Generation Before Acquisition: Decision-Relative Capability Transitions in Costly Information Systems

A system can spend more on measurement without becoming able to answer a decision that its current observational repertoire cannot resolve. We develop a decision-relative framework that separates the cost of generating measurement capability from the cost of acquiring evidence with that capability. A generation budget G induces a reachable experiment closure C_G; the decision kernel K_D(G) contains decision-incompatible worlds that remain indistinguishable under every reachable experiment. For finite deterministic systems, we prove that exact finite resolution is possible if and only if this kernel is empty. The framework distinguishes null, intensive, and extensive capability transitions, with extensive transitions able to change a problem from non-resolvable to resolvable. We separate nonadaptive batch acquisition from adaptive decision-tree acquisition, establish persistence monotonicity, characterize finite discrete capability frontiers as lower-envelope step functions, and decompose capability leverage on bounded-variation frontiers into absolutely continuous, jump, and singular-continuous components. Exact computational audits cover all 256 experiment-family subsets of a finite audit universe, test persistence and its failure, verify frontier boundary cases, and reproduce the leverage identities. The resulting perspective separates more evidence from more measurability: capability investment matters not only when it lowers the cost of an answer, but when it changes which decisions can be resolved at all.