Episodes hold the contract
Each episode identifies the environment, task, initial condition, participants or agents, start and end, terminal result, and related attempts.
Streams reference one episode timebase and declare modality, encoding, sampling behavior, schema, coordinate frame, units, and validity masks.
- Stable release, episode, and stream identifiers
- Declared master clock and stream offsets
- Checksummed raw and derived artifacts
- Rights and quality decisions with evidence
Lineage connects every derivative
Normalized tensors, annotations, clips, and shards should link back to raw inputs through checksums and transformation records.
Each operation names code version, parameters, inputs, outputs, and review status. This supports reproducibility and targeted withdrawal.
Download the machine-readable files
The public JSON Schema and synthetic example can seed internal tooling. Croissant metadata and a Citation File Format record support discovery and attribution.
The example contains synthetic values only. It is not a sample from a commercial dataset.