Designing evaluation beyond public leaderboards

Revision note, October 16, 2026. This essay was substantially revised to add primary references, narrow unsupported claims and distinguish proposals from measured results. The original address and publication date are retained.

An evaluation program needs a way to change as models and tasks change. The useful design question is how to combine reusable comparisons, held-out work and targeted investigation within an explicit budget.

Different tests answer different questions

A new benchmark is not necessarily a harder edition of an old one. GPQA asks expert questions in biology, chemistry and physics. Its task distribution matters when interpreting its results; it should not be treated as a direct successor to a grade-school mathematics test. GPQA.

Likewise, publication of a benchmark does not mean every evaluation item is publicly available. Humanity's Last Exam described a public question set alongside a private holdout. An evaluation program should record which material is inspectable, which material is held back and how repeated use might affect validity. Humanity's Last Exam, section 3.

There is no defensible universal date on which public evaluation expires. A claim that a test has stopped separating candidates needs a defined model cohort, versioned tasks, comparable evaluation settings and evidence about score uncertainty. Close scores can indicate a limited measurement range, but they do not by themselves establish that all remaining differences are noise.

Give each layer a job

Use a public benchmark layer for comparisons that others can inspect. Preserve the exact protocol and avoid turning a broad score into a claim about an unrelated workflow.

Use a held-out task layer to test the intended application. Separate examples used to tune the system from examples used to judge it. Include difficult boundaries and cases in which the correct outcome is to abstain, request clarification or stop an action.

Use a targeted investigation layer for consequential uncertainties. This may require specialist review, adversarial attempts or inspection of a long sequence of tool actions. Record the access, time and expertise available to the evaluator. A successful investigation establishes evidence about the tested conditions, rather than a guarantee about every possible use.

Budget from the workload

Estimate evaluation cost from task count, repetitions, model and tool consumption, specialist hours and adjudication time. A constructed planning estimate should be labeled as such and kept separate from a measured invoice or sector-wide forecast.

Allocate effort according to the decision. A low-consequence formatting task and an agent that can alter important records need different evidence. Keep a reserve for investigating surprising failures instead of spending the entire budget on a fixed average score.

External evaluation already exists

Independent technical evaluation is not confined to model developers. The UK AI Security Institute described testing before and after deployment in its December 2025 frontier AI factsheet. That establishes existing external work, not universal access, unlimited capacity or authority to approve every release. AISI factsheet.

When proposing a larger evaluation institution, specify what is missing: a particular kind of model access, an evaluation method, staff capacity, legal authority or enforcement. These are different constraints and require different remedies.

Define the update rule

Record which changes trigger reevaluation and which failures trigger a deployment decision. Keep protocols and results versioned so that a system improvement can be distinguished from a changed test. This turns evaluation into an inspectable operating process, with public benchmarks as one useful component.