docs(humanizer): reconcile with blader/humanizer upstream, v0.3 -> v0.4

[yulonglin]

Reviewed upstream blader/humanizer (now 33 patterns, up from 24 at first analysis) and folded in the genuinely new ones.

25 -> 31 patterns, 7 -> 8 categories.

New patterns

Pattern	Confidence	Category
Knowledge-cutoff disclaimers	97%	Chatbot Artifacts
Promotional language (puffery)	84%	AI Vocabulary
Superficial -ing significance clauses	86%	Structural Tropes
Decorative emoji in headings	88%	Formatting Tells (new)
Excessive boldface	78%	Formatting Tells (new)
Title Case headings	72%	Formatting Tells (new)

Category 8 (Formatting Tells) covers presentation-layer habits rather than word choice. These survive paraphrasing, so they still fire on lightly-edited LLM prose.

Key decision: patterns yes, philosophy no

blader/humanizer is a rewriter that transforms text and mandates stripping every em-dash. This is a detector that flags and scores. Patterns port across that boundary; the em-dash policy does not.

Kept the frequency-based stance: 1-2 em-dashes per page is normal human writing, 3+ per page is the signal. Adopting zero-tolerance would tank precision against the >90% target.

Added a "Deliberately Rejected" table so the reasoning survives:

• Zero-tolerance em-dash elimination - conflicts with frequency-based scoring
• Diff-anchored writing - coding-agent artifact, out of scope for draft review
• Copula avoidance - deferred, not rejected; no characterisation precise enough to hit the precision bar yet

Drift fix

The three artifacts had diverged:

Artifact	Was	Now
claude/docs/humanizer-patterns.md	25 patterns / 7 cats	31 / 8 (canonical)
ai-safety-plugins agent prompt	claimed both 25 and 23	31
ai-safety-plugins patterns JSON	stale at 15 / 5	31 / 8

This doc is now the canonical source and names its two derived consumers inline, so the fix is structural rather than a one-time sync.

This PR contains the dotfiles doc only. The two derived artifacts in ai-safety-plugins are regenerated and validated locally but uncommitted - that repo was never isolated into a worktree, so it needs a look before committing.

Verification

Counts checked programmatically across all three: per-category declared vs actual, category totals, and header claims all agree at 31 / 8. JSON parses.

https://claude.ai/code/session_01HZxdsbK7sYFNJCP8VyBSDj

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 6ab3397b-15b6-47fc-bc72-57c12eba4a07

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch worktree-humanizer-v04

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}