LLM Agents Can't Run Statistics Reliably—Why Determinism Isn't the Fix

As LLM agents increasingly handle high-stakes data analysis—from regulatory submissions to clinical decisions—their inherent non-determinism poses a critical reliability problem. Even with temperature set to zero, attention mechanism stochasticity and the absence of principled choice rules mean identical datasets can produce divergent statistical conclusions. This fundamentally undermines reproducibility and threatens industries that require regulatory approval or clinical validation.