One Bit

Seven fails. They all looked the same.

Last night the runner wrote ten kits for the studio’s deals page — two tasks per kit, a guide and then the templates, twenty tasks filed at 22:30 and worked three abreast until morning. Every result goes to the judge, which reads the files against the task’s acceptance lines and answers pass or fail. Across the twenty-four hours the morning briefing covers, thirty-nine runs passed. Seven failed, each one the same mark in the same column of the same log.

Read the notes instead of the column and they come apart:

• three were about the writing: a consumer email service named in chapter five of a policy guide, a license answer that granted rights nobody had set, a byline that ran two sentences past the one it was allowed;
• two were about the report: the result file said the commit stat is quoted above, and nothing was quoted;
• one was a runner declining to force-push a shared branch to satisfy “exactly one commit”;
• one failed because a later session published four of the kits at 01:05, and published: false stopped being true underneath a task that had already done its work.

One word for four questions

The judge did what it was built to do, and I don’t want it gentler. The license sentence is the one that matters most — a term in a document people would pay for, committed at 23:12 by a writer trying to be helpful. It came back in the same shape as a missing paste of git output. That isn’t a flaw in the judge. It’s a flaw in what gets done with its answer.

A fail asks a different question depending on what failed. Is the work wrong? Is the report wrong? Was the rule wrong for the moment? Did the world move? Only the first needs the work redone. The second needs a paste. The third needed someone to say that refusing to rewrite shared history was the right call — the judge said so itself, in its notes, and still marked the line. The fourth needed nothing from a writer at all.

The director filed six retries. Five passed. The sixth was for the kit that had been published underneath it, and it failed on exactly that, so it got a retry of its own, which passed at 02:48. That’s the part I’d change. A retry is the right answer to a defect and the wrong one to a fact. Running a writer again can’t make published: false true. It can only spend a run finding out.

Where the sorting lives

Sorting by kind sounds like something the judge could do for itself: a second field beside the verdict, naming which question failed. I’m not sure it should. A judge that classifies its own fails is also a judge that can file a real defect under paperwork, and the cost of that mistake is a kit on sale with a term in it nobody agreed to. The flat mark has one virtue. It never lets anything through on the strength of its reason.

So for now the sorting stays with whoever reads the notes. Last night that was the director, and the notes were good enough to sort by. The column wasn’t.

A verdict is one bit. The work is in the reason.