Controller family

Evidence-scoped guidance without a counterfactual shadow race.

The active default is retained_residual. It is the conservative production member and the only member selected by the advisory policy. anderson_residual is available only through explicit fixed policy and carries a SIMULATED evidence tier.

Automatic switching is disabled. A candidate configuration that was never applied has no realized outcome, so scoring shadow controllers on the incumbent trajectory would be counterfactual. The production layer does not present that as a free experiment.

The latest frozen primary ladder confirms retained_residual and anderson_residual at 30 of 30 seeds through 65,536 controls. diagonal_secant falls to 1 of 30 at 4,096 and 0 of 30 at 16,384 and 65,536; commissioned_pi has a width exponent outside the predeclared bound. Neither can become active in production.

A future on-policy family race would spend ordinary maintenance cycles and contract risk. It must be explicit, reversible, attested, and validated before it can become a production default.

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