model-pro(vider)

One endpoint and one model list in front of every backend. A request for an unloaded model boots the right engine, loads it, and evicts whatever holds the GPU; cloud models use no VRAM and never evict. It never silently substitutes a model: typed errors, not surprise models. Context tiers pick the smallest configuration that covers a request instead of loading an oversized one, and out-of-band reconciliation adopts whatever is actually serving when a backend restarts behind the proxy's back.

It is the sole holder of provider keys, so applications name models and never see a secret, and every caller must identify itself. Virtual names resolve down an ordered preference list with fallback. An evaluation harness replays real sessions as independent rollouts per candidate model and flags tool descriptions where the models agree with each other but not with the person, because that is usually the tool's fault.