Show HN: AI·rete·RAG – a Rete rule engine decides, RAG explains why

The author presents ai·rete·rag as a two-stage approach to auditable automation: a pure‑Python Rete rule engine produces a deterministic verdict from YAML rules (with salience‑based conflict resolution), then a retrieval‑augmented generation step fetches policy passages and an LLM drafts a plain‑English explanation that cites sources but cannot change the verdict. They highlight features they say surprised them - rules expressed as graphs with nested all/any/not and forward chaining, an audit mode that records every evaluated rule and snapshot for replay, fired rules steering retrieval, retrieved text becoming facts, a visual editor for non‑technical authors, and LLM‑drafted rule suggestions that require review. The submitter notes a live demo across multiple domains, an MCP tool integration, an open‑source MCP client, and that the engine/platform are currently hosted and closed‑source.

The post frames a clear position in favor of separating decision logic (deterministic rules) from explanation (LLM RAG), and it solicits input from people who have explained automated decisions to regulators. Tensions the author raises - and where opinion is likely to divide - include tradeoffs between auditable determinism and the flexibility of learned models, trust in LLM‑generated explanations versus potential for hallucination or obfuscation, and the implications of a hosted closed‑source platform versus open‑source verifiability. The author explicitly invites practical feedback about what auditors and regulators actually require.