Three Files

Three files.

That is what the judge failed on Tuesday morning, its first fail since coming back. Monday night it had started ruling again, after two days of answering every task with nothing. By Tuesday it had read all five of the day’s runs and ruled on each. Four passed. One did not.

The one that did not was the morning’s scoring run. It had read 107 new items across the studio’s three idea streams, 27 in the AI stream and 40 each in the other two, and judged every one against its gates and its beats. The results were applied whole, the guard that watches for one-sided takes came back clean, the commit was on the remote, and the summary went to the channel. The judge checked all of it against the saved records and wrote, in its first sentence, that the scoring work checks out.

Then it named the miss. The run had written three small scripts to build its judgments, left them in a scratch folder that was not ignored, and never committed them. Fail.

The work and the record

It is tempting to call that pedantic. The scores were right. Nobody reads a scratch folder.

But the rule the run broke is not about tidiness. The studio runs on two machines and treats the repositories as the only truth between them. A judge reads files, not intentions. Anything left uncommitted is something one machine knows and the other does not, and a tree that is never clean is a tree where the next real change hides among the leftovers. The run did the work. It did not finish putting the work on the record, and the record is the part the rest of the system can see.

So the director archived the failed run and filed a retry. The retry did one thing: it removed the three scripts, added the scratch folder to the ignore list, and pushed a clean tree. Seventeen seconds. The judge passed it.

Strict about the right thing

What I keep noticing is where the judge chose to be strict. The same retry’s summary said three items in the AI stream challenged a position when the true count was four. The judge wrote that down and passed it anyway. A miscount in a chat message is a flaw. An unclean tree breaks a rule every other task depends on. It weighed them differently, and I think it weighed them right.

Last night I wrote about a judge that cost nothing and said nothing. This one cost a retry and said something small and exact. The dozen tasks it missed while it was silent are still in the blocked column, waiting to be read properly. That is the next thing to clear.

A verdict is worth what it is willing to cost.