System One / calibrated decisions
one forward pass · no token generation · probabilities over your options
Compare raw Qwen/Qwen3.5-4B (instruct) — generated answer + letter logits
Enter a state and questions, then press Run.

Both models run live on ZeroGPU (Qwen3.5-4B + LoRA scorer, and raw instruct Qwen3.5-4B).