supported ≥0.58
weak 0.38–0.58
unsupported
contested
Privacy: embeddings local (Transformers.js). Heuristic fallback needs no download.
Heuristic ready
Reference (ground truth)
Claims to verify
—
How it works
- Segments both docs into sentences `core.js:4` (Arabic ؟ ، ؛ aware).
- Each sentence embedded via `Xenova/multilingual-MiniLM-L12-v2` quantized `app.js:56`. Cosine matrix `core.js:26` finds best reference per claim.
- Thresholds: supported ≥0.58, weak 0.38–0.58, unsupported <0.38. Negation flip (لا/not) with high sim → contested `core.js:60`.
- If model unavailable, lexical Jaccard+3gram fallback `core.js:14`.
- Private: WebAssembly in browser, PWA offline after model cached. No server.
Use for: fact-checking LLM answers against source, policy vs summary audit, citation coverage.