A fire-district after-action report assistant finds an oversized foundation model that exceeds the city’s per-token budget. What should the team do?

Rejecting the oversized model that violates the city’s per-token budget and selecting a cost-fitting alternative is the correct response for a fire-district after-action report assistant. Municipal model selection is a constrained optimization across quality, latency, and appropriations: an FM that exceeds the per-token budget is out of policy even if marketing claims are attractive. Teams should choose a smaller or differently priced model that still meets after-action summarization quality bars within the authorized spend. Rejecting the oversized model that violates the per-token budget and selecting a provisioned throughput deployment fails because provisioned throughput changes hosting capacity, not the per-token rate; it does not guarantee a cost-fitting alternative and can add fixed capacity cost while remaining out of policy for civic GenAIOps. Rejecting the oversized model that violates the per-token budget and selecting a fine-tuned version fails because fine-tuning improves accuracy but does not lower the per-token rate; the fine-tuned variant remains over budget and adds training cost without providing an evaluated cost-fitting alternative. Rejecting the oversized model that violates the per-token budget and selecting a long-context variant fails because long-context variants typically carry equal or higher per-token rates; it does not solve the budget exceedance and the team has not evaluated a cheaper cost-fitting alternative. Exam caveat: cost ceilings are legitimate rejection criteria on exam scenarios. When a stem states a per-token budget is exceeded, the correct action is reject-and-select-alternative, not “quality at any price.” Operational check: estimate tokens per after-action report times expected volume, compare candidate models to the city rate card, reject those above budget, and document the selected cost-fitting model with eval evidence.