ZeroProof —Zero-Token Proof-Carrying Routing Agent

ZeroProof routes every task to the cheapest source that can be trusted: fast deterministic solvers for the provable categories (math, code, logic, named-entity recognition) and a small bundled local model for the linguistic ones (factual, sentiment, summarization) — all running inside the container on CPU, at zero Fireworks tokens. Why this wins Track 1: zero tokens is now table stakes, so the real battle is generalized accuracy in the zero-token group on the hidden, refreshed task set, where memorized/overfit entries collapse. ZeroProof answers are proof-carrying — code is executed, math is re-derived and cross-checked, logic is constraint-solved, and entities are rule-corrected — so a reworded prompt cannot break a correct algorithm and a wrong "free" answer is rejected before it is written. A first-class internal "Proving Ground" measures generalization on unseen paraphrase/harder/adversarial task variants (never the public samples), proves 0 Fireworks tokens, simulates the 2 vCPU / 4 GB / 10-minute / 30-second limits, and guards against hardcoded answers. Measured results: 100% on deterministic categories across unseen variants, ~95% overall including the bundled model, 0 tokens, all runtime limits green, image ~1.6 GB. Licensed cleanly (MIT code + Apache-2.0 model).