Routing
`llama3.1:8b` proved good enough for current ShowcaseAgent routing work.
Current result: smaller local models are already enough for meaningful routing and some grounded use, while exactness-sensitive downstream handoff remains a sharper boundary where stronger models or stronger policies still matter.
The Role Question
Project Phoenix does not treat local-model evaluation as a single ranking problem. Routing, grounded domain use, machine-facing protocol work, and repair-assisted pipelines are different operational roles. Models that look weaker in one role can still be the right answer in another.
Three Practical Results
Grounded Use
Smaller models crossed into grounded TourAgent usefulness once the harness carried the answer path.
Protocol Pressure
`gemma3:27b` remained materially stronger when the answer had to survive stricter downstream machine-facing requirements.
Why It Matters
This paper is the role-boundary piece of the local-model details layer. It explains why raw size is not enough as an evaluation language and why repair policy and handoff strictness change the answer before the model leaderboard does.