Slop lint

slop-lint

Purpose

Deterministically measure how much a piece of prose relies on AI-writing tells — negative parallelism, significance inflation, uniform rhythm, formatting reflexes, era-tagged vocabulary, leaked chatbot artifacts — and report span-anchored findings plus a 0-100 cluster score. The script is pure Python stdlib: no network, no LLM, same input → same output. It fixes the observed failure of prompt-only "humanizer" packs: judgments by feel, no numbers, no CI gate, and single-tell overreactions (an em dash is not a verdict).

When to use

• "score / scan / lint / check this text for AI tells or slop"
• "how AI-sounding is this?" — anything wanting a number or a report
• Gating docs or marketing copy in CI before publishing
• Producing the findings a rewrite pass will edit against

When NOT to use

• Rewriting or de-slopping text → use deslop (it runs this lint first).
• Checking whether an edit changed facts/quotes/meaning → use deslop-verify.
• Judging whether a specific person or student used AI — this is a style lint, never an authorship detector; refuse accusation framings.
• Linting code, grammar, or spelling.

Workflow

1. Pick the genre profile: general (default), academic, technical (docs/READMEs), or casual. When unsure, say which you chose and why.

2. Run the script (span-anchored findings + score):

   python3 "${CLAUDE_SKILL_DIR}/scripts/slop_lint.py" FILE... --genre technical
   python3 "${CLAUDE_SKILL_DIR}/scripts/slop_lint.py" --json < draft.md   # machine-readable
   

3. Report to the user: the score and band, which families are active, and the top flagged lines (the report includes line numbers). Always state that the score reflects a cluster of co-occurring tells — never conclude anything from one family, and never frame results as proof of AI authorship.

4. If confidence is low (under 150 words / 8 sentences), say the text is too short for a reliable score and report values without a verdict.

5. For CI gating, use the exit-code contract:

   python3 "${CLAUDE_SKILL_DIR}/scripts/slop_lint.py" docs/*.md --genre technical --fail-above slop-cluster
   

   Exit 0 = pass, 1 = usage/input error, 2 = a confidently-scored file reached the --fail-above band. Recommend starting with --fail-above heavy-slop (warn-first rollout) before tightening to slop-cluster.

6. Sanity check after any script edit: --self-test must print all PASS.

Output spec

A report per file: score (0-100), band (clean / mild / slop-cluster / heavy-slop), confidence, active_families, and per-family value, band, and hits with line numbers. JSON mode emits the same as schema: 1 JSON. Bands: <25 clean, 25-49 mild, 50-74 slop-cluster, 75+ heavy-slop. A single active family caps the score at 24; two cap it at 49 — except leaked chatbot artifacts (oaicite, "As an AI", utm_source=chatgpt...), which force ≥50 because their false-positive rate is near zero.

Gotchas

• Em dashes are weak evidence. Human baseline is ~3.23/1k words with a huge range, and current models are vendor-suppressed below it. The dash family carries minimal weight; never present dash counts as a conclusion.
• Hedging is only scored when stacked ("may potentially"). Academic prose is legitimately hedge-dense; flagging raw hedges punishes real scholars.
• Fairness: the underlying signals overlap legitimate ESL, formal, and neurodivergent writing styles. The cluster gate and genre profiles exist to protect those writers — do not bypass them, and never report a score as evidence about the writer.
• Thresholds decay. Lexical tells die when vendors patch them (delve is gone; em dashes are fading). Thresholds and the era lexicon are dated engineering priors — see references/thresholds.md for the 6-12 month recalibration protocol before trusting them long-term.
• Code blocks are stripped before scoring; a file that is mostly code will come back low-confidence. Score the prose, not the repo.
• The heuristics are tuned for English prose (ASCII word tokenization); scores on other languages or heavily accented text are unreliable — say so rather than reporting them as fact.

Pointers

• scripts/slop_lint.py — the scorer (--help, --self-test).
• references/thresholds.md — evidence per family, threshold provenance, era-lexicon sourcing, fairness rationale, recalibration protocol.