How should a sample be delivered so it loads cleanly into my pipeline?

Agree a delivery layout before the first sample, so loading is routine. One workable shape:

- `manifest.json` listing files, record counts, schema version, release label and generation settings - `schema/` holding field definitions and enumerations for each record type - `records/` with one JSONL file per type: approvals, conditions, evidence, gated actions, audit events - `policies/` with the rule text and a machine-readable rule table - `tasks/tasks.jsonl` with prompts, expected state and expected-trace constraints - `snapshots/` with one folder per task for reset

For larger tables, Parquet can replace JSONL, with the same field names and types. Keep identifiers stable across revisions, such as `APP-2187` and `COND-2187-A`, so a diff between releases shows what changed.

Ask for a data dictionary that marks each field as fixed, derived or free text. Ask also for any seed or generation setting that lets a release be reproduced, if that is something MirrorFirm can supply for your engagement.

On arrival, validate the schema, check foreign keys, load a snapshot into your environment and run one task end to end before anything else. Record the release label in every result file so later runs can be traced to the exact material used.