HCD.18:4.6 - Try the specification on contrasts and establish its continuation

Have an evaluator use the same specification on its adequate, deficient, and outside-use contrasts. Examine the resulting decisions and repairs, not merely agreement on labels. Check whether a known central defect can pass because an example is ambiguous, whether a correct alternative is rejected, and whether a missing observation is mistakenly converted to a low material score.

Where an AI reader or evaluator is used, test the contribution it is actually able to make. Subject-grounded counterexamples can calibrate defect diagnosis; simulated novice behaviour cannot establish human learning. A fresh reader is useful when private knowledge of the intended defect or answer would defeat the selected inference.

Stop construction when another evaluator can identify the object and tasks, collect the specified evidence, return all selected coordinates with their limits, and decide the declared next move. An unresolved criterion or discriminating-evidence gap is a legitimate returned result.

Name the specification version when results will be compared later. A changed value meaning, task denominator, or support arrangement may invalidate numeric comparison with earlier results. Preserve earlier results under their original meanings; establish a justified comparison of meanings or decline equivalence. Reopen only the affected specification content when a new task, audience, source premise, assistance change, or recurrent calibration disagreement changes it.