A municipal ice-rink booking pack botches a three-rule fee ladder in one shot. The tester splits the work into “list rules → apply one booking → check the total” and reviews each reply before the next. Which mitigation is named?

Municipal ice-rink fee ladders often stack multiple rules in one booking. A single-shot pack that botches the three-rule ladder shows classic reasoning-error risk on multi-step work. The mitigation named in the stem is dividing the task into manageable segments with verification between steps—list rules, apply one booking, check the total—reviewing each reply before feeding the next. Early human or structured checks stop a faulty intermediate from poisoning later steps.

Raising temperature fails by concept: wider sampling increases variety; it does not fix multi-rule reasoning and often worsens fee inconsistency. Omitting context so the model invents fewer rules fails because removing basis material increases invention risk rather than reducing it. Treating the LLM as the system under test with neuron coverage fails because that framing belongs to CT-AI-style model evaluation, not Chapter 3 prompt-task mitigation for generating municipal testware.

Exam caveat: segmentation is not the same as lowering temperature or adding more context. Those are sibling mitigations; the stem specifically demonstrates divide-and-verify on a multi-rule fee task.

Operational check: break fee, waiting-period, and eligibility ladders into ordered prompt stages; require a pass/fail check against the posted schedule after each stage; refuse to chain the next stage until the intermediate is accepted; archive stage outputs for audit when rink fees change mid-season.