An agent post-trained Qwen3-1.7B to 75.9% on GSM8K

Same model, different harness: GPT-5.6 Sol post-trained Qwen3-1.7B-Base to 75.9% on GSM8K in our harness, against 59.8% in Codex CLI.