Epistemic Calibration and Latency Advantages of System One Probabilistic Agents in Limit Order Book Microstructure: An Empirical and Multi-Agent Evaluation

Abstract

Autoregressive Large Language Models (LLMs) operating under the deliberative System Two paradigm encounter severe structural bottlenecks when deployed in electronic financial markets: token-by-token sequential decoding induces unacceptable execution latencies (3,500-12,000 ms), unstructured text outputs trigger syntactic parsing failures, and coarse sentiment classifiers fail to identify semantic camouflage in corporate disclosures. This paper formulates, deploys, and empirically evaluates a System One probabilistic decision agent powered by TypeSafe AI's foundation model (jev-1.13.0) and optimized via Reinforcement Learning from Classifier Decisions (RLCD). By restricting the action space to a strictly typed simplex A4 = {HOLD, LONG, SHORT, CLOSE} evaluated in a single forward pass O(1), the agent achieves sub-second inference (mean latency 241.18 ms, compliance 98.8% < 500 ms) and rigorous epistemic calibration (ECE = 0.0228, Brier Score = 0.0274). On empirical Level 2 (L2) tick data from Binance Futures (BTCUSDT), the bidirectional agent delivers +739.70 bps net of friction with a Sharpe ratio of 14.92, identifying the critical latency extinction threshold at 1,485.00 ms.