participles

An advisory by default: it explains, and blocks nothing, on a text that hangs present participial clauses on its sentences at a generated-writing rate.

What it catches

A text that hangs present participial clauses on its sentences at a generated-writing rate.

Dimension

economy: words that carry nothing. A finding is counted on the dimension of its rule, the cost it names to the reader.

Why

Reinhart et al. (PNAS 2025) measured GPT-4o using present participial clauses at 5.3 times the human rate. The rule counts a comma before an -ing word that opens a clause, and an -ing word opening a sentence whose comma closes the clause, per 1000 words, over texts of 200 words or more. Measured 2026-10-10: 808 agent pull request bodies from the AIDev set ran 3.9 per 1000 pooled (p50 3.7), and 30 percent run over the cap of 5; this repository's 96 hand-written docs pages ran 1.5 (p90 3.1, 2 percent over) and its 61 commit bodies 0.5 (none over), both written by people and agents together. Review comments ran 0.6 by people and 2.6 by bots, pooled. The docs are a different genre from a pull request, and the rule cannot tell a participle from a gerund without a tagger, so it advises. Measured 2026-10-11 over the adjudicated labeling round 2: 81.0% precision (34 of 42) before, 85.0% (34 of 40) after. A clause naming an issue ("addressing #2782") and an open task item a template wrote are no longer counted. The rest turn on whether a tail claims a benefit or states a fact (", renaming the types"), which the rule cannot read. It stays off until a fresh labeling round measures it. A finding quotes each clause it counted, so its reader can check them, and a platform's setup tips are not the author's text and are not counted. Labeling round 5 measured 60.0% (24 of 40, lower bound 0.45); its wrong firings counted a coding agent's quoted prompt and pasted summaries, so those are blanked with their words, and nouns, so an -ing word opening a bare item of a series (", logging methods, and"), the subject of a verb (", caching is") or a bold label is no longer counted. Labeling round 6 measured 82.5% (33 of 40, lower bound 0.68). Six of its seven wrong firings counted an adjective before its noun ("existing code", "outgoing calls"), a gerund opening the main clause after an opening element ("Currently, setting X behaves") or an item of a series of -ing topics, which are no longer counted. Labeling round 7 measured 95.0% (38 of 40, lower bound 0.81 at z 2.24); a person spot-checked ten agreed-correct labels, so it advises.

Default decision

The code is PRF4020. The decision depends on the kind of text judged:

Changing it

A decisions table sets this rule to deny, advise or off, for everything judged or for the files a glob matches. Proofread reads the table from the file its -decisions flag names:

{
  "rules": {"participles": "deny"},
  "paths": {"blog/**": {"participles": "off"}}
}

rules applies to every file. A glob in paths applies to the files it matches, and the last match wins. A rule set to off never reports.

Seeing it

A finding names the rule, its code and the decision that applies, and links here:

{"rule": "participles", "code": "PRF4020", "decision": "advise", "url": "https://eli.gladman.cc/magus/reference/proofread/participles/"}

proofread rules prints this entry with every other rule's.

See also

• All proofread rules - every rule, grouped by what it checks
• Writing rules - how the rules apply to this repository's pages and pull requests
• The findings contract - the fields every finding carries