The regime classifier is frozen before any strategy question

Regime-conditional strategies are the first proposer, and a regime classifier carries parameters of its own — lookback, threshold, index-level or per-instrument. Tuning those until a strategy works is a search that search_space_size never sees, because it happens inside a single hypothesis: the declared space says 12 while the real space was 12 times the classifier space. Consumed-observation and search-budget accounting cannot detect this hidden search. We therefore commit the classifier in its own registration, before any strategy hypothesis exists, on a priori grounds, calibrated only on an early definition period that is then permanently excluded from strategy measurement.

Considered options

• Classifier parameters inside each strategy's declared search space. Honest, and sacrifices no history. Rejected because the multiplied search size makes the per-test alpha tiny under correction, which would leave the first real question either unaffordable or underpowered.
• Classifier declared free per hypothesis. Cheapest and most flexible, and precisely the failure this apparatus exists to refuse: nothing prevents choosing the classifier that makes the answer come out, and no guard can see it happening.

Consequences

A slice of history is spent on definition and can never be measured on.

Two constraints follow and are not optional. The label must be causal — computable at time t from data no later than t — enforced through the same point-in-time gate as everything else, not checked by eye. And whether regime is index-level or per-instrument is itself a declared parameter, because an index-level regime makes every concurrent trade share a common factor. The resulting intra-cluster correlation must be measured and used in the power plan. The donor programme asserted that correlation at 0.9 and measured it at 0.38.