Hardware Entropy Injection for Behavioral Divergence in LLM Inference: The PSE Framework on IBM POWER8

We present a method for injecting hardware-sourced entropy into LLM inference to produce provable behavioral divergence using the IBM POWER8 mftb (Move From TimeBase) instruction. Key results: Provable divergence: 3 runs with identical seeds produce 3 distinct MD5 hashes 0.2% overhead: burst strategy (every 4th token, top-512 only) is nearly free 8.81x combined speedup (16.74 to 147.54 t/s) with full PSE stack 4 behavioral metrics defined: NOI, DR, ACS, MCI for entropy-mediated quality Grounded in Hebbian learning theory and biological stochastic resonance. Part of the Proto-Sentient Emergence (PSE) framework.