NEXUS: Duplicate-Charge Triage on AMD

When a customer writes "I was charged twice", a support agent has to compare orders, captures, pending authorizations and earlier refunds against policy. Many are not duplicates at all (an authorization hold, two captures that add up to the order total, a separate subscription order), and a wrong answer costs money twice. NEXUS is built for that agent. The agent picks the customer and the disputed order and pastes the complaint. NEXUS stores a hashed evidence snapshot, and a deterministic policy engine computes any over-capture and the amount still owed. Anything policy says needs a person (open chargeback, pending refund, restricted account, unsupported currency, incomplete history, returns, amounts above a threshold) goes straight to a durable human-review queue with zero model calls. For every other case, Gemma-4-31B-IT, served on AMD GPUs through AMD Token Factory, reads the free-text complaint, maps it to the right charge in the selected order, flags contradictions and returns structured JSON. The model can never change eligibility or amounts: a validator checks every cited record and that any refund equals the policy figure, and code writes the customer reply from validated fields. Badges show honestly whether a result came from a live AMD call, a replay, or a policy skip. Measured on 12 frozen held-out synthetic cases: NEXUS 12/12 correct outcomes versus 10/12 for a rules-only baseline, reply checks 21/21 versus 19/21, zero critical failures for both. The baseline was fooled by an injected "refund $500" instruction and missed an off-topic login complaint; NEXUS kept the exact policy amount and sent the login case to a person. Twelve cases are descriptive, not a general accuracy claim. All data is synthetic and the refund policy is fictional. No payment, refund or email is ever executed; NEXUS recommends and drafts, and a person completes every review.