slop-meter-cli

Marks each paragraph human-ish, machine-ish, mixed or can't tell, with calibrated odds. It's the model behind slop-meter.com and the Chrome extension: 56 KB of weights that run locally, with no network calls and no API key.

It says can't tell on most paragraphs. That is on purpose: it only makes a call when it is sure, and it calls about 1 human paragraph in 1,000 machine-ish. Treat a mark as a guess about style. It can't tell you who wrote something.

CLI

npx slop-meter-cli post.md
cat draft.txt | npx slop-meter-cli

Installed with npm i -g slop-meter-cli, the command is slop-meter.

post.md
  ¶1  machine-ish  99%  In today's fast-paced digital landscape, it's important to note…
  ¶2  can't tell        I rebuilt the bench vise last week. It took two evenings and…
  2 read · 1 machine-ish · 0 human-ish · 0 mixed · 1 can't tell

| Option | Effect | |---|---| | --json | Print every paragraph as JSON | | --why | List the rules behind each call | | --fail | Exit with code 1 if any paragraph is machine-ish, for CI |

Markdown headings, code blocks and HTML comments are skipped. Paragraphs under 25 words, or not in English, are counted but not scored.

Library

import { score, scoreText } from "slop-meter-cli"

const reading = await score("In today's fast-paced digital landscape…")
reading.verdict // "machine"
reading.label // "machine-ish"
reading.confidence // 0.99
reading.reasons // [{ id: "r-006", name: "abstract metaphor nouns", spans: [[25, 34]] }]

const readings = await scoreText(longText) // one reading per paragraph

A reading has text, words, verdict (human, machine, mixed or unsure), label, confidence, probabilities, reasons and skipped (too-short, not-english or null). Each reason's spans are character ranges of the words that set it off.

It works in Node 20+ and in the browser. The models page has every number, and the rulebook explains each reason.

License

AGPL-3.0. Source: github.com/zaidmukaddam/slop-meter.