Add AI-iness density pre-check to gate pass strength (closes #93)#115
Closed
adelaidasofia wants to merge 1 commit into
Closed
Add AI-iness density pre-check to gate pass strength (closes #93)#115adelaidasofia wants to merge 1 commit into
adelaidasofia wants to merge 1 commit into
Conversation
Counts Tier 1 dead-giveaway tells per 100 words and picks light/mixed/full mode before applying patterns. Stops the skill from over-correcting human-first text (personal journals, rough drafts, meeting notes) where all-or-nothing rule application was stripping out the quirks that made the writing sound like a person. Tier 1 = the seven dead-giveaway rules already in SKILL.md (1, 4, 7, 20, 21, 22, 25). Density score is Tier 1 hits per 100 words: - 0 tells: light mode, skip the rewrite (apply only specific Tier 1 hits) - 1-2: mixed mode (default), Tier 1 in full + others where they clearly fire - 3+: full mode, all 29 rules Strictly additive to SKILL.md. No new tools, no infra changes. Closes blader#93
duathron
added a commit
to duathron/humanizer-ext
that referenced
this pull request
May 22, 2026
- Add DETECTION GUIDANCE section (false positives, human-writing signs, LLM idiolects) so editors know what NOT to flag (PR blader#113) - Add Tier-1 AI-iness density pre-flight in Full mode; auto-drops to Quick when density = 0 to protect human-first drafts (PR blader#115 adapted) - Expand blader#7 with era-specific vocabulary clusters (GPT-4 / GPT-4o / GPT-5 eras) and figurative-vs-literal caveat (PR blader#111) - Expand blader#9 with "rather than" dismissals + on-the-table test (PR blader#85) - Expand blader#14 with paired em dash bracketing + 4 fix options (PR blader#85) - Expand blader#21 with speculative gap-filling ("maintains a low profile" template detection) (PR blader#111) - Expand blader#23 with three more didactic disclaimers (subsumes pattern 34 from PR blader#112) - Expand blader#25 with structural "## Conclusion" section note - Add pattern blader#35 Debunking-Pose Headings -- heading-level AI tells that slip through prose-only passes (PR blader#116) - Add patterns blader#36 Conditional Frame Stacking and blader#37 Miscalibrated Epistemic Confidence (PR blader#85) - Add patterns blader#38 Reference-Markup Artifacts, blader#39 Phrasal Templates / Placeholder Text, and blader#40 Markdown / Wikitext Contamination -- three chat-UI copy-paste tells that confirm AI involvement (PR blader#112) - Extend domain overrides for blader#35-37; blader#38-40 are universal - Extend final AI audit from 9 to 13 points - README: pattern count 34 -> 40, three new section rows, updated fork-differentiator table, 3.2.0 version-history entry Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
Hi @adelaidasofia , your PR got me on this. Built a sibling version that moves counting from the model into a Python script (attached). Different tradeoff than your "no new tools" scope, not a replacement. Two things that fell out of going deterministic: -compare input output for the audit pass. Word-count delta (catches silent truncation, #78), tells removed, and new tells the rewrite introduced as a regression check. Stdlib-only, handles Hebrew. Cross-pinging @duathron since humanizer-ext is where adaptation is happening. |
Owner
|
Closing — the principle (don't over-edit human-first text) is good, but a 41-line tier-scoring procedure is more ceremony than the prompt should carry, and the v2.6.0 PERSONALITY gating + DETECTION GUIDANCE already steer away from gutting human prose. Thanks for the work. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Per #93. Adds a pre-flight density check that counts Tier 1 dead-giveaway tells per 100 words and picks pass strength (light / mixed / full) before applying patterns.
This stops the skill from over-correcting human-first text (personal journals, rough drafts, meeting notes), where all-or-nothing rule application strips the exact quirks that made the writing sound like a person.
What changed
SKILL.mdonly. Strictly additive, no new tools, no infrastructure.## Pre-flight: AI-iness Density Checkinserted before## Process. Defines Tier 1 as the seven dead-giveaway rules already in SKILL.md (1, 4, 7, 20, 21, 22, 25), the density thresholds, and the pre-flight announcement format the user sees.## Processupdated: step 2 runs the pre-flight check; the rest of the steps renumber accordingly.41 added, 8 modified (renumbering only).
Density thresholds
Why these seven rules
Rules 1, 4, 7, 20, 21, 22, 25 are the patterns that almost never appear in authentic human writing (broader-trends framing, promotional language, AI vocabulary, "Certainly!" / "Of course!" openers, knowledge-cutoff disclaimers, sycophantic tone, generic positive conclusions). Restricting the density score to these keeps the signal clean. Lower-tier rules can fire as false positives on human writing, so counting them would inflate the score and trigger over-correction.
Verified
Running in production on a fork (adelaidasofia/humanizer) for several months on personal journal and rough-draft content. Eliminates the "humanizer made my writing worse" failure mode on human-first text. AI-generated text still routes to full mode and gets the comprehensive rewrite.
Notes
This is intentionally scoped to the issue. Other improvements in the fork (Spanish-language rule library, voice fingerprinting, ROI-ranked tiers for all 29 rules) are separate concerns and can ship as later PRs if you want them.