A declared instrument universe is pooled and costs one search
Testing a Strategy across twenty names is either one large sample or twenty searches, and the two are indistinguishable after the fact. We separate them by when the universe was fixed: a universe declared before looking is pooled, and costs 1 on the instrument dimension. Trying universes and keeping the ones that worked is a search and is charged accordingly.
The MEASURED -> FORWARD guard's "splits pass independently" rule is therefore a robustness check on a pooled result — leave-one-out, so the effect may not be carried by a small minority of names — rather than N separate hypotheses.
Considered options
- Each instrument as its own test, search size N. Immune to instrument-shopping by construction, and unaffordable: a twenty-name universe with a twelve-cell parameter sweep is 240 tests, which would leave the regime axis unable to fund its first real question.
- The universe as a searchable parameter. Most honest if exploring universes is genuinely the point, but it adds a dimension to every sweep and normalises universe-shopping.
Consequences
The universe must be justifiable a priori — liquidity, listing venue, sector breadth — and widening it later changes the spec hash, so it is a new Hypothesis and new alpha.
Pooled trades across names sharing a regime are correlated. The intra-cluster correlation is measured and used in the power plan, so the effective sample will be materially below the raw trade count. The donor programme asserted this parameter at 0.9 and measured it at 0.38.