TokenRouter the prove-or-escalate agent

TokenRouter is built for one thing: winning Track 1's ranking, the fewest Fireworks tokens once you're past the accuracy gate. Every scored token has to be earned, so the agent won't spend one until it's proven it can't get the answer for free. Each task drops down a three-tier ladder. Tier 0 is plain Go at zero tokens: an arithmetic evaluator, brute-force solvers for logic puzzles (knights-and-knaves, zebra grids, positional puzzles), mutation-based program repair, and a library of solutions already proven out. Each solver only answers what it can verify. That's the whole trick, and it's why a wrong free answer simply can't happen here. Tier 1 bundles a Gemma 4 E2B in the container, still at zero scored tokens, and nothing it says gets trusted on faith. Generated code runs against tests pulled from the prompt, math gets recomputed in Go, formats get checked. Only a miss that survives all that reaches Tier 2, the Fireworks API, which runs the same Gemma remotely with terse prompts, thinking disabled, and prefix caching. On our own evals, that lands about two-thirds of tasks at zero tokens and cuts Fireworks calls by roughly 60%. Every choice was A/B-tested and logged. We even tried a stronger local model and dropped it once the numbers said to.