track1-agent:latest

ZeroRoute treats the leaderboard formula literally: only tokens through FIREWORKS_BASE_URL count, so the cheapest model is the one running inside the container. The agent contains no Fireworks client at all — zero tokens is not a best case, it is guaranteed by construction. That makes accuracy the whole game, and the design spends the entire 4 GB / 2 vCPU / 10-minute budget on it. A zero-cost regex classifier sorts tasks into eight categories, with a local-LLM rescue pass for reworded prompts that dodge keyword rules. Tasks run in two phases so each model loads exactly once. Phase 1 (Qwen2.5-3B-Instruct): factual Q&A with dual-draft merging, sentiment, summarization with hard enforcement of sentence and word limits, and three-pass NER including a product-focused sweep. Phase 2 (Qwen2.5-Coder-3B): math solved by executed program-of-thought where two independently generated Python programs must agree on the number; logic puzzles brute-forced by generated enumeration code under an exactly-one-solution contract, cross-checked against elimination-style reasoning; code debugging and generation syntax-checked with ast.parse and retried on failure. Engineering for the harness: /output/results.json is valid JSON from second zero and atomically rewritten after every task; a pace guard degrades to fast single-pass answers rather than timing out; peak memory stays under 2.6 GB.