humanizelisted

# Humanize (three-step pipeline) Orchestrates the three prose skills that together move an AI-drafted article away from a 100% AI verdict on Pangram. The verified effect (2026-07-29, working gate) is 100% AI -> Mixed: 23.8% AI / 76.2% AI-assisted, mean window 0.993 -> 0.576. Each step exists because the other two cannot compensate for its absence. ## Why three steps | step | what it does | what happens without it | |---|---|---| | structural step (match-outline or tighten-style) | changes enough sentence structure that the downstream paragraph rewriter can clear the mechanical gate — match-outline rewrites at section level against a blueprint, tighten-style tightens paragraph by paragraph toward the author's density floor | match-voice rewrites land at distance 0.0 from the original; the gate rejects every paragraph and nothing changes (unverified — measured against the broken gate, see calibration note) | | filter-tells semantic cleanup | collapse antithesis pairs, remove CoT leakage, cut recap ballast, fix banned words | Pangram score stays at 100% AI even after match-voice, because the rhetorical patterns survive diction changes | | match-voice --no-anchors | paragraph-level diction rewrite via gpt-oss with no voice anchors | prose retains the original model's lexical fingerprint; Pangram detects it | The compound effect: semantic cleanup alone does not move Pangram (rhetorical patterns are not what it measures). Both steps together, with the structural step providing