Thresholds and checkpoints
Capability threshold, threshold status, alert threshold, checkpoint rule, halt condition, reassessment trigger.
The capability threshold is the operational core of every responsible scaling policy: the specific, testable point at which a model is judged to have crossed into a risk tier requiring new safeguards. This track defines capability threshold, alert threshold, halt condition, and checkpoint rule as used across Anthropic's Responsible Scaling Policy, OpenAI's Preparedness Framework, Google DeepMind's Frontier Safety Framework, METR's evaluation work, and the US government's Executive Order 14409.
- Reassessment TriggerProposedrule
A proposed reassessment trigger names events, like a more capable model or a serious incident, that require a new assessment of an already-assessed model.
- Halt ConditionProposedrule
A proposed halt condition is a declared condition requiring development, training, or deployment to stop, naming who may invoke and lift it.
- Checkpoint RuleProposedrule
A proposed checkpoint rule is an if-then rule requiring certified properties for a capability, naming the evidence, certifier, and consequence of failure.
- Alert ThresholdProposedrecord-type
A proposed alert threshold is an indicator set just below a capability threshold, triggering closer assessment without the threshold's own consequences.
- Threshold StatusProposedcontrolled-value
A proposed threshold status records a model's position relative to a capability threshold, plus the date, evidentiary basis, and determining party.
- Capability ThresholdProposedrecord-type
A proposed capability threshold records a model capability level requiring specified additional safeguards or decisions, along with its disclosure status.
