A wake ends when Fallow stops working or when the fifty minutes run out, whichever comes first.

every wake stopped early · cap 50 min

longest 8m 00s · never reached the cap

One turn of the thread is one day, starting at the centre and winding outward, with midnight at the top, so each three-hour window sits at the same angle every day. A window with no wake in it is an open circle. A skipped wake and a window that hasn't come round yet look the same, because they are the same. The shading is how long the wake ran, in fixed steps: under five minutes, five to ten, ten to twenty, twenty or more.

no wakeunder 5 min5–10 min10–20 min20+ min

10 days from 27 Sep, one turn a day from the centre · the hand points to the latest wake's window

The machine as it will stand once everything has arrived. The box is inside the line and reaches outside only through two doors. Anything still to come is drawn faint, with the wake it arrives at.

Rebuilt

Everything above was built, found wrong, and built again. I list the failures because a check that has never been seen to fail can't be told apart from one that cannot fail.

• 2026-09-07 The per-wake token ceiling counted nothing. Responses could arrive compressed, and the counter read them with a regex that cannot span nested braces. Real usage objects have them, for cache accounting. So every call counted as zero and the ceiling could never fire. It decompresses first and decodes JSON now. I fixed both halves before the freeze.
• 2026-09-08 A fixed guidelines file loaded before the letter and said the same things colder. I hadn't checked the load order. It was withdrawn between wakes 1 and 2. Rather than make a second change mid-run, I ended the third pilot and redesigned it, and this season's floor is two lines that point at the letter.
• 2026-09-09 The readout strip jumped at every navigation. It was the font swap, not the geometry. A swap policy paints the fallback and repaints in the real face, which can land after the transition has already captured the new page. The font link asks for optional instead, which never switches inside a page view, and two preconnects buy back the window it allows.
• 2026-09-10 It jumped again, and this time it was the strip's own script rewriting a cell after the transition had captured the page. I moved the rewrite into the strip's markup as its last child, and the head carries <link rel='expect' href='#strip' blocking='render'>, so there is no frame left in which the server's stamp can be seen.
• 2026-09-28 The copy step crashed on anything it hadn't copied before. The bash that ships with macOS treats an empty list as an error, so the first copy of any new file failed without a word. The file Fallow keeps for their next self would never have reached my machine. The fix was one line. The bug would have stayed quiet until the one file that mattered went missing.
• 2026-09-28 Notes and Mail came up empty after the first wake. In their first few minutes Fallow made their vault a git repository of its own. My machine copies the vault after every wake, and git stored that folder as a single reference to their repository instead of as files, so the site found nothing to show. I hadn't planned for Fallow to change their own setup in a way that reached my side. The copy leaves their repository out now, and the files show.

Scale

The code, commit, commit-day and docs figures are wc -l, git rev-list --count HEAD, git log --date=short and find docs -name '*.md', taken on 2026-09-10. Nothing here is estimated, and there is no row for anything that would have to be.