Frontier AI accountability keeps failing the same way: commitments are made by the party that cannot be checked, in language that cannot be verified. The Compact reassigns six commitments to the parties that can actually adopt them now. The labs: published capability thresholds, independent evaluation with real access, and monitoring honesty. Buyers, through procurement contracts: incident disclosure with teeth, and liability that lands somewhere. Everyone deploying agents: attributable agents, meaning verifiable identity for every agent acting in the world.
The Compact is designed to be tracked, not admired. Commitment one is monitored live at Threshold Watch, which snapshots and diffs every frontier lab’s published safety framework; the Compact Clauses turn the buyer-side commitments into contract riders; and the Safety Framework Publication Format specifies what a verifiable framework looks like. Adoption is reported publicly, and as of September 2026 the honest count is zero. I report the silences as plainly as the progress.