5/13
4. Comparability is asserted, not measured
The claim that a model's residual risk is "comparable to the risks of existing models already deployed across the industry," supported by the observation that comparable capability is available from other public systems and used daily by defenders, addresses absolute capability. The quantity that governs an export-control determination is marginal capability: the uplift a deployment confers over the counterfactual in which it does not exist. Formally, this is the difference between the probability of adversary success given the model and that probability given its absence, holding the adversary's other resources fixed. A capability that is ubiquitous confers little marginal uplift; a capability that is differentially superior confers a great deal, irrespective of whether weaker forms circulate elsewhere.
The comparability defense also smuggles an accessibility assumption. "Available elsewhere" bounds the counterfactual term only if the elsewhere-capability is equally accessible to the relevant adversary. Availability of a competing system to a compliant developer is not availability to a sanctioned state actor operating under interdiction; market presence and adversary access are distinct quantities, and the defense conflates them. The statement measures neither the differential nor the accessibility. It substitutes the existence of the capability somewhere for the model's marginal contribution to the specific actor of concern, a substitution that holds only if both the differential and the access gap are zero.
This hinge is symmetric, and intellectual honesty requires saying so: the differential-uplift question is also the load-bearing premise of the case for control. If the differential is null and the capability equally accessible, comparability holds and the directive overreaches; if either fails, comparability fails and the directive is grounded. Both positions stand or fall on the same unmeasured quantities (see §11).