Why the vision route uses a dedicated vision model.

ChironAI routes image-bearing requests through a dedicated vision model as the consultant in the vision slot. That is a specific architectural choice with specific consequences, and it is worth saying explicitly why we made it.

The shape of the choice

The Fusion v5 architecture has three frontier consultant slots plus a parallel vision route. The vision route is the path image-bearing requests take — chest X-rays, MRIs, CT studies, pathology slides, the document-bearing clinical inputs that arrive in radiology and pathology workflows.

Picking which model occupies the vision slot is an architecture decision, not a product decision. The classifier decides which slot a request needs. The reasoner orchestrates. The vision consultant returns structured findings. The choice of which model fills that slot affects two things: open-weight family alignment with the long-context consultant, and provider-diversity posture across the architecture.

Open-weight family alignment

The vision model shares family lineage with the long-context model that occupies our longitudinal-context consultant slot. The two models share alignment, tokenisation, and lineage. When the reasoner delegates a long-context read of a clinical chart to the long-context consultant and a parallel visual read of the imaging study to the vision model, the two consultant outputs reconcile more cleanly than they would across two unrelated provider families.

That family alignment is a small but real engineering win. It reduces edge cases at the synthesis layer where the reasoner pulls consultant outputs back together.

Provider diversity posture

Healthcare procurement asks specific questions about model provenance. Some health systems prefer open-weight models for sensitive routes; some have vendor-restriction rules; some prefer to keep the consultant pool diverse across providers to avoid single-provider dependency. A dedicated vision model in the vision slot keeps the architecture diverse across providers rather than concentrated.

That diversity is structural. It is not a feature we can claim about other vendors; it is a fact about how the Fusion v5 architecture is composed.

What this is not

We are not claiming any single vision model is universally superior. The frontier vision space is competitive and changing fast. We are claiming that the architecture lets us swap the vision-slot consultant when the right answer becomes a different model — and that the swap is a configuration change, not a rebuild. The composition is per-request; the slot is provider-agnostic by design.

When the right answer in 2027 becomes a different vision consultant, we will swap without rewriting the radiology workflow. The reasoner is unchanged. The modality-appropriate framework application (BI-RADS, LI-RADS, PI-RADS, others) is unchanged. Only the consultant changes.