From 26eaaec36c81adebe44c141910d046a0b7e57b09 Mon Sep 17 00:00:00 2001 From: Advay Monga Date: Sun, 4 Oct 2026 22:41:40 -0400 Subject: [PATCH] docs: direction is load-adaptive across all corpus classes; cold start no longer the headline Co-Authored-By: Claude Opus 5.5 (1M context) --- CLAUDE.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/CLAUDE.md b/CLAUDE.md index c4823ea..b44a951 100644 --- a/CLAUDE.md +++ b/CLAUDE.md @@ -20,7 +20,9 @@ ## Models - Hand-written forwards: `models/gemma4.py` (Gemma 4) and `models/qwen3_moe.py` (Qwen3-MoE); `CustomTorchBackend` picks one by the HF `model_type`. Qwen3-30B-A3B is the benchmark model - for MoE cold start (owner's Phase 0 decision, 2026-10-01); its path runs decode CUDA graphs (2026-10-04) but not torch.compile or int8 yet. + (owner's Phase 0 decision, 2026-10-01); its path runs decode CUDA graphs (2026-10-04) but not torch.compile or int8 yet. +- Direction (2026-10-04): one engine that adapts to load across every corpus class; cold start is + one class among them, no longer the headline. The headline metric is not decided yet. ## Strict referee, free player (the bitter lesson) - Harness is code and strict: sandbox, grader, measurement protocol, output equivalence,