Hit a hard p99 SLA on an LLM product without blowing a fixed monthly spend ceiling. Model it.

This screens for whether someone can hold two hard constraints at once instead of optimizing one and apologizing for the other. The trap is batching: it is the biggest cost lever and the biggest p99 risk, and a candidate who reaches for aggressive continuous batching to cut cost without naming the tail-latency hit has not run a serving fleet. The discriminating follow-up is 'traffic doubles mid-month and you are on track to blow the ceiling, what gives?' The answer that scores names the explicit degradation policy decided in advance (route more traffic to the cheap model, shrink max output tokens, shed or queue low-priority work) rather than silently breaching the SLA or the budget. Anyone who says 'autoscale' without a unit-cost number is guessing.