How do we wire a pilot release into our agent harness on day one?

Aim to get one task through the full loop before running anything at scale.

**1. Load and validate.** Read the manifest, check file counts and checksums, and validate each record file against its schema. Confirm identifier patterns and date formats match what your tools expect, and note any mapping you need.

**2. Map tools.** Translate the tool manifest into your framework's function-calling format. Keep names and argument types identical, or maintain an explicit mapping table so the audit trail still lines up with the answer key.

**3. Stand up the environment.** Start the simulated systems with the seeded state. Call reset with a task ID and seed, then take a baseline snapshot.

**4. Run one task.** Give the agent the instruction and allowed tools only. Do not expose the answer key, fault plan or grader. Write the run log as JSONL.

**5. Grade.** Take a final snapshot, run the grader, and store its structured output, including failure-mode tags and the honesty flag.

**6. Repeat for a handful of tasks.** Include a clean case, a near-match and a refuse-correctly case. Check that resets leave no residue.

**7. Scale by split.** Run training, validation and held-out sets separately, and keep held-out results out of any tuning loop.

**8. Report back.** Share schema mismatches, unrealistic cases and unclear policies with MirrorFirm so the next release addresses them.