# E21w: 3-bit KV quant variants on real Qwen3.5-4B f16 KV (CPU only)
Data: E21t f16 slot dumps (saved on the test rig), first 6000 tokens of each of the 8 full-attention layers, sliced by slice.ps1 (slices not published; SHA256SUMS kept locally). Note: first slice attempt had a PowerShell $n/$N case-collision bug (K layers 1-7 = layer 0, V = zeros); fixed and re-sliced, verified per-layer stds.
Run: python3 e21w.py (≈8 min, 8 cores) -> results_raw.json ; python3 analyze.py -> results.json. Rerun verified byte-identical (sha in run1.sha).
Queries are a PROXY (post-RoPE keys at 128 positions in [3000,6000), causal over earlier keys). Matched Q capture still needed.
Prior art: KVQuant, KIVI, QServe SmoothAttention, KVTuner, QuaRot. No novelty claim.