openbot-data 0.0.3 · updated 2026-08-12

Why a Sample Audit Is Not Full Dataset Readiness

Compare sampled and full-integrity robot dataset audits without overstating what a fast preflight actually verified.

Run the check

openbot-data audit ./lerobot_dataset \
  --format lerobot \
  --integrity sample \
  --out ./sample.audit.json \
  --fail-on error

openbot-data audit ./lerobot_dataset \
  --format lerobot \
  --integrity full \
  --out ./full.audit.json \
  --fail-on error

Machine-readable output: sample.audit.json and full.audit.json (`openbot.dataset_audit.v1`)

What this proves

A sampled audit can quickly expose structural and selected media problems.
The output records coverage and skipped capabilities instead of converting missing checks into a pass.
A full-integrity run expands evidence across the complete local source available to the tool.

What it does not prove

That an unexamined frame, episode, or shard is valid after only a sample run.
That no semantic labeling or task-quality problem exists.
That a zero-error audit is equivalent to policy-specific readiness.

Decision checklist

01Use sample integrity for bounded triage and label the result as sampled.
02Use full integrity before release, merge acceptance, or an expensive training run.
03Run a policy profile separately when training compatibility is the actual decision.
04Preserve both artifacts if the sampled run determined whether deeper inspection was warranted.