ProofGraph

Courts have now dealt with more than 2,200 cases of AI-hallucinated content, and the industry’s answer is citation checking: does the case exist, is the quote real. That catches fake sources. It misses the error that reaches a sign-off every day inside companies: a real clause that doesn’t say what the sentence claims. “Under our NDA, Kestrel can’t share our pricing with anyone.” The NDA does say “shall not disclose”. The same sentence then says “except to its Representatives”. Every citation checks out. The claim is wrong. ProofGraph is the check before you say yes. Paste the claim, drop in the documents. Each part of the claim gets an answer and the clause that decides it, boxed on the page. When the documents are silent, it says so instead of guessing. It’s for in-house legal and contract reviewers who get “can you confirm?” from Sales every week and re-read the whole packet to be safe, because one wrong yes costs more than the afternoon. With ProofGraph they answer in minutes and send back the clause itself. How: a span-grounded evidence graph, not RAG. No embeddings, no vector store. Muse Spark reads every page, scans included, into rule-shaped statements with exact quotes, links the claim to the packet, and labels each claim-and-rule pair. Code pins every quote to its page, walks a typed graph, applies which document wins, and decides. Muse Spark Open, open-weight on Baseten, re-reads the deciding links without seeing the verdict. On 30 held-out contract pairs, it never passed a real contradiction as supported: 0 of 10. I spent two years building document AI for billing reviewers. The AI rarely found nothing. It found a real line and drew the wrong conclusion. ProofGraph is the tool I wanted then.