A published commitment that says, in advance, which capability levels trigger which safeguards, and that development stops if the safeguards are not ready. It converts safety from a matter of judgement in the moment into a pre-committed rule that can be pointed at later.
Technical view
Defines capability tiers with associated evaluation triggers and required security and deployment measures at each, plus commitments to pause if measures cannot be implemented. Similar frameworks exist across several developers under different names and informed some regulatory drafting. Weaknesses are that they are voluntary, self-assessed, and revisable by the organisation that wrote them, so external verification remains the open question.
Related
Part of AI Terminology Index, section: Security and Privacy.