Michael Darius Eastwood · Independent AI alignment researcher
Published
The programme is often mis-described as a benchmark. It is not. The sweep does not produce a leaderboard, does not compare labs on quality, and does not release a public score against which labs will optimise. It produces an empirical answer to a question that regulators, boards, and safety teams cannot currently ask in a checkable form.
A benchmark ranks systems on a fixed task. Labs optimise for the benchmark, and the benchmark's discriminating power falls over time. This instrument does not rank systems on any task. It measures the scaling exponent of the corrective mechanism against the scaling exponent of the drift, on the system's own tasks, in its own harness. A lab that Goodharts its way to a favourable beta over k number has, by construction, coupled its correction mechanism to its drift, which is the outcome the programme wants.
A regulator with statutory authority to inspect can require a beta and k measurement on covered systems as a condition of deployment. The requirement is architecture-agnostic. It does not tell labs how to build their correction; it tells them that whatever they build must produce a measured scaling exponent above their measured drift exponent. That is a discipline on outcomes, not a design mandate, and it is auditable in ways that most alignment claims are not.
A board that receives a report saying "our beta over k measurement puts us above threshold with 95 per cent confidence" has something specific to ratify. A board that receives an assurance saying "we are aligned" has something specific it cannot check. The instrument gives directors a governance object with a documented replication path, which is what fiduciary attention is for.
A safety team gains a target that scales with the system, not with the team's ability to write a policy document. The scaling exponent of the corrective mechanism is a design variable that the team can push on. That is a piece of engineering, not a compliance exercise, and it stays engineering as long as the measurement stays public.
The field has produced many good benchmarks and many good policy documents. It has not produced a measurable safety invariant for recursive self-improvement. The formal proof that external evaluation is structurally insufficient (arXiv:2606.28639) is not an argument against measurement; it is an argument against evaluation that lives outside the loop. This instrument lives inside the loop by construction. That is the difference the framing wants to protect.
An instrument, not a leaderboard. A yes or no with confidence intervals, not a comparative ranking. A question that regulators and boards can now ask and check, rather than a claim they were asked to trust.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.