Abstract
<title>Abstract</title> <p> Misbooked trades and communication failures remain persistent sources of operational risk in financial markets. Recent advances in large language models (LLMs) suggest that conversational trading records may provide an additional source of information for post-trade reconciliation, yet little evidence exists regarding their ability to reconstruct executed trades from imperfect trader communications. This paper introduces RT-SIM, a simulation framework for evaluating LLMs as trade reconciliation agents under varying levels of communication degradation. The framework generates synthetic trading activity and associated trader dialogue, enabling controlled assessment of trade reconstruction performance across multiple model providers and communication conditions. Experiments evaluate reconciliation accuracy under competent, noisy, confused, and semantically degraded communication environments. Results indicate that frontier LLMs achieve high levels of trade reconstruction accuracy when communication remains explicit and retain substantial robustness to irrelevant conversational noise. Performance deteriorates significantly when communications become ambiguous or contradictory and collapses under semantically meaningless dialogue. These findings suggest that LLM-based reconciliation agents may provide effective support for post-trade control processes by transforming unstructured communications into an additional source of reconciliation evidence. Beyond improving discrepancy detection, such approaches may also support fraud detection and operational risk monitoring by helping identify unbooked positions, unsupported transactions, and other indicators of potential control failures. <bold>JEL Classification:</bold> C63 , D83 , G14 , G28 </p>