AI CompendiumA public chart of the language
Home / Safety & Alignment / Responsible Scaling Policy
Safety & Alignment

Responsible Scaling Policy

also called RSP, Frontier safety framework

Capability thresholds that trigger heavier safeguards before release.

Definition

A published commitment that defines capability levels, the evaluations that determine whether a model reaches them, and the safeguards required at each level — so the decision to deploy with weaker controls has to be made explicitly rather than by default.

The structure is a safety-case style argument applied at organisational scale, and the different labs have converged on recognisably similar shapes. For anyone adopting frontier models the practical parts are the deployment standards attached to a tier and the reporting obligations.

For consumers of frontier models the actionable content is the tier structure: it tells you which evaluations precede a release and what standards the developer committed to at each level. It is a way to make a deployment decision legible rather than a marketing document.

How people use it
A regulated deployment.
"Check the policy for the tier we are in. It tells you what evidence you are expected to have and what reporting applies."
Worked example

Using a policy as a checklist

Assessing a model for a high-consequence internal use.

Map your use to the capability tier the policy describes, then check what evidence and mitigations the developer committed to at that tier. It converts a vague question about trustworthiness into a specific list of things that should exist.

Sources
Related terms