conorbronsdon/avoid-ai-writing v3.24.0

on GitHub

Added

• Unnecessary hyphenation is now a P2 copyedit with deterministic detector
  coverage (#107). The rule handles three bounded subclasses: welded open
  noun phrases (research-impact aggregator → research impact aggregator),
  compounds with an established closed form (code-base → codebase), and
  attributive compounds used adverbially (in real-time → in real time,
  while real-time analytics stays unchanged). The catalog goes from 61 to 62
  categories and the engine from 47 to 48 types.
• Protected spans and legitimate compounds stay out of the detector. It
  masks fenced and inline code, quotes, Markdown blockquotes, URLs, paths,
  filenames, command flags, identifiers, version strings, YAML metadata,
  Markdown tables, and HTML attributes. Fixtures preserve high-quality,
  family-owned, third-party, real-time dashboard, long-term plan, and
  out-of-the-box support.
• The general grammar call remains editorial judgment. Open-ended compound
  detection would flag ordinary technical writing, so the engine uses a curated
  list and reports suggestions instead of rewriting text. The optional -ly
  adverb cleanup discussed in #107 is deliberately absent from this release;
  the issue agreement allowed it to ship later as opt-out style cleanup rather
  than as an AI tell.

Changed

• Hyphenated-pair overuse is now named hyphenated modifier stacking. Its
  signal is the density of otherwise valid compounds, not the correctness of
  each hyphen. Incorrect but deterministic forms belong to the new rule.
• P2 hyphenation copyedits do not contribute to the AI score. They remain
  visible as editing suggestions without changing the label, class
  probabilities, or trinary classification in short documents.

Fixed

• Path and filename masking remains linear on adversarial input. Bounded
  path components remove the superlinear backtracking exposed by long kebab
  identifiers, with a timing regression fixture covering the failure shape.