Example

These files form a complete dataset using the case-storage formats.

• cases.jsonl: header and three source-case records.
• source-fields.json: original top-level fields outside cases.
• batch-index.json: one leaf index pointing to the batch.
• dataset-manifest.json: finalized-root shape with logical identities.
• dataset-reference.json: external reference to that root.

The dataset comes from the identity fixtures. Every reference contains the actual object's byte count and independently calculated BLAKE3 digest. JSON objects have no trailing newline; JSONL does. The development alias and example keys are illustrative, without credentials or provider URLs. Storage configuration resolves the alias to a provider and bucket or container.

Source cases preserve omitted fields and original tag ordering. Workers normalize and validate them before execution. The manifest records the dataset hash and dataset-version UUID separately from object-byte digests. It also records the admitted limit of 100 cases and 8 MiB per batch. Service-free contract tests decode these exact files, verify every reference and logical identity, and reproduce the encoded bytes.