A hard-to-eval product is hard for users to verify

When a team says “our product is hard to eval,” that is usually a product smell: an artifact that is hard for the builders to verify is hard for users too, and in the worst case users redo the work from scratch to check it. The lever is to design for verification before building evals, not after.

The fastest way to make an output checkable is provenance, showing where each part came from with a link back, paired with progressive disclosure so the sources do not overwhelm (see: agent actions and planning must be observable). Break the output into smaller units the user can accept, edit, or reject, a diff against a trusted anchor or a finding with a citation, which shrinks the surface area evals must grade and lowers the cognitive load of judging a whole artifact at once. Products already doing this, a data agent showing the query and the metric definition behind a number, or a coding agent recording a video of its UI changes, are easier to trust and easier to evaluate (see: eval systems create a flywheel for iteration).

Before AI, verification happened incidentally while creating the work; with AI, verification is the bottleneck, so it has to be designed explicitly (see: evals define what success means for an agent). The same provenance that makes an output checkable for a user is what makes knowledge durable for an agent (see: OKF gives every agent durable company knowledge from a folder of typed markdown).