How it works

Everything here is deterministic and rule-based. There is no language model, no external API and no account. The same text with the same settings always gives the same result, and nothing you submit is stored or logged.

The pipeline

1. Mask – quotations, citations, math, URLs, code spans and your protected terms are swapped for opaque placeholders so they cannot change.
2. Lexical – cliché phrase rewrites, meaning-checked synonym swaps (with a minimum-confidence slider), determiner and contraction handling.
3. Structural – active/passive voice, clause and information reordering, nominalization, sentence fusion and splitting.
4. Discourse – connective rotation, neutral transitions, appositives for well-known technical terms, optional conceptual bridges.
5. Audit – a source-aware n-gram check: Karp–Rabin rolling hashes over 5-word windows, winnowed with a window of 4. Any run of 8 or more words that still matches the original is broken.
6. Unmask – protected spans are restored exactly. In Word files the text is edited in place, run by run, so formatting, tables and images survive.

The intensity slider

The slider is mapped on an exponential curve, so the low end is fine-grained. The default of 75% corresponds to an internal aggressiveness of 20%; every stage's firing rate scales from that.

Rules are data

All word lists, synonyms, phrase rewrites and the citation taxonomy live in JSON files, separate from the code.

Honest limits

• Rule-based rewriting is conservative. Some sentences have no safe rewrite and are left alone; the statistics panel shows what actually changed.
• Always read the result. Nothing here guarantees any outcome with a particular detector or similarity checker.
• The AI checker is a statistical estimate, not a copy of any commercial tool.
• Citation suggestions are real, well-known works chosen by keyword match. Confirm that each one supports the claim before you rely on it.