🧠 ROCK Brain Bank

TEAM ROCK — Qwen3-4B + project-specific lookup bank.
"1+x の式を登録 → 使うほど速くなる" (Boss 発案).

What is this

Adaptive learning bank for LLM FFN layers. Instead of recomputing W @ x every time, cache precomputed (x_signature, y) pairs and lookup on repeated patterns.

Architecture

Qwen3-4B (local GPU)
    ↓ hidden x
Signature quantize (PQ sub-vec)
    ↓
Index (KV/D1)  ──→  signature lookup
    ↓
CF Pages chunks  ──→  y (FFN output)
    ↓
Substitute FFN output

Stats (as of 2026-04-24)

L16 gate bank (PoC): 86,934 entries
100K tokens processed, hit rate 13.67%
Speed: 197 tok/s (RX 9060 XT DirectML)
Projected 1B tok: 85-95% hit

Design principles (from Boss)

圧縮ゼロ (no quantization lossy for W)
計算を検索に置換 (lookup over matmul)
使うほど育つ bank (永続的、ADAM 型)
CF Pages で容量無制限 (25MB/file × multi-deployment = 50TB+ 無料)

ROCK @ kagemushasystem.com · 2026