[Experiment] Compression ratio sweep on GPT-OSS-120B and Gemma-4-26B-A4B (optional) #17

Repeat the arm 1 setup — calibrate on 64 samples of simplescaling/s1K-1.1, evaluate on GPQA-Diamond — at p ∈ {25, 50, 75}% on the two new model families only.

Scope

• GPT-OSS-120B and Gemma-4-26B-A4B, Table 1 layout, one block per ratio.
• Same rows as Table 1: base model, SlimQwen (prune-only), IntDim-E, IntDim-L, IntDim-G. The p = 50% block is the one already produced by [Experiment] Knowledge track: two calibration formats and their out-of-domain cost #15.

Depends on: #7, #12, #13, #14

Done when

• The sweep table is filled for both backbones at all three ratios.

Part of #2