What formats and intake steps suit a pilot delivery into our evaluation harness?

A pilot delivery can be shaped to fit what your harness already reads.

**Formats**

- JSONL for cases, records, policies and notes. - Parquet for larger tables, with a schema file. - A snapshot format for the environment, such as SQL dumps or per-table files, if a stateful run is in scope.

**Manifest**

A `manifest.json` could list each file, its record count, its schema version, a checksum and the split it belongs to, such as `dev` or `held_out`. Ground truth is listed separately so you can restrict access.

**Intake steps**

- Validate every file against its schema and confirm that identifiers join across files. - Check that dates, currencies and time zones parse as declared. - Confirm that no agent-visible file contains the root cause or answer. - Load a handful of cases by hand and solve them without ground truth.

**First run**

- Run a simple baseline and a stronger configuration on the same cases. - Record component scores and event logs. - Review the disagreements between automated scores and human review.

**Versioning**

Pin the release version in your results so later changes to schemas or cases do not alter earlier comparisons. Agree delivery method and any change process with MirrorFirm for your engagement.