Minimal reciprocal disclosure in shared semiconductor screening depends on memory pressure, billing, timing masking and division of work

A group screening two-dimensional semiconductors by density functional theory and the computing centre it relies on could each gain from disclosure, yet each guards its knowledge (candidate list; settings). Federated learning and encrypted computation protect a shared model but do not ask how little each party must disclose. The hypothesis was that agreement requires jobs filling much of node memory, billing for reserved memory, masked completion times, and paired grading, and that under these conditions one bit per job flagging high memory demand lets the centre gain. Four coordination patterns and five disclosure levels were evaluated in 1,000 simulated scenarios and four preregistered measured comparisons, Quantum ESPRESSO screens of 29 transition-metal dichalcogenides for a 1.0–2.0 eV Perdew–Burke–Ernzerhof gap. Lacking any condition, no agreement, even full disclosure, benefited both parties; run-time brokering, not a shared logbook, provided the value. With all four, the bit raised the centre’s valid results per unit charge 1.13-fold, as predicted, but the group reached a qualifying compound more cheaply in only 5 of 15 paired campaigns, and completion times still revealed each compound. On shared nodes the gain fell to 1.07-fold, with an interval including no gain. The conditions are necessary; sufficiency is not yet shown. Among research groups sharing gap and vacancy-barrier results (simulated on measured values), exchange benefited no group; a neutral broker dividing the work, with decoy refusals, benefited every group with lists hidden in a quarter of scenarios. Before disclosing, a group should check memory pressure, reserved-memory billing and completion-time visibility.