← Projects·AI Engineer
slop-scanner
A deployable Streamlit app that scores a draft on 13 literature-grounded quality metrics (vocabulary entropy, hedge-word density, generic-template phrases, repetition n-grams, etc.) and outputs a single SLOP index from HEAVY to CLEAN. The stdlib scoring core needs no LLM and no API key.
Why this exists
The “AI wrote this” smell is detectable in measurable ways — repetition, hedge-word density, generic phrases, low vocabulary entropy, template openings. slop-scanner operationalizes that instinct into 13 numeric metrics you can compute deterministically, no LLM needed.
The 13 metrics
A combination of:
- N-gram repetition rates (overlap with a generic-phrase corpus).
- Hedge-word density (“could potentially maybe”).
- Vocabulary entropy (Shannon) over content words.
- Sentence-initial template phrases (“In today’s fast-paced world…”).
- Bullet-list overuse.
- Specificity score (proper nouns, numbers, cited sources vs vague placeholders).
- And 7 more, each with a paper citation in the README.
What’s transferable
The scoring core is deliberately tiny and dependency-free — it drops directly into any CI step, including the reflect-revise loop and the eval-mcp-server exposed MCP tool. A real draft went from HEAVY (index 81) to CLEAN (index 3) over one revision cycle.