Hybrid Routing Agent — Local First, Tokens Last

Track 1 scores accuracy first, then ranks survivors by Fireworks tokens spent — and local tokens count as zero. So the winning agent isn't the one with the best model; it's the one that most reliably knows when its free answer is already correct. Every task climbs a cost-ordered escalation ladder and exits at the first trustworthy rung. Rung 0 classifies the task with zero-cost heuristics, plus an exact Python solver for explicit arithmetic (free AND certain). Rungs 1-3 are local Gemma attempts gated by a strict free verifier (numbers must parse, MCQ letters must be among the options, Python must compile, summaries must honor stated length limits) and honest logprob confidence, then self-consistency voting across independent samples. Rung 4 is the cheapest paid call: a remote judge grades the local winner with a single output token — verification costs 10-50x less than generation. Only then come real paid generations: a cheap model with a compressed prompt and the local draft as a hint, then a strong model as last resort, under a hard per-task token budget. Weak-verifier categories (math word problems, code debugging, logical deduction) never ship on local confidence alone — the 1-token judge must approve, and a judge-rejected answer can never win a self-consistency vote. That discipline came from dry-running the judged contract end to end: all eight capability categories answered correctly at a fraction of all-remote token cost. It ships as a single public linux/amd64 Docker image implementing the harness contract — tasks.json in, results.json out, atomic writes after every task, 10-minute budget management — reading FIREWORKS_BASE_URL and ALLOWED_MODELS from the environment at runtime. Plus a live dashboard that visualizes every task climbing the ladder with a token meter, 161 offline tests, and graceful degradation when the model file or API key is absent.