We currently expose two paraphrasers in this playground: humarin (fast, ~5s, local T5-base) and DIPPER-XXL (slower, ~60s, 11B adversarial paraphraser). Research-only methods — Claude-rewrite and translation pipelines — are documented with full benchmark numbers at /research/methods.
Humanise
Editorial humaniser for drafts that need a human voice without losing their citations. Paste the draft, pick a mode, read the diff. The MVP ships Editorial mode only; Adversarial and Localise modes follow in v0.4.
Humanise — editorial rewrite
Rewrites your draft while preserving meaning. Re-scores after to show the drop.
Modes
- Editorial — available now. Rewrites for sentence-length variance, hedge density, and concrete-detail recovery. Preserves citations, quoted passages, code blocks, and named entities verbatim.
- Adversarial — closed-loop rewrite against the detector, shipping v0.4.
- Localise — register-aware rewrite for a target EU language, shipping v0.5.
AST-aware editorial mode
Most humanisers feed your draft to a paraphraser and return whatever comes
back. That breaks two things: citations and quoted material. A paraphraser does
not know that
(Liang et al., 2023) is load-bearing prose; it sees a noun phrase
and rewords it.
humaniser.eu parses the draft into an abstract syntax tree before the rewrite. Citation nodes, block quotes, fenced code, named-entity spans, and footnote markers are pinned. The rewriter only edits the prose nodes between them. The rewrite arrives with a diff view that shows what changed and what was held.
Citation preservation
The citation lock recognises author-year (Smith, 2024), numeric
([14]), and footnote formats. The pinned span is reinserted
verbatim at the matched position in the rewrite. If the rewriter cannot place
a pinned citation cleanly, the rewrite is rejected and the original sentence
is returned untouched.
What is not rewritten
- Direct quotations inside curly or straight quotes.
- Block quotes, including indented and fenced.
- Code fences and inline code.
- Citations and references in any recognised format.
- Named entities — people, organisations, products, court cases.
The detector page documents the verdict format; this page documents the rewrite. They share a model SHA, which appears in both receipts so an auditor can confirm the rewrite was scored by the same detector version that flagged the draft.