I wanted my model routing to become more predictable. By October 30, there were enough choices in the setup that reading a preference list wasn’t enough to understand where a task would go or what had happened when it came back.
One of my requests was to make delegated work visible in T3 threads. I wanted a history I could inspect when reviewing a result. If a coordinator handed work to another model, the delegated attempt needed to be part of the record rather than something I had to reconstruct from a final summary.
The routing review found problems that were more basic than which model should rank first. Unavailable routes could consume attempts, and model aliases needed consistent handling. If the same model appears under different names, even a simple usage summary can become misleading. If an attempt goes to a route that can’t answer, changing the preferred model doesn’t address the failure.
The reported fixes addressed unavailable-route handling and alias normalization. Route and syntax checks passed, but complete runtime verification was still open. No models were promoted or demoted in that review. That restraint made sense: the evidence concerned the routing machinery, so it didn’t justify a new model leaderboard.
There was related environment work that day after provider roots changed. I approved tiny live prompts to check the routes from the setup where I would actually use them. The thread reported successful replies from Claude, Codex and Grok. Gemini’s quota prevented a complete equivalent retest.
Those prompts were availability checks. They showed that a route could respond to a small request at that moment. They didn’t establish coding quality or how it would behave during a longer delegated task. The missing Gemini check also needed to stay missing, rather than being counted as a pass or interpreted as a quality failure.
I was trying to make future decisions easier to explain. Before comparing models on finished work, I needed to know which route was attempted and whether it ran. Visible delegation history and small live checks gave me evidence for those narrower questions while leaving the larger performance questions open.