Skip to content

v4.6.0 — Humanizer Korean Writing Module

Choose a tag to compare

View all tags
@Dannykkh Dannykkh released this 09 Jun 18:52
· 161 commits to master since this release
v4.6.0
03dedca

Humanizer Korean Writing Module

The humanizer skill could only handle English AI-writing patterns (Wikipedia's "Signs of AI writing") and missed the real signals in Korean text — translationese, nominalization, and comma habits. This release adds a dedicated Korean module grounded in Korean translation-studies scholarship, bumping humanizer to v2.5.0.

  • Korean translationese module (67 patterns, 10 categories A~J) — translation-influenced grammar, AI clichés, rhythm uniformity, over-modification, and more. Top signal: comma-after-connective (4.84× the human rate), plus ~성/~적/~화 nominalization, progressive-aspect overuse, and pronoun literalism.
  • Quantitative first-pass scan — diagnose Korean AI-ness by the numbers (comma ratio, nominalization count, progressive rate, pronoun density) before reading patterns.
  • Genre guardrails — essay / report·paper / blog / script / formal-address, each with its own allow-forbid matrix. Humanize without shifting genre.
  • Severity tiers (S1/S2/S3) across all 24 English + 67 Korean patterns, plus a procedural over-editing guard: do-not masking, risk-ordered rewriting (D→A→I→G→H→F→B→C·J→E), live change-rate tracking with safe-version rollback, and meaning-drift checks (causal reversal, quantifier weakening, modality shift, polarity flips).

Taxonomy and procedures adapted from im-not-ai (MIT) v2.0. Deployed to all three CLIs (Claude / Codex / Gemini).

Docs

  • Corrected the api-tester retention rationale in the agent/skill audit (spawn 0 + skill superset, not "passive").
Assets 2
Loading