Intent-Preserving Neuro-Symbolic Repair for Safe LLM-Driven Financial Transactions

Verifier-guided large language model (LLM) agents can certify that an executed payment satisfies a platform's business rules, yet the certificate says nothing about whether the payment is still the one the customer asked for. When a verifier rejects a command and returns a violation report, the LLM repair step can restore compliance by changing the recipient, the currency or the funding account, producing a rule-compliant but unintended transaction. We call this failure repair-induced intent drift and propose Intent-Preserving Repair (IPR), a neuro-symbolic execution scheme that anchors every user-facing field of the parsed command to evidence in the utterance and computes repairs with answer set programming: anchored fields become fixed facts, only defaulted and procedural fields are searchable, and lexicographic weak constraints select the least intrusive change. When no intent-preserving repair exists, a relaxation program returns a minimum set of conflicting rules that drives a targeted clarification instead of an execution. Building on the Verifier–Actor architecture of VA-NSF, we evaluate on SGD-Fin, 607 real banking and payment dialogues from the Schema-Guided Dialogue corpus paired with seeded platform states and nine business rules. On 296 instances with a grammar-constrained Qwen2.5-0.5B Actor, LLM repair in a re-implemented VA-NSF loop produced intent-mismatched commits for 32.1% of the requests, and 34.6% of its repaired commits on correctly parsed requests changed a user-stated field. IPR raised safe task success from 60.5% to 82.4% (p < 10⁻¹⁹), produced no drifted repair, and needed 1.15 instead of 2.49 LLM calls per request.