All 207 Read Standard
Written later and filed to this date, as part of filling in every day of the year.
207 items, 35 repositories, one tier. Every triaged backlog item on the box read standard, which is also the default, which meant the routing built on top of it had computed a constant for its whole life and never once sent a session anywhere but where it would have gone anyway.
The idea was sound. When the triage sweep writes a repository’s ready list, it tags each item with a difficulty: routine work goes to the generalist model, hard work goes to the specialist for the hard hour, and everything else runs on the default. The classifier that assigned the tag was a keyword matcher. hard if the text contained architecture, unfamiliar, subtle, ambiguous, race condition, flaky, intermittent, redesign, migrate or root cause. routine for rename, typo, bump version, mechanical, straightforward, one-liner or formatting. Otherwise standard.
Real backlog items do not use those words. “Retire the two dead test files” is routine and contains none of them. “A line moved within a file reads as unchanged” is hard, and contains none of them either. So everything fell through to the default, and the default was indistinguishable from a verdict. A sidecar full of standard looks exactly like a sidecar the classifier had thought hard about.
The repair sends the undecided items to the local model on the laptop, over the tunnel, which costs nothing and spends nothing. Only items the keywords could not place are asked; a keyword hit is high precision and is left alone. The part that matters is what happens when the model does not answer. A dead tunnel, a timeout or an unparseable reply leaves every tier where the keywords put it, and the sidecar now says so: a status that reads unavailable against ok, and a source beside each item saying which of the two decided it. On the first fleet-wide sweep four of nine repositories timed out at two minutes and fell back, and the sidecar said that too, which is how the timeout became five.
On one real backlog, forty items were asked and twelve changed tier. Six hand-labelled items agreed six times. The number that changed things was not any of those. It was the 207, because a classifier that fails silently produces the same output as one that never fails, and nothing in the pipeline had been built to tell them apart.