AI safety · compute governance
SplitComp: Stress-Testing Compute Governance Under Strategic Evasion and Jurisdictional Fragmentation
Modelling how a lab chooses between compliance, single-jurisdiction evasion, and splitting a training run across jurisdictions — and how audit probability, detection accuracy, penalty severity and cross-border coordination shape that decision.
Why it matters
Governance mechanisms fail when formal rules ignore the strategic incentives of the actors expected to comply.
Overview
This project studied how a lab weighs compliance, evasion within a single jurisdiction, and splitting a training run across several jurisdictions — and what combinations of audit probability, detection accuracy and penalty severity can sustain compliance.
One of the main findings is that stronger domestic enforcement does not necessarily produce compliance. When cross-border coordination is weak, it can instead make jurisdictional fragmentation more attractive.
What survived
A durable interest in strategic behaviour, institutions, comparator design and the conditions under which powerful systems remain governable.
Research pages preserve changed claims rather than silently rewriting them. Last revised .