ZeroFire — Zero-Token Local-First Routing Agent 02

Track 1 ranks agents by how few Fireworks tokens they spend. ZeroFire takes that to its limit: it spends ZERO. Every task is answered inside the container by a compact local model (Qwen2.5-1.5B-Instruct, 4-bit) wrapped in deterministic verification — so the cheapest API call is the one we never make. How it works: - A zero-cost regex router classifies each task into one of the eight capability categories. - Math is computed, not guessed: the model writes a tiny Python program, we execute it in a sandbox, and independent programs must agree. - Logic puzzles are solved in code: deterministic ordering, assignment, and truth-table solvers parse the constraints and brute-force the answer with regex and Python — often with zero model calls. - Code tasks are compiled AND executed, then differential-tested against an independently written reference implementation. - Sentiment labels are computed in code from extracted aspects, so a label can never contradict its justification. - Factual answers are cross-examined for agreement and completeness, with a small offline gazetteer for geography. - Format constraints (exactly one sentence, word caps) are enforced programmatically. An answer-then-improve scheduler banks one solid answer for every task first, then spends the remaining wall-clock re-verifying the weakest ones. A watchdog guarantees a complete, schema-valid results.json and exit 0 well inside the 10-minute limit. Engineered for the judge's 2 vCPU / 4 GB / 10-minute box and validated on a clean 2-vCPU machine: all 19 tasks answered in about 250 seconds with 0 placeholders and 100% on our held-out eval, at 0 Fireworks tokens. Because the heavy lifting is done by deterministic solvers, a small fast model is enough.