DROID-100 quality audit v0.1.0
This release publishes a reproducible correction for the official droid_100/1.0.0 action description and a media-free quality-audit sidecar for all 100 episodes.
Deterministic finding
- Every decoded action vector—32,212/32,212—exactly equals
action_dict.cartesian_position(6 values) concatenated withaction_dict.gripper_position(1 value); maximum absolute difference is 0. - The published
features.jsoninstead describes the first six values as joint velocities. The release includes an RFC 6902 patch and the full all-step audit. - Under the explicitly declared path-token mapping, terminal rewards match on 100/100 episodes; this is a consistency check between related source fields, not independent relabeling.
Provisional semantic screening
- 100 consensus records, each backed by at least two distinct accepted model audits.
- Gemini and Terra cover 100/100 episodes; Claude covers 99/100. The one missing model/episode pair is disclosed rather than coerced.
- Model-derived episode labels remain candidates requiring human review. The release includes 533 typed attempt events and 52 sanitized terminal-failure invocation records, but no source media or provider response bodies.
Integrity
- Full pinned-source, telemetry, decoded-frame, raw-response replay, consensus-reconstruction, supplemental-value leak, and release-checksum gates passed.
- 170 offline tests, Ruff, strict mypy, JavaScript syntax, wheel build, and clean-wheel installation passed.
Read the report, methodology, and dataset card. Verify the downloadable bundle with release-assets-SHA256SUMS; its contents also carry an internal SHA256SUMS manifest.