Cross-border data swaps face uncertain rules: LCNB turns compliance into a computable game and cuts violations below 1% in two stress cases
Synopsis
When machine-learning agents in large organizations exchange data across boundaries, which rules bind the transfer is itself uncertain, depending on where data subjects live, where processing happens, who owns the agent and which rule version is in force. LCNB models one exchange as a regime-uncertain transfer game with an enforcer, has agents hold a Bayesian belief over applicable rules, pay to unlock safeguards or verify claims, then bargain over the remaining lawful bundles with a Monotonic Concession Protocol. In three stress cases where residence and citizenship differ, a citizenship-based rule violates in 21–37% of shared records in two of them, and LCNB at or below 0.3%; under default economics it shares almost only aggregates, which partly explains the low rate.
Interpretation
Compliance for a cross-border data exchange splits into two steps: infer which rule actually binds this transfer, then negotiate a transfer form both sides accept inside the lawful set. Prior practice either applied the sender's host-region rule to every record or treated residence as equal to citizenship, and each reads only one feature. In three stress scenarios where residence and citizenship differ, the citizenship rule violates in 37% of shared records for non-Europeans living in the EU and in 21% for EU citizens abroad with no EU rule binding, while LCNB stays at or below 0.3% with diagnosis F1 of 0.93 or higher.
Rules enter as hard constraints, each giving a minimum protection level per dimension, and overlaps take the strictest level per dimension, so the lawful set is described by one minimal requirement vector per record. Agents see tags that may be missing or wrong and a counterparty that may overclaim its safeguards, so they first infer a posterior over true attributes and accept only bundles whose violation probability stays within a tolerance. Tests run in a synthetic simulator whose rule library encodes GDPR, DPDP, CCPA/CPRA and a stylized stricter regime, with 5 seeds and 60 ticks per scenario and 10 records per tick.
The bargaining step uses a Monotonic Concession Protocol to approach the Nash bargaining solution, each side opening with its best Pareto offer and the lower-risk side conceding the least amount that makes its opponent's risk smaller. The protocol matches the Nash point on a 400-point frontier test and takes 11–26 rounds per tick in the experiments. Pooling sender values of 0.2, 0.4 and 1.0, the Monotonic Concession Protocol gives the same welfare as the Nash solution, with Kalai–Smorodinsky slightly lower.
Perspective
The conclusions rest on synthetic records and stylized rules drawn from public texts but not legal advice, and the authors make no claim about legal compliance in practice; the experiments use one sender–receiver pair and runs of 5 seeds and 60 ticks.
The authors report that verification never fired in any configuration, so there is no evidence it helps; unlocking only gave a small welfare gain at higher tolerance; under default parameters LCNB shares almost only aggregates, making its zero violations weak evidence; at low audit rates a violating method can beat the oracle on welfare, which the authors report without treating as a case for violating.
