Evals will conduct an interview at three in the morning with the same patience it had at nine. It will apply the fortieth rubric criterion with the same attention as the first. These are things machines are simply better at, and pretending otherwise is nostalgia.
A score is not a decision
A decision weighs things no transcript contains: what the team is missing, what the last hire taught you, the risk you are willing to take on someone unusual. The examiner does not know any of that, and we are not going to pretend it does.
The instrument measures. The person reading it decides. Confusing the two is how black boxes get built.
So the product draws a hard line
The AI scores against the rubric and files its reasoning, criterion by criterion, next to the transcript. A person reads why before trusting it, and can overrule any score with a note that becomes part of the record. The override is not a failure mode. It is the point.
watchglass (2026). What the examiner must never decide. The notebook, volume № 01, § 04. watchglass.app.