Friengineers

WarrantyParts Logistics is a warranty-claims adjudication agent that knows when not to call a model. Warranty claims look like AI problems, but about 80% of the work is deterministic policy: coverage dates, serial matches, exclusions, recall bulletins, and SLA clocks. Most agents send all of that to a model anyway, burning roughly 18,000 tokens per claim and producing citations no one can verify. We invert that ratio. A Python rules engine resolves everything computable at zero tokens. A router sends only one of six named ambiguities to Meta AI: muse-spark-1.3 for normal ambiguity, and muse-wonder-1.0 when a claim is above $1,500 or carries commercial risk. The model receives a closed evidence pack the order, 4–5 relevant policy clauses, and the supporting evidence with stable IDs. A server-side guard rejects any citation that is not in the pack, retries once, then escalates. Code, not the model, applies the policy consequence. A human approves every outbound message and parts order, and every transition is appended to a tamper-evident hash chain. On our 24-claim demo corpus, 14 claims (58%) close with zero model tokens. The full queue uses 26,138 tokens versus 441,600 for a naive RAG agent 94% fewer at $0.0226 total and $0.0009 per claim. Every citation is machine-checked, and all 24 audit chains verify. The result is a claims co-pilot that is cheap, explainable, and safe enough for a regulated workflow. It runs fully offline with a deterministic mock provider for demos; the live Meta Model API path is implemented. Next: real Shopify and email connectors, plus a golden-case regression suite.