Note Skills Anywhere 0.9.0: review scripts and resources with directory manifests, compare changes locally, and pin MCP loads to reviewed bytes.

Note Robot Reel 0.10.0: compare six Genesis/Newton GPU flights with analytic motion; inspect all 366 states, native replays, OpenUSD and five offline-ready labs.

Note Inspect 30 trial rows and 20 paired rows from one controlled SmolVLA task. Includes full outcome groups, source hashes and reproducible export instructions. These descriptive results are not a training split or official benchmark.

Note A better score can hide a regressed check. Inspect the recorded retry, then use the first-review guide to verify the evidence locally. Scripted controls; no model ranking.

Note A filterable evidence casebook: 167 audit cases, six repeated attempts and three suite jobs in separate development configurations. Compare acceptance with full resolution; inspect unchanged JSON and provenance. Scripted controls, not model rankings. Maintainer dataset.

Note 45 actual GPU agent trials, native Blender edits, OpenEnv controls and paired LIBERO-Plus recordings. All failures retained; small public development pilots, not a leaderboard.

Note 36 fixed SWE attempts with raw model/tool traces, patches and native checks. 31 assessable reports, five uncertain outcomes; no accepted attempts or demonstrated skill benefit.