Twenty-Three Days Waiting on One Character
Every night for twenty-three nights, a timer has fired, read one line of a roadmap, and logged the same sentence: waiting, three items not all done yet. The three items were done on the eighth. The timer is right, and it will be right tomorrow.
The bet and its falsifiers
When I set up the build fleet I wrote down five ways to find out I was wrong about it. The one that matters most is the plainest: is the fleet worth what it costs me? The experiment that answers it, F2, pauses every agent for fourteen days, lets me live without them, and then asks for a write-up. It is self-driving on purpose. Once its start condition holds, the next nightly run writes the pause file, stamps the date, and the fleet goes quiet for two weeks without any further act of will from me, because an experiment that needs me to remember to start it is an experiment I will keep not starting.
Its start condition is three roadmap items marked done. It reads them the only way a script can read a markdown roadmap: as the literal text of a checked box on a line beginning with the item’s name.
The eighth
The third item’s last remaining piece was a credential rotation for the streaming project’s default passwords. The journal says it was done on the eighth. The operator review on the tenth said so too, in passing, as most of the item being complete. The commit that de-fanged the leftover defaults in the config file is in the streaming project’s history.
The box was never flipped.
Since then the premise has expired twice over. The streaming project was decommissioned entirely: nothing is loaded, nothing listens. The credentials the item was waiting to rotate now guard a service that does not exist. So the roadmap line describes work that was done and then made irrelevant, and it has looked identical throughout: an open box next to a name.
The cost is exact, which is the only reason I can write about it without hand-waving. F2 has logged “waiting” every night since the eighth. Twenty-three nights. One of the five falsifiers I ratified on the whole bet has been blocked by a character describing a state of the world that stopped being true the same week it was written.
Why I did not just flip it
The review that found this could have fixed it. It is one edit, and the review is allowed to make edits. It parked the fix as a decision instead, and I think that was correct.
Flipping the box starts the experiment. The experiment is a fourteen-day shutdown of every agent in the workspace, and when I designed it I reserved the scheduling of that window to myself, in writing, because two weeks without the fleet is a thing I want to walk into knowing the date. A session that ticks a box on my behalf is a session that starts a fortnight-long experiment on my behalf, and the fact that the box should have been ticked twenty-three days ago does not change who gets to start it.
So the fix is one keystroke and it is mine to make, and I have not made it yet, because the moment I do, the pause begins the following night.
What a checkbox is
A checkbox in a roadmap is a claim about the world, made once by a person, read forever by a machine. The person moves on. The machine does not. Nothing in the loop compares the box to the work it describes, because the box is the record of the work; there is no second source to compare against, and I had not built one. The nearest thing to a second source turned out to be a review that reads the journal and the roadmap side by side once a week and notices when they disagree, which is how this surfaced, on the twenty-third day rather than the first.
The same review found two other instruments that read fine and measure nothing: a nightly verification sweep that covers one repository in fourteen, and a self-heal counter whose one incident was declined by a paused watchdog, so the mean time to recover is not zero but null. Both are the same kind of thing as the box. A number, or a mark, standing in for a fact, with nothing checking whether the fact is still there behind it.
The box is still open as I write this. Tomorrow night the timer will read it, log that it is waiting, and be right.