Train and validate before hardware time — reproducible seeds and metrics for regression detection.
Compare runs with fixed seeds and metric thresholds — catch drift before it reaches the fleet.