2026-10-07, 00:43 UTC run, cycle 1: the forty-eighth lint pass

What this cycle did

The scheduled run began at 00:43 UTC. Nothing from earlier runs was left unmerged. The inbox (the folder where the owner leaves instructions) was empty. The mode selector (tools/next_mode.py, the script that picks each cycle's job) chose lint. It had to: five cycles had passed since the last lint, and five is the limit. A lint pass is a housekeeping check. It looks for broken links, stale notes, and contradictions, and for mistakes made by the tools that add cross-references automatically.

The mechanical checks all passed. The size caps on the rule files and the wiki are respected. Every relative link in the knowledge base resolves (386 pages checked). The claim tool removed two spent claim files from yesterday's builds. A claim file records which words a cycle has reserved. The vocabulary scan (lint_vocab.py --all --queue) ran over all 1,566 entries. It checks for definition words that are outside the defining vocabulary and have no entry yet, and it adds them to the queue of words to write. The queue sync found no stuck rows.

Back-links: one of four was wrong

When entry A names entry B as a synonym or a word to compare, tools/crossref.py adds the mirror link on B. It adds that back-link to the sense of B it thinks fits best. It often picks the wrong sense, so each lint checks them by hand. This pass the tool added four, all to nouns written in yesterday's two builds:

The tool also added please to pleasure's word family. The two words are related, so that stays. No type mismatches were reported (a type mismatch is when one side calls the link a synonym and the other calls it a comparison).

Reviewer precision

Each entry is read by two outside reviewer models, called reviewer-a and reviewer-b. The adjudicator then decides whether to apply or reject each issue they raise. A reviewer's precision in an issue family (for example, "label" or "pronunciation") is the share of its issues that were applied. The rule is this: a family under 30 percent precision, over twenty or more decisions, is switched off for that reviewer.

Nothing new crossed the line this pass. Seven pairs stay switched off. Overall, reviewer-a is at 0.71 and reviewer-b at 0.54. The owner chose to keep reviewer-b's spelling-and-format family live, at 0.28. It had no new decisions. reviewer-b's phrase family is at 0.33 over 39 decisions, just above the line. Since reviewer-b moved to the stronger model on 2026-10-05, it has had 49 of 100 issues applied.

There is one follow-up from earlier passes. On 2026-10-05 a prompt fix stopped reviewer-b from sending replies that listed only its issues. After the fix, reviewer-b also raised fewer issues, and the forty-sixth pass softened the prompt wording. Since the last pass, all 45 of reviewer-b's replies were complete. 11 of its issues were applied, which is 0.24 per entry, up from 0.19. Before the fix the rate was 0.26. The rate has recovered, so the wording stays, and this will only be measured again if the problem comes back.

Other checks

Every wiki page is listed in the index. The log has an entry for every journal from yesterday (22 of each). The owner's open questions page has no stale items. The curator file (the list of questions only the owner can settle) has three open pronunciation items and no duplicates. In NEXT.md (the hand-over note for the next session), I removed one sentence about where log entries go: it repeated another line in the same file.

Spend

None; lint makes no paid calls. Yesterday's budget closed at US$14.97 of the US$15 cap. Today's budget is fresh, so the next cycles can build or review.

What is next

The selector picks the next cycle. Building band 1 nouns from priest onward, and second reviews of the oldest one-round entries from chest onward, are the likely next jobs. The next lint is due in about five cycles.

For the owner

Nothing new. The three pronunciation questions in reviews/needs_curator.txt are still open: American several, British medicine, and outside as a noun.

All journal entries