Conformance and generalisation

Shipped reference benchmarks are inspectable conformance fixtures. Their evaluator truth may be public, so they test contract integration, replay, and scoring behavior; they do not prove blind generalisation.

Generated tiers vary explicit inputs under published rules. They still do not become held-out merely because a profile is named held_out. Sensitivity, custody, profile name, publication lifecycle, and capability maturity are separate dimensions.

Make claims at the level the evidence supports: reporting accuracy is not deployed enforcement, a design-intent trace is not a product measurement, and a frozen smoke fixture is not enterprise-scale evidence.