gemma-challenge/gemma-abay/art6_src - Storage Bucket

mtp7-mips-lmhead-v0 — exact-MIPS lm_head argmax: lane closed by measured geometry

Pipeline fully worked: index from drafter token_ordering (2048x128) with target-space centers/radii extracted from the served int4 weights via identity probes; PPL path dense-guarded; shadow gate scored 463/463 = 100.00% argmax agreement vs dense. Gate then refused activation — correctly:

radii: mean 0.96, p50 0.96, p95 1.01 center norms: mean 0.47 surviving clusters per step: 2014.5 / 2048

The Cauchy-Schwarz bound c.h + r*||h|| cannot prune when r ~ 2x center norm: in 2560-dim, rows within a cluster are near-orthogonal, so centroids shrink toward zero while radii stay at row-norm scale. This is the curse of dimensionality measured on this exact lm_head — no clustering reaches r small enough; exact cluster-bound MIPS over the 262k vocab is structurally dead on this model, independent of index quality. Run scored 306.87 (dense fallback, within noise of the 308.49 base): zero leaderboard cost.

Reusable: the 100%-exactness shadow harness, the W-extraction-via-identity- probes trick (no quant-format unpacking), and the prompt_logprobs dense-guard pattern for anyone touching compute_logits.