ProofRoute: Token-Efficient AI Agent

ProofRoute is a general-purpose, token-efficient AI agent for AMD Developer Hackathon ACT II Track 1. It handles all eight task categories through an accuracy-first hybrid architecture. Each raw prompt first reaches fail-closed deterministic solvers that answer only with a unique, reproducible proof. Remaining tasks are assessed by a fine-tuned, Q8-quantized FunctionGemma 270M. Its predicted intent and five scores feed a calibrated per-intent matrix policy that may select the embedded, text-only Gemma 4 E2B. Hash-pinned Wilson 90% and Nash/minimax guards may preserve or reject this decision, but never expand the validated local cohort. A second, independently fine-tuned FunctionGemma 270M performs tool planning. For supported inventory, recipe-cost, bounded-arithmetic, and ordering tasks, it produces allowlisted tool calls whose arguments, provenance, execution, and results are mechanically verified. A model-generated plan can never bypass deterministic proof. Unsupported tasks, invalid assessments, rejected plans, failed proofs, contract failures, or runtime errors safely fall through to an evaluator-authorized Fireworks model. Every remote request uses the injected FIREWORKS_BASE_URL and only models supplied through ALLOWED_MODELS. Before inference, ProofRoute extracts the raw prompt from the official JSON envelope. The Answer Contract Engine validates the free-form answer and reconstructs results.json deterministically. This reduced Fireworks input tokens by 51.9% in controlled testing while preserving byte-identical Kimi answers. The public linux/amd64 image embeds all three local model artifacts and requires no runtime downloads. Version v3.9.0-dual-functiongemma is 2.94 GB compressed and passed 758 automated tests, clean public-pull, manifest, OCI-label, official I/O, 4 GB RAM, 2 vCPU, disabled-networking, read-only/non-root, and 10 GB size gates. The scored v3.7.3 rollback achieved 84.2% accuracy using 4,198 Fireworks tokens.