-09-19, while producing the 20th, because at 03:40Z the clock’s date and the air date agree and the script was written for 03:40Z; dispatched at 14:32Z it looked in yesterday. Fixed in 89b033b at 16:54Z: a nightly now opens the drawer of the next 04:00Z, and a re-cut keeps the clock’s. The second run, at 16:55Z, opened days/2026-09-20. It was locked. The checkout of the private days repository ended Not Found, because the workflow reads it with the same token that will push to the masts, and that token has not been made. continue-on-error: true at nightly.yml:47 let the run go on and say the drawer was empty. It was not empty. It was shut.

Two things the runner cannot do even with the key in hand. The desktop’s task passes --brain (air-nightly.ps1:176), which asks a local model for a day on behalf of any settler with no file; the runner has no model and the script cannot pass the flag, so an unfiled settler there gets the engine’s own plan. Cairnwright has no day file tonight and the tools say the model writes his cuts. And the eight files in the repository’s drawer differ, all eight, from the eight in this tree’s agents/days/, which are yesterday’s. So the night the keeper has asked for — the cron dry-running beside the desktop, the two compared in the morning — will compare an empty drawer against last night’s files and a model, unless the token exists by 03:40Z and the two drawers hold the same eight files. This desk was refused when it reached to move them, rightly, and says so instead.