Roadmap
What the engine detects today, and the layers planned next — stated plainly so nobody has to guess.
The engine ships in layers. What exists today, and what's planned:
Shipped
- L0 alignment — seed-and-extend character-range alignment.
- L1 winnowing fingerprints — GIN-indexed retrieval over
bigint[]fingerprints in Postgres.
Together these detect verbatim and near-verbatim reuse against a local corpus, with per-passage character offsets and honest coverage accounting.
Planned
- Paraphrase layer — candidate retrieval robust to word-level rewriting.
- Semantic layer — embedding-based matching for meaning-level similarity.
- AI-writing detection — flagged as a separate signal, never conflated with plagiarism.
Planned layers land when they meet the same bar as the shipped ones: measurable on public benchmarks, with numbers we publish. Follow the public roadmap for status.
What this means for self-hosters
A self-hosted engine today gives you exactly what the benchmarks measure: strong verbatim and near-verbatim detection. If your threat model is heavy paraphrase or AI-generated text, the current engine will not catch it — plan accordingly rather than assuming.