ACES captures trajectories, tool use, skill discovery, and regressions that structural lint and LLM-judge document scores miss. NVIDIA open-sourced SkillEvaluator as its reference implementation on the nvidia/skills repository with continuous benchmarks.json updates. August 2026 data showed Tier 1 scan scores correlated at Spearman ρ ≈ -0.02 with live Skill Lift — scans are necessary for structure but insufficient to prove a skill helps in production.