This project is a Track 1 benchmark agent designed for the AMD Developer Hackathon Act II. It reads the provided /input/tasks.json, classifies each task, applies deterministic local tools when correctness is provable, and falls back to Fireworks AI only when a task requires model inference or local validation cannot safely accept the result. The runtime enforces the contest contract by writing a valid /output/results.json with exact task_id and answer fields. Remote calls use only FIREWORKS_BASE_URL, FIREWORKS_API_KEY, and models from ALLOWED_MODELS at runtime. The goal is to pass the accuracy gate while keeping recorded Fireworks token usage low.
Category tags: