What delivery formats, splits and versioning should I agree before a first sample?

Agree delivery mechanics during scoping so the first sample drops into your tooling without rework.

**Formats.** JSONL is the default for records and tasks because it streams and diffs well. Parquet can be requested for large tables. A schema file in JSON Schema form describes each record type, and a manifest lists files, row counts per file and a version tag.

**Splits.** Tasks can be delivered with a `split` field (`train`, `dev`, `heldout`) and a seed so you can regenerate the assignment. Keep held-out tasks in a separate file and out of any training job.

**Environments.** A simulated environment can be delivered as a container or a service exposing the tool actions over a defined interface. A tool-definition file in your preferred format (function-calling JSON, or an MCP-style tool list) describes names, arguments and return types.

**Versioning.** Each release carries a `release_id`. Tasks and graders reference the release they belong to, so results from two releases are not mixed by accident.

**Licence fields.** The manifest can carry a `licence_scope` note covering permitted uses. The wording of that licence comes from the agreement, not the manifest.

Final choices, including volume and delivery method, are decided in the pilot.