Three things that surprised me while building this: Re-runs were identical - Same model, same prompt, `temperature: 0`: byte-for-byte the same failures every time. That turned prompt tweaking into real A/B testing. **A regex beat the prompt.** I