Neuro-symbolic compliance monitors (SMT-solvers enforcing causal DAG constraints derived from regulatory rule graphs) will reduce hallucination propagation in multi-agent financial KYC/AML systems by ≥40% compared to RAG-only baselines, as measured by false-positive compliance flag rates. Amended 24 August 2026. As published, the DAG constraints were specified as "derived from ZNF740-BRD3/BRD4 transcriptional programs" — a biological anchor this hypothesis never needed and could not use. The MS targets it referenced did not replicate (CTSS non-significant; ZNF740 and DNMT1 sign-reversed in the independent GSE138614 cohort), so the anchor is now also false. The compliance claim is unchanged and stands on its own.

In a multi-agent LLM-based KYC/AML pipeline, augmenting or replacing a RAG-only compliance-flagging architecture with a neuro-symbolic monitor — an SMT solver (Z3 or CVC5) enforcing causal DAG constraints encoded from a formal regulatory rule graph (FATF/BSA/EU-AMLD rule sets) over agent outputs — will reduce the false-positive compliance flag rate by ≥40% relative to the RAG-only baseline, measured on a fixed held-out corpus of synthetic + adjudicated real-world KYC/AML cases, at equal or higher true-positive recall (non-inferiority margin ≤2% absolute recall drop), with statistical significance at p<0.05 (paired bootstrap, n≥2,000 cases).

Minimum viable test: 2-arm comparison (RAG-only vs. RAG+SMT-symbolic monitor) on a frozen multi-agent KYC/AML simulation harness, using identical underlying LLM agents (same model, temperature, prompts) so the only variable is presence/absence of the symbolic constraint layer.

• Arm A (baseline): Multi-agent pipeline (retrieval agent, entity-resolution agent, risk-scoring agent, flag-decision agent) with RAG grounding only.
• Arm B (treatment): Identical pipeline + SMT solver post-hoc verification layer that checks agent-asserted facts and flag decisions against a causal DAG of ~150–300 encoded regulatory rules (FATF 40 Recommendations, BSA/AML CFR Title 31, EU AMLD5/6 subset).
• Test corpus: stratified synthetic case generator (calibrated against SEC/FinCEN enforcement action patterns) + a held-out set of anonymized adjudicated cases with known ground-truth compliance labels.
• Primary metric: false-positive flag rate (flags raised that adjudicators reject) with paired bootstrap CI over cases.
• Blind adjudication: 3 independent compliance-SME raters, majority vote, Cohen's kappa reported.

# Arm B pipeline sketch
class ComplianceDAG:
    def __init__(self, rule_source):
        self.nodes = parse_rules_to_typed_nodes(rule_source)
        self.edges = build_causal_edges(self.nodes)
        assert is_acyclic(self.nodes, self.edges)
        self.solver = z3.Solver()
        self.solver.add(*compile_to_smt(self.nodes, self.edges))

    def verify(self, agent_claims: dict) -> VerdictTrace:
        local = z3.Solver()
        local.add(self.solver.assertions())
        local.add(*translate_claims_to_predicates(agent_claims))
        result = local.check()
        if result == z3.unsat:
            return VerdictTrace(flag=True, reason=extract_unsat_core(local))
        return VerdictTrace(flag=False, reason=None)

# Multi-agent loop
for case in corpus:
    retrieval_out = retrieval_agent(case)
    entity_out   = entity_resolution_agent(retrieval_out)
    risk_out     = risk_scoring_agent(entity_out)
    raw_flag     = flag_decision_agent(risk_out)

    if ARM == "RAG_ONLY":
        final_flag = raw_flag
    else:  # RAG + SMT
        verdict = compliance_dag.verify(assemble_claims(retrieval_out, entity_out, risk_out, raw_flag))
        final_flag = raw_flag and verdict.consistent  # suppress flags that violate DAG constraints
        log(verdict.reason)

    log_result(case, final_flag)