How slop-prose works

The density model behind detecting AI-written prose — machine artifacts, style tells, the sliding window, and why it's a nudge rather than proof.

slop-prose is a separate analyzer for the linguistic tells of LLM-written text, scoped to prose (.md/.markdown/.mdx/.html) — never code. It's on by default (Straitjacket runs at its max); disable it with --skip slop-prose.

Unlike an emoji in source, "reads like an LLM" is a probabilistic claim, so its density gate warns before it fails — the design reflects that.

Two mechanisms

Machine artifacts → hard fail

Copy-paste residue that a human essentially never types:

oaicite / contentReference
turn0search0
utm_source=chatgpt.com
As an AI language model
unfilled placeholders (PASTE_URL_HERE, [Your Name])
2025-XX-XX dates

A single hit is an error, regardless of document length. These are language-agnostic — they work no matter what language the surrounding text is in.

Style density → warn, then fail

Everything else carries a weight:

AI-vocabulary words like delve / tapestry / pivotal
stock phrases like "stands as a testament"
negative parallelisms like "not just X, it's Y"
spaced em dashes, curly quotes

No single one means much — the signal is co-occurrence. Straitjacket slides a fixed window (the --prose-window, default 400 characters) across the text and scores the densest span. Elevated density → warning; high density → error.

Why divide by a fixed window

Dividing by a fixed window (rather than by total length) is deliberate: it keeps short text lenient, so a lone "not X, but Y" in a commit-style line can't spike the ratio. The score and the contributing phrases are shown:

CHANGELOG.md:14:4  [slop-prose]  density 0.100 (score 40/400)
   AI-prose density 0.100 over a 400-char window — reads like LLM slop:
   "stands as a testament", "rich cultural heritage", "showcasing", "vibrant", …

English only, for now

The wordlist, stock phrases, and templates are all English — slop-prose doesn't know what LLM slop sounds like in any other language. If you want a specific language and are willing to help verify what actually reads as sloppy in it, file an issue requesting it. (The machine-artifact hard-fails above are language-agnostic and work regardless.)

Caveats, by design

LLMs are trained on human text, so these distributions overlap — humans write "not X, but Y" and use em dashes too. Treat slop-prose as a nudge for review, not proof; the thresholds are v1 calibration guesses (FAIL_DENSITY / WARN_DENSITY in src/slop_prose.rs). Grounding, methodology, and the full tell taxonomy — including what isn't statically detectable — are in notes/detectability-tiers.md, derived from Wikipedia's Signs of AI writing.

Exempt a doc that legitimately quotes these things (like those notes) with straitjacket-allow-file:slop-prose, or a single line with straitjacket-allow. See Suppression markers.