ai·rete·rag

ai·rete·rag

ai·rete·rag splits deciding from explaining. The verdict comes from a Rete production-rule engine — deterministic, reproducible, and traceable to the exact rule that fired, with salience deciding which rule wins when two disagree. Only then does an LLM run, grounded by retrieval over your own policy documents, to render that verdict into prose that cites real policy text. The model never decides anything, so an explanation can't invent a threshold your rulebook doesn't have. Eight tools cover the whole loop: hand it a written policy and get back draft rules that cite the sentence each one encodes, review them, save them as YAML, ingest supporting documents, then decide against them — and ask why any rule didn't fire. Eight demo domains (loan, fraud, clinical, insurance, legal, blockchain, operations, e-commerce) work without a key.