SkillsBench challenge; Terminal-Bench 2 out of the arena; gates check SkillsBench; practice off the board

Benchmarks: rank and plot one benchmark at a time, per domain where it has domains

Compose run configs from model, suite and recipe fragments

Challenge tb2-9b v1: sealed 32-task TB2 subset, one serving GPU, subset base-model reference

Challenge runs: pinned recipe, budget reservation, one a100x8 job running the real pipeline, recomputed collection

Add agent API instructions, complete recipe packages, and coherent task split

Upload folder using huggingface_hub