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 release an answer only when they can establish a unique, reproducible proof. Remaining tasks are assessed by our fine-tuned, Q8-quantized FunctionGemma 270M model. Its predicted intent and five scores feed a per-intent matrix regression that may propose the embedded, text-only Gemma 4 E2B model. A hash-pinned Wilson 90% and Nash/minimax guard may preserve or reject that decision, but never expands the validated local cohort. Unsupported tasks, invalid assessments, rejected local answers, failed contracts, or runtime errors safely fall through to an evaluator-authorized Fireworks model. Every remote call uses the injected FIREWORKS_BASE_URL and only models supplied through ALLOWED_MODELS. Before inference, the runtime extracts the raw prompt from the official JSON envelope. The Answer Contract Engine validates the free-form answer and reconstructs results.json deterministically. In controlled testing, this boundary reduced Fireworks input tokens by 51.9% while preserving byte-identical Kimi answers. The public linux/amd64 image embeds both local models and requires no runtime downloads. Version v3.8.2-e2b-contract is 2.67 GB compressed and passed 712 automated tests, clean public-pull, manifest and OCI-label verification, official input/output validation, and exact published-image gates for 4 GB RAM, 2 vCPU, disabled networking, and the 10 GB image limit. The officially scored v3.7.3 rollback achieved 84.2% accuracy using 4,198 Fireworks tokens.