Available regulatory mutation directions improve finite-budget adaptation from matched states in a developmental-network model

The organization of a genotype-to-phenotype map can alter which variants become available to selection, but comparisons among independently evolved networks often confound present mutation access with ancestral performance and architecture. We used a 16-trait developmental-network model to ask whether the number of adjustable post-training regulatory directions changes adaptation from matched states under a fixed mutation-and-selection budget. Historically trained cores were copied into nested treatment arms with either 2 or 32 adjustable regulatory modes (`d2` and `d32`). The modes were mathematical directions in a 16-by-16 regulatory matrix, not genes or additional network nodes. Arms began with identical physical matrices, phenotypes, scores, direct states, and realized wiring costs. Mutations were proposed from paired random tuples, and each trajectory received 1,000 candidate slots. An initial matched-core study found a `d32-d2` final raw-benefit difference of 0.1336 (95% paired-unit bootstrap interval 0.1227 to 0.1442) in a separate 48-unit cohort. A structural follow-up retained positive effects across two target families and dense versus single-edge direction bases, but its nominal cumulative-displacement bound never rejected a proposal. We therefore prospectively tested smaller, actively exercised bounds and matched midpoint interventions. In the completed 48-unit-per-family validation, the prespecified higher-order-family, single-edge, radius-0.50 primary contrast was 0.0103132613 (ordinary 95% interval 0.0082893286 to 0.0124080783) and the bound was active in both capacities. All 16 prespecified capacity contrasts and all eight separately adjusted midpoint contrasts were positive. Opening 30 zero-valued directions after an identical `d2` prefix improved subsequent performance, as did retaining the mutability of directions 3–32 after an identical `d32` prefix. In the four exploratory within-C48 comparisons, the R050-minus-BOX capacity contrast was negative with every ordinary 95% interval below zero; this describes R050 attenuation in those paired cells, not a monotonic constraint-strength effect. Exact target attainment remained rare in difficult cells, and realized cost and historical retention were not uniformly improved. These results identify a conditional effect of available mutation directions and their associated proposal probabilities within the specified nonlinear model, constraints, costs, and finite horizon. They do not show that dimensionality is universally advantageous, that the intervention uniquely explains the full capacity effect, or that regulatory capacity originates or is maintained by natural selection.