What is a practical first technical step after receiving a sample?

Before running any model, prove the plumbing with a scripted agent that knows the answer.

**1. Load to staging.** Read each JSONL file, or Parquet table if supplied, into a separate schema from anything production-like. Keep the manifest and its hashes with the load log.

**2. Run integrity joins.** Check that every `payments.account_id` and `open_items.linked_item_id` resolves, that outstanding amounts are never negative, and that each task points to an existing reference.

**3. Execute one task by script.** Pick a task with a one-to-many match. Call `apply_allocation` with the amounts from `reference_outcomes` and confirm the grader returns a pass with no failure codes.

**4. Break it on purpose.** Run the same task with a misdirected allocation and confirm the expected failure code appears. This tests the grader as well as the data.

**5. Check reset.** Call `reset(task_id)` twice and compare table hashes with the manifest, including the audit log.

**6. Check repeatability.** Run the scripted agent twice with the same seed and confirm identical results.

**7. Review the records.** Scan remittance text and identifiers for grounding and disclosure risk, and note what you find.

Only then connect a model. Report any schema mismatches to MirrorFirm, since adapting field names in a pilot is cheaper than adapting your harness.