Token Frontier Model

Token Frontier Router — a general-purpose AI agent for AMD Developer Hackathon ACT II, Track 1. Routes 8 task categories (factual Q&A, math, sentiment, summarization, NER, code debug/gen, logic) through Fireworks AI models, minimizing token spend while clearing an 80% LLM-judge accuracy gate. Built with a record-then-replay evaluation harness (frontier tracing, per-category gate policy, judge-proxy validation) to calibrate model choice and prompt design before every submission — architecture and gate survived two scoring-rule changes (accuracy-first, then no-local-tier) intact. Verified 89.5% (17/19) at 6,351 tokens on live evaluation; an isolated logic-template fix is staged and tested pending resubmission. Zero-dependency Python, Dockerized (linux/amd64), all inference routed through FIREWORKS_BASE_URL per competition rules.