# Safety Contract

The book distinguishes **prediction research** from **financial execution**. A well-calibrated probability is evidence for a research decision; it is not a trading command.

The reference implementation is limited to data collection, model training, evaluation, deterministic replay, and paper trading. It must record source timestamps, model versions, configuration hashes, replay assumptions, and every safety decision.

An agentic monitor may take only reversible, bounded actions:

1. stop a data consumer when a contract fails;
2. quarantine a corrupted batch;
3. mark a model candidate ineligible;
4. rerun a named validation job;
5. open an incident with evidence; and
6. request human review.

The monitor must not hold credentials capable of placing, canceling, resizing, or changing live orders. Every remediation action requires an allowlisted runbook, an audit event, a rollback path, and a human escalation condition. If evidence is incomplete, the default action is to stop or abstain.