Describe the feature

Right now, the router config uses model endpoint, and assigns signals, without a transparent and consistent way of evaluating the model or routing policy.

Why do you need this feature?

We need an model onboard procedure that does the following:

• Evaluate the model through curated benchmark to infer the latency, accuracy, reasoning, and token consumption.
• Update the config if necessary especially if the model's eval results are not covered by the (e.g. accuracy, latency) thresholds used by signals
• Update routing models, e.g. the embedding based clustering.
• Update the config and continuously monitor the new model's performance through online learning feature: enhance online learning with bidirectional signal extraction #1140

Additional context

No response