ReproBridge fixes ML's reproducibility crisis

🔥 What happened Google researchers dropped ReproBridge, an agent workflow that turns scattered paper code into reproducible artifacts. Its benchmark spans 1,029 tasks from 387 experiments across 36 papers. 💡 Why it matters Refined artifacts lift reproduction success from 76.4% to 92.6% across four models and cut API spend by ~20%. The cross-model spread collapses from 21.6 to 3.9 points — execution knowledge beats raw model strength. ⚡ Our take Anyone still preaching "code is code" is losing. The real innovation isn't the model, it's how the knowledge is packaged — and that's exactly where labs will save money next.