There's a difference between mapping rules in a general fashion, and applying them to actual situations. Invariably, reality always has elements that differ from the model, and thereby make it inaccurate. This is why reductionism doesn't generally map back accurately to the big picture. Understanding a component is different from understanding how the component integrates into the whole, and how other components within the whole interact with, and alter, the component.
Human behavior is very nearly impossible to accurately model at the individual level. You hear "I never would have expected him to do that" when people speak about their next door neighbors who go on homicidal rampages. Husbands or wives are shocked when the spouse that they've lived with for years or decades has an affair, buys a sports car, or has a second cup of coffee after dinner. Even crowd dynamics can be altered by incredibly subtle, and unpredictable, circumstances. (Someone in a crowded line says "He's got hairy wrists", someone else hears "He's a terrorist", and soon 8 are trampled to death in the ensuing panic.)
Then, add people deliberately gaming the system into the equation. (i.e. someone else wants to steer the crowd in a different direction than your model would normally predict, perhaps even using your model to predict the best way to disrupt it.)
Raw computing power just can't compensate for all the variables. At some point of hyper-precision, they become recursive anyway, and Gödel gets to have posthumous a laugh at your expense.