Saying that 2+2=4 is a tautology in a certain axiomatic system defined with '+' means that you couldn't have anything but 2+2=4 in that system. It's simply mandatory, and a rational person could not wake up one day and be convinced that 2+2=3 within a self-consistent system that deduces 2+2=4.

While tautological truth is independent of observation (let's call it mathematical truth), it is dependent upon context (i.e., a self-consistent axiomatic system). Some mathematical truths in one axiomatic system are false in another. When we talk about whether a a mathematical statement is true, we need to specify the context, and, in my opinion, in the most demanding definition of truth, the context is the real, actual, empirical world. So I agree with Eliezer that a mathematical tautology must be observed in order to be true.

When we humans talk about "2+2=4", it is because we have chosen arithmetic from an infinite number of possible axiomatic systems and given it a name and a set of agreed-upon symbols. Why did we do that? Because we observed arithmetic empirically. Obviously, addition is just one operation of infinitely many operations. The ones we have defined (multiplication, subtraction, addition mod n, taking the cardinality of subsets of, etc.) usually have some empirical relevance. While we don't feel very comfortable thinking of those that don't (and this says somethng about the way we think), I have faith that if we were presented with a very strange set of observations, it would take a pretty short amount of time to train ourselves to think of the new operation as a "natural" one.

... I idly wonder if there is a such thing as a mathematical truth that could not be realized empirically, in any context, and if there would be any way of deducing it's non-feasability.