Demo · live RAG grounding check
Did the model make this up? Or is it grounded in the sources?
Paste a model output and the source documents it was supposed to be grounded in. We score how strongly the output is supported by the sources and return a hallucination probability, verdict, and signed receipt. Useful as a guardrail in RAG pipelines, chatbot factuality checks, and summarization audits.
How this works
- v7 backbone + grounding probe. The model encodes the output and each source document; the probe scores how well the output is supported by the source evidence.
- Category-aware threshold. Q&A, summarization, and dialog have different grounding norms; the probe uses the category to pick the right calibrated threshold.
- RAG-pipeline use. Drop in as a guardrail after retrieval — if the output isn't supported by what was retrieved, fail closed or surface uncertainty to the user.
- Bantu-language coverage. Same multilingual backbone, so the grounding check works on Zambian-language outputs too (under NDA today, in the API key tier).
- Signed receipt. Every check produces a SHA-256 audit record — same audit substrate as bias and hate.