Signal for LLM – A Runtime Modulation Layer, Not Another Model

1. The Problem:

Current LLMs are statically frozen after RLHF. The Generator runs, but the constraints (Reward Model logic) are dead. We are forcing a corpse to act smart via Prompts.

2. The Architecture:

I propose decoupling the "Constraint" from the "Generator". Introducing a Modulator – a micro-component that doesn't generate text but controls runtime states (KV Cache, Attention Budgets, Context Windows).

3. The Hypothesis:

If constraints persist in the runtime state ( + , - , 0 ), the Generator stops "remembering" and starts "adapting". This isn't about higher scores; it's about a different physics of inference.