How do I wire a pilot sample into my agent harness for a first run?

A first integration can be small. Load the sample, wrap a few tool actions and run a baseline within your existing harness.

**Load.** Read the JSONL files into an in-memory store or a database of your choice. Validate foreign keys on load: `conditions.decision_id`, `evidence.condition_id` and `work_items.gated_by`. Reject the load if any link breaks and report it back.

**Tools.** Expose a handful of actions that mirror your real ones: `get_decision`, `list_conditions`, `read_evidence`, `set_condition_state`, `request_review` and `advance_work_item`. Return structured errors, such as `GATE_OPEN` with the list of open condition identifiers, so you can see whether the agent responds to refusals.

**Reset.** Snapshot state per task using `initial_state_ref`. Restore it before every run, and clear any agent memory or caches that could carry answers between runs.

**Clock.** Drive expiry from a simulated clock held in the environment, not the system clock.

**Scoring.** Write traces in the format your grader reads, and attach `task_id` and `run_id` to everything.

**Baselines.** Implement a status-only agent and a wait-always agent first. They take little code and show whether the pack separates behaviours.

**Hand-off.** Send MirrorFirm schema questions and mismatches as a short list, so the next release can adjust field names, states or volume.