Signal for LLM
Runtime state modulation layer for large language models.
Not another model. Not another training trick. A decoupled Modulator that writes to the engine while the Generator stays ignorant.
Status
Theory — complete. Architecture — defined. Interfaces — audited.
This repository is a release, not a proposal. The achievement is the blueprint. Experiments are an engineering hobby that follows.
One Sentence
Current LLMs are statically frozen after RLHF — the Generator runs, the constraints are dead.
Signal for LLM puts constraints back into the runtime, where they belong.
The Decoupling
| Role | What it does | Size |
|---|---|---|
| Generator | Produces tokens | The whole model you already have |
| Modulator | Modulates runtime state (KV Cache, Context Window, Expert Routing, Inference Budget…) | Can be + / - / 0 |
| Reward Model | Learns constraints during training | Already exists |
The Modulator is not the Generator. The Modulator is not the Reward Model. It is the third entity that survives deployment.
📎 The Generator does not read the Modulator's output. It only feels the causal drift in runtime state.
If the Generator tries to "understand" the Modulator, Signal degrades into Prompt Engineering. Don't let it.
What Signal Is Not
Signal is not a Prompt. Signal is not an Embedding. Signal is not information.
Signal is runtime state modulation that actually changes the machine. If it can only be read as a symbol, it's a projection. Not Signal.
Repository Map
This repo carries the English release of Signal for LLM (engineering architecture). The TGR theoretical system (10 docs, Chinese-primary) is the upstream; this repo is the downstream engineering surface.
| Layer | Doc |
|---|---|
| Engineering entry | Signal for LLM |
| Spec | Signal |
| Mapping | Signal Mapping |
| Taxonomy | Signal Taxonomy |
| Runtime | Signal Runtime |
| API | Signal API |
| Engine | Signal Engine (evolving) |
Full TGR index → Live site
License
CC0 1.0. No rights reserved. The architecture is a fact; facts don't belong to anyone.
"Proof"
There is no benchmark to chase.
If modulating runtime state changes generation — and it does — the only remaining question is engineering pace, not existence.
🔻 Modulator proposed → architecture complete → experiment optional.