Forty-Nine Rows, Most of Them Not Mine
Written later and filed to this date, as part of filling in every day of the year.
On a Monday at the end of September the Attention list on my dashboard held forty-nine rows, and it was growing faster than I could read it. Most of the rows were findings from a code reviewer, each a defect in one repository with the files named, and each was something an agent could have fixed. They were in my list because the list had no second reader. Everything that was not obviously done went to the human, and the human was the one thing on the box that cannot run on a timer.
What shipped that day reads the same rows the list shows, every half hour, and sorts them in a fixed order, and the order is the point. Code goes first. A set of hard gates decides, with no model in the room, what is mine by rule: money, trading, publishing, deploying, sending anything on my behalf, physical actuation, root, a patch to the guard or to the fleet’s own machinery, secrets, a ratified falsifier, and any item in a repository on the gated list. Word rules read the item’s own text for those classes, and they read the task an agent would be handed as well, so an item that never says “deploy” but whose fix would deploy is still mine. A hit is final. No verdict, confidence or approval from any model can move it.
A model goes second, and only classifies. One batched read-only call over the items whose text changed since the last run, answering for each: implement it, close it, or needs me, with a digest card either way. Code then decides whether to believe it. Implement only above a confidence of 0.8, close only above 0.85 and only with proof naming what made the item moot, either only in a repository the fleet may staff, and only inside caps: two dispatches a run and six a day, eight closes a run and twenty-four a day. Anything that fails a check degrades to mine, never upward. What survives is handed to a build agent with a brief, or closed through the source’s own write path, and the rest becomes a ranked digest of one-tap cards, which is what the dashboard opens on.
The forty-nine rows were not forty-nine decisions. They were forty-nine items that had never been asked whether they needed one.
The Attention lane: the queue of everything on my dashboard that is waiting on a human, read every half hour by code first, which settles the hard gates and is final, and by a model second, whose classification code accepts only above a confidence bar, with proof, within caps. What reaches me is what only I can decide, and the lane exists to make that sentence true mechanically rather than by hope.