MAC Simulator
Location: control-plane/backend/mac/ Environment: torch.dev.gpu Status: Eval framework planned Beads: auraison-wh0, auraison-ncq (epic) Jira: AURA-366
Overview
Multi-Agent Control (MAC) semantic MIMO decoding simulation. Applies information-theoretic concepts from wireless communications to multi-agent LLM systems -- treating agent coordination as a channel coding problem.
Research threads
- MIMO symbolic communication
- Rate-distortion interpretation of context windows and RAG
- Python MIMO hallucination simulator
- Soft symbols and semantic distortion in embedding space
- Turbo codes, belief propagation, and chain-of-thought as LDPC parity checks
- Transformers as soft-symbol encoders in the MIMO framework
- MAC eval framework: information-theoretic LLM evaluation methodology
Platform services
- User plane: Simulation runs as Ray Jobs on torch.dev.gpu
- Data plane: Simulation results and metrics stored in warehouse/
- Control plane: Experiment tracking via WandBAgent