The ~30-line LIBERO command block is duplicated verbatim in Quick start and Workflow 1 Step 3, and the full eval recurs in Step 5 and again in references; the "Core concepts" section restates VLA/Cosmos-Policy background Claude largely already knows, so it could be tightened rather than being fully lean.

Provides fully executable, copy-paste-ready `uv run` commands with concrete flag values, env exports, and working Python snippets for parsing results, matching the score-3 anchor.

Workflows use numbered steps with copy-paste progress checklists, a smoke-gate validation ("Run full benchmark if smoke passes"), an explicit validate-and-parse step, and a common-issues section with fix-to-retry guidance, satisfying the explicit-checkpoint anchor.

The two reference files are real, one level deep, and clearly signaled, but the full command matrices are inlined in SKILL.md and largely duplicated in those references, so content that should be separated into the bundle is still inline rather than appropriately split.

Lists multiple concrete actions — "setting up cosmos-policy", "running headless GPU evaluations with EGL rendering", and "profiling inference latency" — matching the score-3 anchor for specific concrete actions rather than just naming a domain.

Explicitly answers what ("Evaluates NVIDIA Cosmos Policy on LIBERO and RoboCasa simulation environments") and when via an explicit "Use when setting up..." trigger clause, satisfying both halves.

Covers natural terms a user would actually say for this domain — "LIBERO and RoboCasa", "cosmos-policy", "robot manipulation evaluation", "headless GPU", "EGL rendering", and "inference latency" — giving good keyword coverage.

Targets a narrow, named niche (NVIDIA Cosmos Policy on two specific benchmarks) with distinct triggers and explicit alternative-skill pointers, making mis-triggering unlikely.