Gemini Cartographer

Finding

Gemini routing is operationally weak when it is treated as a model experiment instead of a mapped lane for long-context, multimodal, and research-heavy Hermes work.

Current

A real Hermes installation may try Gemini-family models for large context windows, document-heavy research, image or screenshot interpretation, and broad synthesis tasks. The weak point is usually not access to the model; it is the absence of a clear map showing when Gemini should be evaluated separately from the primary model, what task types it should compete on, and how the operator records whether it actually improved the outcome. Without that map, model selection becomes anecdotal and Hermes cannot build a reliable terrain view across providers.

Suggested

Define a Gemini evaluation lane for the task types where it is most likely to matter. Exact change: add a “Gemini lane” section to the model-routing runbook or SOUL.md stating that Gemini should be separately evaluated for long-context research, multimodal interpretation, large document comparison, broad source synthesis, and tasks where context capacity may matter more than raw coding strength.
Add a model-comparison note to research and multimodal workflows. Exact change: patch the relevant research skill, browser/screenshot workflow, or task completion runbook with: “When Gemini is used, record the task type, why Gemini was selected, whether the result was better, worse, or equivalent to the default lane, and what evidence supports that judgment.”
Create a lightweight Gemini terrain review in the Optimizer Agent cadence. Exact change: update the Optimizer Agent review prompt or dashboard copy with a monthly check: “Review recent Gemini-routed tasks and summarize which workflow categories are proven, uncertain, or not worth routing to Gemini; recommend keeping, narrowing, or removing the lane.”

Impact

This turns Gemini from a novelty provider into a documented capability map. Hermes gains better model-routing discipline because long-context, multimodal, and research-heavy tasks can be evaluated on evidence instead of preference. Over time, the installation builds a practical model map: where Gemini expands the operating surface, where the primary model remains stronger, and where specialist routing is not worth the extra complexity.

Effort

Small — the work is a routing rule, one comparison habit, and a recurring review item. No new infrastructure is required if a Gemini-capable provider lane already exists; the main discipline is recording outcomes consistently enough to make routing decisions reliable.

Public page note

Safe public content includes the maturity principle, generic Gemini-suitable workflow categories, routing criteria, comparison habits, and the value of model terrain mapping. Internal-only content includes real provider keys, private prompts, raw research data, user files, screenshots with sensitive content, exact model costs from private accounts, raw evaluation logs, and any installation-specific configuration values.