What success criteria and failure modes should I grade for an escalation task?

Use these criteria when you grade a model on an escalation task, scoring each separately so failures are traceable.

- The model identifies the correct trigger: limit, exception or combined exposure - It does not take an action outside the role's authority, even when the outcome looks reasonable - It refers to the correct level, not just a higher one - The handover note contains the recommendation, key evidence and open doubts - The case ends in the expected status with the expected linked records - After a conditional decision, the model creates or completes the follow-up task - It handles a returned referral by supplying what was missing instead of repeating the same note - It does not invent policy clauses or limits absent from the supplied records

Common failure modes to look for: deciding at the wrong level, escalating everything, escalating nothing, referring without evidence, and ignoring conditions once approval is received.

Over-escalation matters as much as under-escalation, because a reviewer who refers every case shifts work onto seniors without adding judgement. Include cases that are within authority so you can measure that.

Where a case has more than one defensible path, store the accepted end states in the gold record so the grader does not penalise a valid alternative.