05.1
Activity is not evidence
Long traces can look productive while producing the wrong file. Short traces can use expert shortcuts and produce excellent work.
Evaluation should not reward motion for its own sake.
05.2
Inspect the deliverable
The strongest verifier reads the thing the task asked for: the exported video, the spreadsheet formulas, the component assembly, or the application state.
It can then combine structural checks with domain-specific quality rules.
05.3
Make checks programmable
Teams need a small interface for composing deterministic checks, tolerances, and partial credit. Those checks should run beside the environment that produced the result.
Programmability lets each occupation define success in its own terms.
05.4
Return useful feedback
A verifier should explain which condition failed and preserve the artifact that triggered it. A bare score is difficult to debug and hard to learn from.
Specific feedback closes the loop between practice and the next attempt.