Suppressing Sycophancy Cascades and Trauma Lock-In in Autonomous Financial Multi-Agent LLMs: Bounded Adversarial Deliberation, Exponential Time-Decay Episodic Memory, and Foundation Model OOD Veto

Multi-agent Large Language Model (LLM) architectures deployed in non-stationary financial markets exhibit two fatal cognitive failure modes when transitioning from static backtests to continuous live execution: (1) Sycophancy Cascades, wherein adversarial agents (Bull and Bear researchers) prematurely converge toward a shared narrative echo chamber when exposed to high-salience macroeconomic headlines, inflating position sizing immediately prior to regime reversals; and (2) Trauma Lock-In, wherein episodic memory banks that store historical stop-loss events without temporal attenuation permanently paralyze the execution engine after a tail-risk shock, causing the system to miss subsequent V-shaped recovery rallies. In this paper, we formulate a three-layer cognitive control architecture implemented within the open-source lumina-quant framework: (i) a Bounded Adversarial Deliberation Protocol equipped with an explicit Anti-Sycophancy Penalty that discounts consensus conviction whenever semantic overlap between opposing agents lacks independent evidentiary support; (ii) an Exponential Time-Decay Episodic Memory Retrieval kernel (half-life tau = 14.0 days) that preserves acute post-crash caution while mathematically guaranteeing asymptotic recovery from historical trauma; and (iii) a Zero-Shot Time-Series Foundation Model Out-of-Distribution (OOD) Quantile Veto (10th to 90th percentile trajectory cone) that acts as a physical reality check against LLM narrative hallucinations. Across a 300-episode controlled multi-regime evaluation spanning crash shocks and post-shock recoveries, the unified architecture reduces maximum drawdown by 54.2% while eliminating post-crash paralysis and increasing net risk-adjusted return (Sharpe ratio +0.94) over unregularized multi-agent baselines.