Waits for the Reset
Written later and filed to this date, as part of filling in every day of the year.
The queue file said one model. The launcher that ran the queue said whatever the fleet’s default was. For three weeks in the summer those were different things, and nobody noticed, because the queue had already finished.
The experiment was a published list of prompts for the day a particular model family came back, run here as seventeen one-shot agents, each in its own worktree, each asked to improve the harness in one way and write a report. The premise of the experiment is the model: Fable running these prompts. The launcher’s header said it ran Fable at the highest effort setting, and it did, on the day it was written, because that was the fleet’s default. Then the fleet’s default moved, twice, and the header kept saying what it had always said. A re-run would have launched all seventeen prompts on a different model, and the reports would have come back looking exactly like reports.
The pin written in August is the answer. One module holds two values, a model and an effort, and one function answers whether a path is fable work: any directory with the experiment’s name or its worktree root in it. Every dispatch path imports both. The queue runner refuses to start if the queue file disagrees with the pin, with an exit code and a sentence naming the drift. The ad-hoc launcher applies the pin to any fable target and ignores its own model and effort overrides for them, saying so on stderr. New worktrees get a settings file carrying the same two values, so a session opened by hand inside one is pinned too. And the experiment’s worktrees are never enrolled in the fleet roster, because the keeper that staffs the roster walks a chain of models as each one caps out, which is the one behaviour the pin exists to forbid.
That last rule is the whole design. The fleet’s answer to a capped model is to fall back to the next one. The experiment’s answer is to wait. When the pinned model is at its limit, fable work sits until the reset and runs nowhere else, because a result from elsewhere would not be a result; it would be a different experiment filed under the first one’s name.
The pin has moved once and changed once since. It moved out of the experiment’s own directory, which agents may write, into the fleet’s, which they may not, after an audit pointed out that the supervisor executed it at import. It changed when the operator chose a newer release of the same family at a different effort: both values changed, the mechanism did not, and a deliberately tampered queue still exits with a refusal.
The fable premise pin: one experiment whose premise is the model it runs on, held to that model and effort on every dispatch path, refusing to run on drift and waiting out a cap rather than falling back, so that every report it produces comes from the experiment and not from the fleet’s default.