Catch what your model made up.

📊 0.948 accuracy on HaluEval/QA (standard benchmark, at its tuned threshold) — above ChatGPT's published 0.63 · 0.66 on summarization (the shipped JavaScript build, measured on held-out pairs) — deterministic, offline, zero-cost.

Paste the source (what is actually true) and your LLM's output. Every sentence that hallucinates content, invents a number, or flips a claim’s negation or direction is flagged — with the reason.

🔒 Runs entirely in your browser. Your prompts and outputs never leave your machine — nothing is sent to any server.

This is the deterministic auditor from the NEURON7X Cognitive Kernel, running client-side: it grounds every claim against your source, rejects fabricated quantities (digit- and word-level, unicode-safe), and detects meaning inversions (negation / direction flips). It is red-teamed (8 attacks blocked across 4 breach classes) and benchmarked on HaluEval, a standard external set: 0.948 accuracy on QA — above the published ChatGPT baseline (0.63) — and 0.66 on abstractive summarization for this JavaScript build, deterministic and offline. An audited claim is GROUNDED, never asserted true: truth is a separate step. Boundary: relevance is lexical (bag-of-words), so arbitrary paraphrases, rare antonyms, and argument-order reversals (“A causes B” ↔ “B causes A”, identical tokens) can pass — the full engagement adds a labelled eval and CI gate.

The one coordinate nobody else holds: deterministic · offline · $0 · sub-millisecond · red-teamed — and still beats an LLM judge on HaluEval QA. The fail-closed first pass for CI, honest that it is a lexical filter, not a truth oracle.