slop-detector

Pattern detector for inbound text: emails, writing samples, proposals, deck copy. It prints an evidence report where every finding cites a byte span and a verbatim snippet, and the reader decides what the findings mean.

What it detects

The report splits findings into three categories, plus input measurements.

• paste_residue: residue left by copying out of a generation surface. Provider-attribution lines, chat-export citation artifacts, chat-tool tracking parameters, turn markers, citation-delimiter codepoints, invisible unicode, and positional typographic spaces.
• injection_patterns: phrasing that addresses an assistant, such as demands to disregard prior guidance or reveal hidden configuration.
• quality_patterns: formulaic-writing patterns, from the measured excess-vocabulary set through business-register density, structural scaffolding, agent-loop vocabulary, and verbatim self-duplication.
• stats: word_count, byte_len, and the per-class densities block with raw and residual rates per 1000 words.

Read each paste_residue and injection_patterns finding individually. Read quality_patterns as densities against the word count. Every finding carries a container label (prose, fenced-code, blockquote, quoted, heading) from a linear heuristic pre-pass, so the reader can discount fenced or quoted material. The label annotates and never suppresses: the tool scans raw bytes with no markdown segmentation.

Every rule is data, loaded from data/inbound/inbound.toml. The reference lexicons in data/words/ and data/policy.toml are vendored from the ai-slop crate. The inbound selection and its edits live in data/inbound/.

CLI

The slop-detector binary reads file path arguments, or stdin when no path is given, and prints the report as JSON on stdout.

slop-detector inbound.txt
cat inbound.txt | slop-detector
slop-detector deck-v1.txt deck-v2.txt
slop-detector --allow-term beekeeping paper.txt

One input produces the single-document report. Two or more paths produce the bundle report: a full per-file report for each input plus cross_file_duplication, the verbatim runs of ten or more words shared between files, each occurrence cited by path and span. The repeatable --allow-term flag labels findings matching a stated topic word with topic_term: true. Labeled hits stay in the report and leave the residual density figures.

Exit 0 means a report was produced. Exit 1 is a read or encoding error. Each input over 4 MiB is rejected with exit 40. An input with zero matches yields empty arrays. A closed output pipe ends the run quietly.

Library

let report = slop_detector::analyze(&text);
for finding in &report.paste_residue {
    println!("{} at {:?}: {}", finding.rule_id, finding.span, finding.snippet);
}
let bundle = slop_detector::analyze_bundle(&docs); // (path, text) pairs

analyze is a pure, total function of the input text, and each per-file report inside analyze_bundle equals the analyze output for that text.

The coupled skill

The crate pairs with an agent skill, in the repository at skills/slop-detector/, that directs a human-driven agent through the read: state the purpose, run the tool, read residue and injection findings per hit, compute quality densities per class, and report with cited spans, leaving the conclusion to the human. Per-rule interpretation notes live in the skill's references/rules.md.

License

MIT OR Apache-2.0.