Token-Miser Router Agent v9

Token-Miser Router Agent is our Track 1 entry: a containerized agent that answers all eight task categories of the benchmark — math, logic, sentiment, NER, summarization, factual QA, code generation and code debugging — while treating every Fireworks token as a cost to be engineered away. Track 1 scores in two stages: an LLM judge gates on accuracy, then survivors are ranked by tokens spent through the Fireworks API. That flips the usual objective — the winning move isn't a smarter model, it's not calling the API at all. Every task walks down a ladder of tiers. A regex classifier tags the task for free. Provable math and logic are parsed into a syntax tree and solved deterministically in pure Python — a path that is structurally incapable of hallucinating. Fuzzy tasks (sentiment, entities, summaries) run on a local Gemma model baked into the image; local inference never touches the scored API. Only when a cheaper tier cannot produce a validated answer does the task escalate to Fireworks — and even there we squeeze: hidden reasoning disabled per model, price-ranked routing to the cheapest capable model, terse prompts, and hard per-category output caps (a sentiment answer is capped at four tokens). The core invariant: cheap tiers may save tokens, but they never gamble the accuracy gate — every local answer is validated before it ships, and validation failures escalate. Measured worst case (no local model at all): 9/9 correct on the sample set, 627 total tokens, about five seconds end to end. Ships as a public linux/amd64 image on GHCR with a public GitHub repo and a README that documents setup and usage; nothing is hardcoded — the harness injects credentials at runtime. Categories / Track / Technologies — you've already got Virtual Assistant, the "Hybrid Token-Efficient Routing Agent" track, Gemma, and AMD Developer Cloud selected. If the Technologies dropdown offers them, also add Fireworks AI, Docker, and Python — more accurate tags, and judges filter by them.