The mayor’s office wants multimodal models for press-photo captioning but text models for legislative Q&A. Which planning principle should guide that split?

The mayor’s office should match modality to the task: multimodal models for press-photo captioning and text models for legislative text Q&A. Foundry planning treats modality as a first-class selection criterion. Photos need vision-capable models; ordinance and bill text are language tasks that do not require image understanding. Splitting deployments by task avoids matching input format to the task for photo captioning and legislative text Q&A, which misses the deciding factor because generic format wording does not specify vision capability for press photos. Matching input format to the task for photo captioning and legislative text Q&A fails as a planning principle—describing inputs alone does not explain model selection, since only a vision-capable multimodal model can interpret images, while legislative text Q&A does not require image understanding. Matching data type to the task for photo captioning and legislative text Q&A fails as a planning principle: generic data-type wording does not give a text model vision for press photos and does not capture the modality capability each workload requires. Matching content type to the task for photo captioning and legislative text Q&A fails as a planning principle: content wording alone does not determine vision capability, so it cannot explain why press photos need multimodal models and legislative text needs text models. Exam caveat: shared platforms can still host multiple model deployments—modality matching does not forbid centralized Foundry governance. Operational check: route press-photo jobs to the multimodal endpoint and legislative Q&A to the text endpoint, compare quality and cost, and prevent accidental cross-routing in app configuration.