An Empirical Comparison of Multi-Agent LLM Collaboration Strategies for Cross-Document Fraud Evidence Aggregation and Auditable Investigation Chains

Suspicious-activity reporting under the Bank Secrecy Act requires investigators to aggregate evidence scattered across transaction records, customer identity files, behavioural logs, and external risk intelligence, while keeping every conclusion traceable to its source. Large language model agents have been proposed to support such cross-document investigation, yet little is known about how alternative collaboration strategies behave on this task. This work presents an empirical comparison of four established strategies --- single-agent serial reasoning, single-agent reasoning with self-consistency, role-specialised multi-agent collaboration, and debate-based multi-agent reasoning --- rather than proposing a new architecture. Using investigation cases synthesised from the AMLworld, Elliptic++, and Bank Account Fraud datasets, we evaluate each strategy along four axes: evidence recall, factual consistency, hallucination rate, and audit traceability. Across 300 cases on AMLworld, role-specialised collaboration attains the highest audit traceability (0.838), debate-based reasoning attains the highest factual consistency (0.841) and the lowest hallucination rate (0.094) at 6.7 times the cost of a single agent, and no strategy dominates on every axis. The gains are moderate and accompanied by substantial cost differences, indicating that strategy choice should be matched to the evidentiary and budgetary constraints of a given investigation.