Model evaluations
Evaluation as a governance gate, not just a benchmark.
Evaluation becomes governance when results drive release decisions, restrictions, escalation and control requirements.
1. ScopeSystem, deployment context and material capabilities.
2. EvaluateCapability, safety, misuse and robustness testing.
3. CompareAssess against predefined thresholds and tolerances.
4. DecideRelease, restrict, escalate, remediate or stop.
5. EvidenceArchive results, rationale, approvals and exceptions.
Evaluation domains
What to evaluate
- Capability uplift and frontier capability indicators.
- Autonomy and ability to execute multi-step plans.
- Tool use, permissions and external action capacity.
- Misuse potential and circumvention resistance.
- Robustness under distribution shift and adversarial pressure.
- Monitoring and intervention effectiveness.
Decision outputs
What evaluation should change
- Deployment scope and eligible users.
- Rate limits, permissions and tool connectivity.
- Required human review and escalation level.
- Monitoring intensity and logging retention.
- Need for independent assurance or executive approval.
- Whether deployment should be delayed or blocked.
Evaluation without a decision rule is only measurement.
Governance requires thresholds, owners and consequences.
Capability governance →