Eligibility as a Constraint, Not a Footnote

Most planning systems treat export control as a note attached to the answer. This post argues that it belongs inside the question, and shows what that costs to implement, which is almost nothing.

What the rule actually says

[OBSERVED] The governing instrument at the time of writing is the Bureau of Industry and Security rule published at 91 FR 1684 on 15 January 2026, RIN 0694-AK43, which amends 15 CFR 742.6(b)(10)(iii)(A)(1). The preamble names the client’s H200 product. The rule conditions licence eligibility on a certification that aggregate shipments of a given product to China and Macau do not exceed 50 percent of the volume of that same product destined for the United States.

Read carelessly that is a restriction on China. Read carefully it is a coupling between two regions, and the coupling is the part a planning model has to carry.

Three lines of algebra

Written as a constraint, for product p, with the decision variable x indexed by destination region, the rule is a linear inequality on the decision variables themselves.

For every product p in the controlled set:
Σ x[p, r] for r in {China, Macau}  ≤  0.50 × Σ x[p, r] for r in {United States}

Three properties follow, and each of them matters to the design.

It is linear. The constraint does not change the solver class. A linear programme stays a linear programme, so the cost of correctness here is one row in the constraint matrix.
It couples two regions that nothing else couples. United States allocation silently caps China allocation. The shadow price on the coupled constraint is the figure a planner needs to see and the figure a post-solve filter can never produce.
It is a ratio rather than a ceiling. A reduction in United States volume reduces permitted China volume without any regulatory change occurring. That is the failure mode a filter applied after the solve would never surface, and it is the reason the rule belongs inside the model.

WHY NOT A FILTER

A filter applied afterwards produces a plan that was optimal for a market the client may not serve. A constraint applied inside the solve produces a plan that is optimal for the market it does serve. The two differ by more than presentation.

Why a price test alone will not do

The optimizer proposed in this project uses bid prices to value the next unit of constrained capacity. Akan and Ata prove that under a bid price control the bid price process is a martingale (Akan & Ata, 2009, Theorem 5.1, p. 921), which is the formal result behind treating a shadow price as that value.

They also show, by counterexample, that no pure price test can be optimal on its own, and that a capacity usage limit is required alongside it (Example 2, pp. 918 to 919; Definition 3.3, p. 919). The 50 percent export ratio is a second such usage limit. The regulation and the theory arrive at the same structural requirement from opposite directions, which is a reassuring place for a design to sit.

Withdrawn from enforcement, not rescinded

[OBSERVED] A second instrument remains live and is recorded because its status is unusual. The Artificial Intelligence Diffusion rule published at 90 FR 4544 was withdrawn from enforcement by agency announcement rather than by a published rescinding rule. The text remains in the Federal Register.

A planning model should therefore carry the Diffusion rule thresholds as a dormant scenario rather than delete them, because reinstatement would require no new rulemaking. Deleting a rule that still exists is the kind of tidiness that costs a quarter to undo.

A decision cannot be audited against a rule that has since been edited

[CONVENTION] Scope parameters are versioned and dated in the same way as model weights. When scope changes, the system selects a precomputed scenario rather than reforecasting, and the run record retains the scope version that produced every historical decision.

That last clause is the point of the whole section. A decision cannot be audited against a rule that has since been edited in place. Either the rule is versioned or the audit trail is decorative.

Provenance and marks

This post is part of the NVIDIA capstone for the M.S. Business Intelligence program at Full Sail University. It originates in BIN520-O, the second course of the program, and therefore predates the other capstone posts published here. Epistemic marks appear before the claim they govern: [OBSERVED] read from a primary source and verified against it, [MODELED] produced by a stated method from stated inputs, [INFERENCE] reconstructed from indirect evidence, [CONVENTION] a customary figure with no published benchmark behind it, [UNVERIFIED] asserted somewhere but not resolved to a locator.

Non-affiliation notice. The author is not employed by, affiliated with, sponsored by or endorsed by NVIDIA Corporation. NVIDIA is named here as the subject of an academic analysis prepared by an independent party, from public disclosure only. No confidential information of NVIDIA Corporation was used, and no representation is made on its behalf.