Domain LLM bridge (MathNN + Gaming RL)

Slow-path LLM helpers around MathNN and Gaming RL. Not used as the per-tick policy. Kernels, A*, and MathNN remain ground truth.

MathNN

llm math route how do I use Stokes theorem
llm math curriculum math_algebra 8
llm math distill math_algebra factoring quadratic 48
llm math explain math_algebra 0.2 1.5

• route — keyword + LLM map utterance → math_* domain
• curriculum — JSONL problems under data/mathnn_curriculum_*.jsonl
• distill — LLM action hints → supervised MathNN warm-start on smoke env
• explain — narrate training metrics

Gaming RL

llm gaming coach deaths=3 kd=0.8 objective=B
llm gaming callout low ammo mid rotate
llm gaming scenario fps 5

Lua

ai_llm_math_route("eigenvalues of a matrix")
ai_llm_gaming_callout("enemy on A site")

Design rule

LLM shapes inputs/outputs of the RL loop (curriculum, coach, routing).
MathNN / Gaming RL act in the closed loop.