/ revision utility · runs locally
A diagnostic for academic drafts. It flags the paragraph-level patterns that read as machine-generated — and then asks you the question only you can answer. It does not rewrite a single sentence, by design.
100% local. Your text never leaves this page — no upload, no server, no LLM. Grammar tagging runs in an inlined offline parser.Nine patterns (P1–P9), weighted by how strongly they carried the change in the source study. Highest-yield: abstract openers, uniform paragraph shape, abstraction where a number belongs.
How it works & what a result means
This tool comes from one real revision — a reinforcement-learning paper whose AI-detector score fell from 94% to below 25%. Two findings from that history shape how it behaves.
An AI rewriting AI-flagged text plateaus. The score fell from 94% to 60% while a model did the rewriting, then stalled — it dropped below 25% only once the authors rewrote the flagged paragraphs themselves. So this tool flags a pattern and asks a question; it never writes a replacement sentence. That is also why it needs no LLM and runs entirely in your browser.
Surface metrics barely moved. Sentence length, "burstiness", and connective density stayed flat while the score halved. What carried the change was paragraph-level reconstruction — different information order, claims fused or split, concrete specifics swapped in for abstractions. The nine patterns target those tells, not the easy-to-game surface knobs.
Reading a result — please note
"No patterns flagged" is common, and usually a good sign. The tool only looks for a specific set of AI-voice tells. Well-edited or genuinely human writing often clears all nine — that is a clean result, not a failure. For the paragraph-level checks (P3, P4) to run, paste a full section and separate paragraphs with a blank line.
Detectors are noisy — a flag is a prompt to look, not a verdict. Non-native-English and heavily edited academic prose are disproportionately flagged; if a paragraph is genuinely yours and still flags, the right response is documentation (drafts, version history), not more rewriting. Methods sections legitimately read procedural — don't "de-AI" precision out of a methodology.
Keep your AI-use declaration. Disclosed assistance plus genuinely authored prose is defensible; a low score obtained by concealment is not. This tool does not detect, score against, or evade any specific vendor's classifier.