Assessment a model cannot sit

Any task that a model can complete end to end, unsupervised, is a task that cannot tell you anything about the candidate. That is a stricter constraint on assessment design than it sounds — it rules out most timed, closed-book, single-output formats currently in use.

The tasks that survive tend to share a shape: they require a decision under conditions the model was not given, evidence of reasoning the model cannot fabricate convincingly on its own, or accountability for a real consequence the candidate has to own afterwards.

Designing for that shape is slower than writing a mark scheme for a finished output. It is also the only way an assessment stays meaningful once the tools involved keep getting better.