unslop-check

Calibrated AI-text detection reference CLI — distilling unslop.run's methodology into your terminal.

"Unslop this paper before submission."

What is this?

Unslop is a free, calibrated AI-text detector for scientific writing. It uses a three-expert mixture (neural + lexical + stylometric) with a false-positive floor at 0.4% — meaning 99.6% of genuine pre-2020 arXiv writing clears it. A study of 12,750 arXiv papers found ~32% of new submissions read as machine-written, ranging from 65% in CS to 0.7% in Mathematics.

unslop-check brings that knowledge to your command line — methodology reference, field data, stylometric analysis, and cross-reference with Hallmark's 57 gates.

Install

npm install -g unslop-check

Or run directly:

npx unslop-check help

Usage

Reference commands

# Show the calibration curve (FPR vs. AI recall)
unslop-check calibrate

# Show per-generator detection rates
unslop-check generators

# Explain the three-expert mixture approach
unslop-check methodology

# Show training data composition
unslop-check data

# List all arXiv fields
unslop-check fields

# Get field-specific data
unslop-check field cs
unslop-check field mathematics
unslop-check field statistics

# Study limitations
unslop-check limitations

Text analysis

# Analyze a text file for machine-writing signals
unslop-check check paper.txt

The check command computes 7 stylometric/lexical signals inspired by Unslop's three-expert approach:

| Signal | What it measures | |---|---| | Burstiness | Variation in sentence length — AI text is more uniform | | Sentence Variation | Standard deviation of sentence length | | Type-Token Ratio | Lexical diversity — AI text can be more repetitive | | Function Word Density | Filler word frequency — higher in AI academic text | | Hedge Markers | Words like significantly, notably, demonstrate — common in AI science writing | | Transition Density | Words like however, therefore, furthermore — structured AI text overuses them | | Paragraph Uniformity | How uniform paragraph lengths are — AI text is very uniform |

Each signal scored 0–1 (0 = human-like, 1 = machine-like), with an overall composite score.

Cross-reference with Hallmark

unslop-check crossref

Shows how Unslop's calibrated detection methodology maps to Hallmark's 57 slop-detection gates and 21 macrostructures — and how to use both tools together.

Complementary tools

| Tool | What it does | |---|---| | hallmark-check | Apply Hallmark's 57 qualitative gates to text | | unslop-check (this) | Calibrated quantitative detection + stylometric analysis |

Recommended workflow:

1. Run hallmark-check → get gate flags and macrostructures
2. Run unslop-check check → get stylometric/lexical scores
3. Cross-reference: high gate count + high unslop score → strong signal
4. Field-context: a flag in Math is more significant than a flag in CS

Data sources

All methodology data from unslop.run:

• How our AI-text detector works
• How we measured AI writing across arXiv

License

MIT — built by shift-zero from unslop.run's openly published methodology.