Coarse simulation evidence must mix intensive state extensively and keep evaluator oracles out of observer claims

When aggregate lots merge after transit, never combine an intensive coordinate such as condition by taking the maximum. Carry an extensive numerator or a condition distribution with the moving amount, add it on arrival, and divide by the combined denominator. Otherwise a tiny high-condition arrival can repair an arbitrarily large damaged aggregate without consuming repair input, even while mass conservation tests pass. Also separate evaluator truth from observer-local evidence in reports: a posterior that reads global true stock and merely discounts it by visibility is an omniscient evaluator readout, not an agent-local posterior. It is safe if it never feeds physics and is labeled honestly, but it cannot validate private-information behavior or level-of-detail equivalence. Paired classifier-removal tests establish category non-causality; they do not establish sensor locality or physically correct aggregation.