Anomaly predictor

Sampling is a trade-off. Sample constantly and you burn resources and money; sample rarely and you miss the seconds that mattered. Most agents pick one rate and live with it.

This one moves between four states. In normal it ticks along cheaply. In watching it has noticed something off. In armed it raises its own sampling rate so the seconds around a failure are captured densely. Then cooldown, so a flapping machine cannot arm forever — capped at six times an hour.

Because it runs on the machine rather than in our cloud, the decision takes microseconds and needs no round trip. By the time the risk engine agrees something is wrong, the fine-grained evidence has already been collected.