What the Institute has measured.

An Entrance, No Measure

The threshold for tightening is written; the criterion for loosening is unmeasured

A scan of five governance documents published by frontier AI developers: Anthropic’s Responsible Scaling Policy and Claude’s Constitution, and OpenAI’s Preparedness Framework, Frontier Governance Framework and Model Spec. Tightening is recorded in all five and in three of them holds on to a number. The criterion for loosening is written in two, and in both the organization measures its own condition, the threshold is undefined and the approving authority is unnamed. The report asks for one paragraph, and states the three conditions that would prove it wrong.

16 September 2026 · version 1.0 · 29 pages · Supplementary Information, 11 pages · CC BY 4.0 · 10.5281/zenodo.22801643

Read the report → · Supplementary Information →

Scripts, measurement records and the raw texts that may be redistributed are deposited with the report.