How this dashboard works
Drupal core has posting rules for AI-assisted contributions: core/AGENTS.md. This dashboard measures how the issue queue follows them. It counts; it does not moderate or report anyone.
What gets checked
Every issue comment, issue summary, and merge-request note with text goes through two checks.
Disclosure: does the text say AI was used? Three forms count. The exact AGENTS.md line. A named tool ("written by Claude"). A plain admission ("I used AI"). Talking about AI does not count.
Writing rules: five checks from AGENTS.md §2.
- over 180 words
- a sentence over 25 words
- stock words
##headings in short comments- template phrasings: "not only … but also", "rather than", "I believe", "whether X or Y"
Code blocks are removed before checking. Word matching catches conjugations: leverage also matches leveraged.
What "Likely-AI" means
A post counts as Likely-AI only when its author disclosed AI use. Writing style alone never puts a post in that count. Style-based guessing was tested and rejected: it missed real cases and flagged humans.
What is public, what is not
Public: aggregate counts, and the list of self-disclosed posts (without author names). Numbers covering fewer than 5 people are hidden from the published data.
Never public: style suspicion about a named person, per-author history, and author names on per-post data.
Per-post API
One post at a time can be checked at api.tresbien.tech/v1/meta/issue/<nid>, /v1/meta/comment/<cid>, or /v1/meta/note/<gitlab note id>. The response shows disclosure, a rating for non-disclosed posts, and the reasons behind it. The API is limited to 10 requests per minute. There is no way to list posts through it.
The API also checks three rules added to AGENTS.md after the experiment started: negative parallelism, em dash pivots, and reader narration. Dashboard numbers stay on the original five rules so the before/after comparison holds. Posts from before 2022-11-30 (ChatGPT release) always rate "no".
Timeline
AGENTS.md was published on 2026-06-27. The experiment clock (T0) starts the following Monday, 2026-06-29. Charts compare three windows: before publication (the baseline), the first week after (settling in, shown grey), and everything since. The ruleset is frozen at its T0 version so the comparison stays fair.
Data
- Posts checked, latest day
- 34
- daily.json — every number on the dashboard, raw.
- stock-words.txt — the word list, verbatim.
- AGENTS.md — the ruleset snapshot from T0.