AMD Router: Local-First Token-Efficient Routing

AMD Router treats remote tokens as the scarce resource they are. Track 1 imposes an 80% accuracy gate — miss 16 of 19 tasks and you don't appear on the leaderboard at all, with token count only breaking ties among teams that already pass. So the objective isn't "minimize remote calls," it's "clear the gate, then minimize remote calls." Every task first goes to a local Qwen2.5-0.5B model running CPU-only inside the container — free tokens. A heuristic classifier assigns a task type and difficulty by regex, with no model call. The local answer's confidence is measured from token logprobs and compared against a per-task-type threshold: math demands 0.80, classification 0.55, because a wrong number is worse than a clumsy label. Only when local confidence falls short does the router escalate to a Fireworks model, and it selects the cheapest capable model from ALLOWED_MODELS at runtime. On a 19-task evaluation mirroring the harness categories, the hybrid scores identically to always-remote (15/15 auto-graded) while spending 3009 remote tokens versus 4526 — a 34% reduction. Six of nineteen tasks never make a network call. The design is measurement-driven. The original policy force-escalated all code tasks as "locally unreliable"; measurement showed the 0.5B model passes them at 0.90 confidence, and removing that assumption saved 691 tokens at zero accuracy cost. The router also survives real-world failure: three of the five allowed Gemma models are undeployed and return 404, so it marks them dead and walks the list rather than collapsing to the fallback model. Reasoning models that return only a reasoning trace, and small models that recite training-data system prompts, are both handled explicitly. The router is the product; base models are interchangeable. Public image: ghcr.io/makabeez/amd-router:v0.12