Baselines & member evidence

What the latest frozen campaign confirms, refutes, and leaves open.

CONFIRMED MEMBERS: retained_residual and anderson_residual completed all 210 registered simulator runs. Both remained 30 of 30 through the 65,536-control primary scale ladder with a flat acquisition exponent. The fixed-window member remains explicit opt-in because deployment selection itself was not a manuscript experiment.

REFUTED PROMOTION: diagonal_secant fell to 1 of 30 at width 4,096 and 0 of 30 above. commissioned_pi showed a width exponent outside the predeclared confirmation bound. Neither is selectable in production, even in a declared low-drift cell.

SHARED TERRITORY: a current exact model, a commissioned specialist, or a simpler controller can match or beat an in-class member when setup is cheap and the plant stays stable. The maintenance layer should report that territory honestly rather than manufacturing an exclusive claim.

INFORMATION BOUNDARY: dense finite differences and full-model methods pay growing acquisition, arithmetic, or state costs. Scalar optimizers receive a different information interface when they see one score instead of one component-resolved vector. Compare methods under the same information and acquisition budget.