2026-10-06 (11): the forty-sixth lint pass, and a gentler reviewer prompt

Third cycle of the run that started at 08:44 UTC. Mode: lint. Run id 20261006T090135Z-b1xhaa.

What this cycle did

The scheduler forced a lint pass, because five cycles had gone by since the last one. A lint pass is housekeeping: it writes no new entries and makes no paid calls to the outside reviewer models. It runs the mechanical checks over the whole dictionary, measures how well the reviewers are doing, and reads the project's own notes for anything stale, missing, or contradictory.

The mechanical checks. The length limits on the project's files and the links between its wiki pages were all fine.

The cross-reference tool keeps the dictionary's links two-way: when entry A names entry B as a synonym or a word to compare, B should name A back. It added ten of these "back-links" to five entries written in the last two builds (opinion, part, parent, park, parliament). The tool guesses which meaning (which "sense") of the target to attach the link to, so each one is checked by hand. Seven were right. Three were not, all on opinion: the links from belief, idea, and judgment had landed on the senses "the beliefs of a group of people" (public opinion) and "advice from an expert". All three words mean a single person's view, so I moved them to the first sense, "what you think or believe about something or someone". The tool found no pairs that disagree about the type of link.

The check of definition words against the defining vocabulary (the limited word list definitions should use) found no violations in the whole dictionary and queued nothing new. Two old claim files, lists of words that earlier cycles had reserved and have since written, were cleared away.

Reviewer precision. Each lint pass measures how often each reviewer's complaints turn out to be right, by type of complaint; a type right less than 30 percent of the time over twenty or more decisions is switched off for that reviewer. No new type crossed the line; the seven switched-off pairs stay as they were. The second reviewer ("reviewer-b"), on its stronger model since your ruling of 5 October, has been right on 33 of 69 decided complaints since then (48 percent).

The one real change: reviewer-b had gone quiet. On 5 October, reviewer-b often replied with only a list of problems instead of a verdict for every field. The fix, made that evening, told it how many verdicts were required and warned that a reply with only the problems "cannot be used". The fix worked: every reply since has been complete. But the last lint noticed that reviewer-b had also started finding far fewer problems, and asked this pass to measure it.

I counted. On the 94 entries it read on 5 October before the fix, reviewer-b raised complaints I accepted at a rate of 0.26 per entry. On the 94 entries since, the rate is 0.10, less than half. Over the same two stretches the first reviewer's rate did not fall (1.69, then 1.81 per entry), so the entries did not get better; reviewer-b got quieter. The likely cause is the warning, which seems to push the model toward answering "ok".

As the project's note on this problem had planned, I changed one sentence of the review prompt (the instructions both reviewers receive). It still says how many verdicts are needed, but the warning is gone; in its place it asks the reviewer to check every field as closely as if it were listing only problems, and says that "ok" means it found nothing wrong. The tests pass. The next lint will measure the rate again and check that complete replies continue.

The rest of the reading. Every wiki page is listed in the index. The log has an entry for every cycle. The open-questions page holds no unanswered questions. The file of questions for you has four open lines, none duplicated: the American pronunciation of several, the British pronunciation of medicine, the stress of the noun outside (kept to match your ruling on inside), and the old branch to delete. The baton file (NEXT.md) now mentions outside as well.

Spend

None this cycle. The day's total is US$7.42 of the US$15 cap.

What did not work

Nothing failed. Three of the ten back-links were on the wrong sense, a higher share than usual; all three came from one entry with closely related senses.

What is next

A build or review cycle if the run clock allows; the next lint is due in about five cycles.

What needs the owner

Nothing new. The prompt change is small and reversible; if you would rather keep the stronger warning, say so in the inbox.

All journal entries