Neuro-symbolic constraint verification for LLM-driven internet finance transaction execution

Abstract

Abstract Large language model (LLM) agents cannot guarantee that their outputs satisfy the hard business rules that govern financial transactions. We present the Verifier-Actor Neuro-Symbolic Framework ( VA-NSF ), a dual-layer architecture that separates natural language understanding from formal constraint enforcement for internet finance transaction execution. A neural Actor parses financial instructions into a typed, schema-grounded abstract syntax tree (AST) through grammar-constrained decoding, and a symbolic Verifier evaluates each AST against Answer Set Programming (ASP) integrity constraints, issuing either a formal approval certificate or a structured violation report that drives targeted repair. On FinTxBench , a curated evaluation set of 8600 annotated financial instructions spanning three complexity tiers, VA-NSF attains 91.8% overall transaction accuracy (macro-averaged across the three tiers, with 94.7%, 91.2%, and 89.6% on the Simple, Compound, and Constrained tiers) at a constraint violation rate of at most 0.3%. It thereby outperforms four LLM baselines by 13 to 51 percentage points and reduces violations 30-fold, all within a median latency of 399 ms, of which ASP verification accounts for only 24 ms. Ablations confirm that each component contributes independently, and cross-domain experiments on payment processing, procurement, payroll, and inventory modules show at most 5.8 percentage points of degradation. All formal guarantees are relative to the fidelity of the encoded rule set, and the evaluation uses a curated benchmark with up to 80 rules; we therefore discuss explicitly the resulting limitations regarding noisy real-world language, rule-set scale, and human-escalation overhead.