Related-task warm-starting

Carry useful state across a stream of nearby problems.

This page documents an optimiser-era simulation. The consumer it serves is an argmax or single-score objective, which the layer's own hardware evidence shows is insensitive to systematic drift (+0.7 points against +34.0 for a numeric consumer). It is kept as a record; see Limitations and Preflight for the consumer-class preconditions before relying on it.

The configuration is part of the loop state. Do not throw it away when the next target, frame, or instance is related to the last one.

Warm starts are useful for tracking, repeated response matching, related-target control, and drift recovery.

When tasks are unrelated, reset explicitly. Retention is valuable because it is honest state, not because every prior configuration deserves to survive.

layer = StabilityContract(key, n=len(x0), mode="regulation", target=target)
x = layer.start(x0)
for _ in range(rounds):
    x = layer.step(measure(x), target=target)
proof = layer.end()