37 S100P jobs for comparing base Qwen3.8-27B with its GLM-5.3-SFT version,
GLM-5.3 as reference: 4 controls GLM passed on the first compile, 26 GLM
passed only after 2-9 compiles, 7 harder probes (pose, depth, instance and
semantic segmentation). Each model runs 2 rollouts on the same harness;
only rollouts started after the 2026-10-08 fixes count, and system
interruptions are rerun rather than scored.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>