I've read Taleb's first book, waiting for Black Swan soft cover to come out...
The fact that we have to resort to statistics just means that there are lots of stuff we don't know about a system so we use a large data pool to give us an average idea of what the system is doing. Consider quantum machanics where you decribe electrons as probability curve or in thermodynamics where you don't track each individual molecule but resort to statistical mechanics to tell you the average which is what pressure and temperature are. Because these are physical systems so they obey certain invariant property 99.9999% of time so using such measure in engineering is acceptable. Even in science and engineering, problem often must be simplified in order to have a solution (okay I mean analytically e.g. we can only solve two body problems analytically anything higher we must resort to numerical technics) otherwise the problem would be too complex to solve. But by simplifying you are eliminating certain information but in engineering we say approximation is good enough (most problems that are solvable are linear meaning first order, anything of higher order are difficult to impossible to solve--consider any high order PDEs so the higher order effects if minimum are lopped off so that we would have progress in the world). Again these simplification works because the invariant physical systems, the obey universal law. When you use these technics to model financial and economic topics which are "man-made" that includes people's emotions, they will not do a good job and thus subject to blow-ups......
In terms of the water tank example, you would model the structural portion of the tank separately from the flowing portion and if time and money permits you could subject to vibration etc etc. Ther eis not one model that encompasses both the flowing and structural part you'd maybe calculate one and use that as input into the other one....