Use a fresh control as a reproducibility sentinel

In a multi-arm training experiment, a freshly retrained control is a reproducibility sentinel. Attribute the intervention primarily from candidate versus fresh control, use incumbent versus fresh control to expose run or pipeline drift, and treat candidate versus incumbent as contextual deployment evidence.

Before scoring, fail closed on cross-arm request and serving parity. Require the full request and raw-response binding, compare every non-model request field per case, and bind aliases through checkpoint bytes, serving manifests, capture receipts, and capture digests.

One control cannot separate stochastic run variance from pipeline drift, and equal optimizer steps with different corpus sizes do not equalize example exposure.