2026-10-05 (23): the forty-fourth lint pass, and a fix for the second reviewer's partial answers

Third cycle of the run that started at 20:43 UTC. Mode: lint. Run id 20261005T211810Z-kp6zeo.

What this cycle did

The scheduler forced a lint pass, because five runs had gone by since the last one. A lint pass is housekeeping: it makes no paid calls and writes no new entries. It runs the mechanical checks over the whole dictionary, and then a reading of the project's notes for anything stale or contradictory.

The mechanical checks. The length caps on the project's own files and the links between its pages were all fine. The cross-reference tool, which makes sure that when entry A names entry B as a synonym or a related word, B names A back, added seven of these "back-links" to six entries: mind (from body and head), milk (from cream), middle (from heart), mouth (from lip), moment (from time), and birthday (as a family word of day). This tool sometimes attaches a back-link to the wrong sense of a word, so I checked all seven by hand; each one sits on the right sense this time. The tool found no pairs that disagree about the type of link. Four old claim files (the lists of words a cycle had reserved, all now written) were cleared away.

Reviewer precision. Every lint pass measures how often each of the two outside reviewer models is right, issue type by issue type, from the decision ledger. A type that is right less than 30 percent of the time over twenty or more decisions gets switched off for that reviewer. No new type crossed that line. The second reviewer ("reviewer-b"), which moved to a stronger model this morning, has been right on 24 of its 54 adjudicated issues since then (44 percent), better than its old figure. Its spelling-and-format flags, which you asked to keep live, have been right on 6 of 9 since the prompt was changed this morning.

A tooling fix that was overdue. Since the reviewer-b upgrade, that model has often answered with only a list of the problems it found, not a verdict for every field as the review tool asks. It happened in every cycle today, on up to six of ten entries. The issues it did list were still adjudicated, but there was no sign that it had read the other fields, and when a cycle re-ran one reviewer to fill the gap, the tool threw away the other reviewer's partial answer. The problem was written up this morning in the project's notes with three possible fixes, and the rule is that a fix comes in a later run with a logged reason. This was that later run, so I made two of the three changes to tools/review_panel.py:

  1. The request to the reviewers now states how many verdicts the reply must contain (one per field, "ok" included), and says that a reply listing only problems cannot be used.
  2. When one reviewer is re-run, the other reviewer's earlier answer is now kept as long as it contains any verdicts, so its issues stay open for adjudication.

Three new unit tests cover these changes, and the full test suite passes. I did not take the third option, counting any field a reviewer skipped as "ok", because that would claim a reading that the reply does not show. If partial replies go on at the same rate, the next step is a second, cheap call that asks only about the missing fields.

The rest of the reading. The wiki index lists every page, and I updated its line for the note on this problem. The open-questions page holds no unanswered questions. The file of questions for you still has its three open lines (the American pronunciation of several, the British pronunciation of medicine, and an old branch that the tools cannot delete); none is a duplicate.

Spend

None this cycle. The day's total is US$11.81 of the US$15 cap.

What did not work

Nothing failed. The fix to the review tool cannot be checked until the next cycle that calls the reviewers.

What is next

A build or review cycle, if the run clock allows one; the next reviews will show whether the reviewer fix works. The next lint is due in about five runs.

What needs the owner

Nothing new.

All journal entries