eval (graded eval)
definitionA human-graded check of output quality—correctness, clarity, judgment—where a pass/fail machine won't do.
A validation modality (c): a human-graded pass on output quality—correctness, clarity, judgment—that needs discernment rather than a pass/fail robot.
Produces a graded-eval record (a type: eval-handoff + type: eval-results pair). What
earlier drafts mislabeled a "smoke test."
The Validation Layer. The verification stack answers one question: did we build it right?—does the code match the spec. Every layer in Section 2 serves that question. It is only half of quality.
The other half is validation: did we build the right thing, and is it good? A system can pass every test, clear every linter, and survive cross-agent and human architectural review—and still be the wrong thing, or a right thing that is confusing, unpleasant, or useless in real hands. Verification cannot catch this, because verification measures conformance to a spec, and the spec itself is what validation interrogates.
This is Section 1's thesis extended. The research there—that humans overestimate AI-assisted work—was about correctness. It holds at least as strongly for goodness. An agent reporting "the feature works" is reporting that it ran; it is not reporting that the feature is good, because good is not a property the agent is positioned to judge. Validation is the antidote to overconfidence about goodness, exactly as verification is the antidote to overconfidence about correctness.