.png&w=256&q=75)
🤓 Latest Submissions
TranscendiantRouter5
TranscendiantRouter is built on a simple observation about how this track is scored. Local tokens are free, so the cheapest way to catch a local model's mistakes is to use the local model itself. The agent works in four steps. First, it sorts each task into one of the eight capability categories using simple rules that cost nothing. Second, it asks a small local model (Qwen3-1.7B, served by llama.cpp with the weights baked into the image) to answer the same task several times, then measures how much the answers agree. Agreement turns out to predict correctness far better than a model's own confidence. Third, if agreement clears a tuned threshold for that category, the majority answer is used at zero token cost. Fourth, if the answers scatter, that single task is escalated through the Fireworks proxy with a short prompt and a capped token budget. The thresholds were tuned on a benchmark harness of 227 tasks drawn from GSM8K, HumanEval, and hand-written questions, with a held-out split that was never used for tuning. The router also carries safety rails. It cannot call any model outside the ALLOWED_MODELS list. A failed remote call falls back to a local answer instead of a blank. A time guard reduces sampling before the ten minute limit is ever at risk. On a replica of the grading environment with 4GB of RAM and two CPU cores, the full run finished in just under six minutes with 92.5 percent accuracy on held-out tasks, and most answers cost zero tokens! (This is the thirteenth iteration.)
13 Jul 2026
