Which file formats and schema conventions suit an escalation dataset?

Choose formats by how the data will be read, and fix the conventions before generation so every table joins cleanly.

- **JSONL** for cases, events and tool traces. One object per line streams well and diffs cleanly in review. - **Parquet** for large tables such as exposure links or balances, where columnar reads help bulk analysis. - **JSON Schema** files for each table, so a loader can validate a release before ingesting it.

Conventions worth agreeing up front:

- Stable identifiers with a type prefix, such as `APP-2187`, `FAC-5530` and `CASE-0412`, used as join keys across files. - Money as integer minor units plus a currency code, which avoids rounding disputes in aggregate calculations. - Timestamps in ISO 8601 with a timezone, so event ordering is unambiguous. - Enumerations for status and event types, listed in the schema rather than free text. - A `manifest.json` listing files, schema versions, row counts, the generation seed and the train, dev and test split assignment.

Keep expected outcomes in their own file, separate from the case files, so a team can withhold them from an agent's context and from any training split.

Formats are negotiable. If your harness expects a different layout, such as a single nested JSON per case, raise it during scoping and test it on a sample first.