| name | ai-residue-detector |
| description | Detect AI-writing residue and author-fingerprint loss in any draft (academic, email, or Chinese-X). Trigger when the user wants to check a draft for AI flavor, robotic tone, missing voice, over-uniform syntax, AI buzzwords, or low judgment density — or asks "does this read as AI" / "where did my voice go." |
ai-residue-detector
LAYER 3 · SKILL · the lightest, highest-frequency tool. Usable on academic drafts, emails, and Chinese-X alike. Detects two sides of one coin: AI flavor and author-fingerprint loss.
Detection signals
- Syntactic over-uniformity — sentence-length standard deviation too small; no rhythmic variation.
- AI buzzword density — leverage / underscore / robust / comprehensive / anchor / converge / salient / multifaceted / delve / pivotal / nuanced / tapestry … (extend the list).
- Low judgment density — how many "so what" sentences per paragraph? Near zero = AI flavor.
- Absent authorial presence — zero judgment verbs, zero meta-judgment sentences ("this much is uncontroversial", "what is striking is…").
- Over-completeness — every claim fully cited, every controversy given both sides, zero "we don't know yet." Human high-quality writing is sometimes deliberately incomplete.
Output
- an AI-flavor score (with sub-scores per signal)
- specific locations (which sentences trigger which signal)
- for each location: "what author fingerprint should have appeared here" (cross-reference the voice spec)
Hard constraints
- RL-1: report on tone/voice only; never touch the substantive argument.
- RL-4: a flagged "imperfection" is not automatically bad. Distinguish AI-flavor imperfection (✗) from strategic/resonance-carrying imperfection (✓, see voice-synthesizer §6).
Self-assessment for scripting
This skill is the STRONGEST candidate for Python — most signals are computable: sentence-length stdev, buzzword frequency, citation-per-sentence ratio, type-token ratio. After a representative run, if the corpus is non-trivial, build a deterministic scorer per SCRIPTING-POLICY.md. Keep the "what fingerprint should be here" judgment LLM-driven; let the script feed it locations.
