Hypotheses come in two tiers: mechanism and implementation

ADR-0001 required a Strategy to cite a resolved Hypothesis without saying how many Strategies one Hypothesis licenses. Both extremes fail. Free lets a single confirmed mechanism justify twenty deployed variants, which is the multiple-comparisons problem re-entering through the deployment door. Strictly one-to-one means any spec change — a different universe rule, a wider disaster stop — requires re-establishing the mechanism from scratch at full alpha, which makes ordinary iteration unaffordable and invites working around the rule.

The donor register already contains both kinds of question. Mechanism: X2-OVERNIGHT-GAP, C2-MAE-PREDICTABLE, X1-COMPRESSION. Implementation: Z04-ORB-OBTAINABLE, Z-ENTRY-IMPLEMENTABLE, E3-COSTS.

A mechanism Hypothesis establishes the effect and carries the large alpha. Each Strategy built on it registers its own implementation Hypothesis — does this spec clear its floor net of costs and obtainability — at a smaller declared allocation, and cannot exist without a resolved mechanism parent.

Consequences

The family stays bounded and countable. Variants cost something, which stops twenty of them appearing; they do not cost everything, which stops the discipline being abandoned as unworkable.

The mechanism and implementation per-test allocations are numbers that must be justified rather than derived, and they are the obvious place for the discipline to erode. They belong in configuration beside each axis budget. Both draw from that one axis budget after search correction; they are not additional top-level pools and are therefore never counted twice. Config load enforces sum(axis budgets) + reserve == total, and registration refuses any spend that would exceed the remaining axis budget.