Preserve information-flow labels through model distillation

Prove that information-flow labels governing tier-restricted scientific knowledge remain intact through model distillation, thereby enabling a valid tiered-release certificate after distillation.

Background

The paper argues that tiered release of scientific foundation models requires information-flow control over datasets, adapters, checkpoints, outputs, and tool boundaries. It notes that non-interference can be imposed across domains in modular models, but that distillation may alter how restricted information is represented and propagated. Establishing label preservation through distillation is identified as the unresolved part of the proposed release workflow.

References

Modular-model IFC shows non-interference can be imposed across domains; keeping labels intact through distillation is the open part.

— Privacy Foundations for Multi-Institutional Scientific Artificial Intelligence  (2609.39787 - Kotevska et al., 30 Sep 2026) in Section 8, subsection “Adaptation, evaluation, and release”