Evals-Driven Policy

Module two ended with a measurement problem. Evals-driven policy turns it into a governance mechanism: define capability thresholds in advance — 'if a model crosses X on bio-uplift or cyber autonomy evals' — and pre-commit to the safeguards that trigger, before the model exists.

The flagship example is Anthropic's Responsible Scaling Policy, which maps models to AI Safety Levels borrowed from biosafety — ASL-2 covers today's models, ASL-3 and beyond demand stronger containment and deployment controls proportional to measured capability. Other frontier labs have published parallel 'preparedness' or 'frontier safety' frameworks on the same logic: measure against the threshold, then apply the pre-committed controls.