The revision in the face of observations relies on them being new observations. You can make an almost infinite number of models that fit existing knowledge; th

The revision in the face of observations relies on them being new observations. You can make an almost infinite number of models that fit existing knowledge; the way you figure out which one to go with is by having the model make testable (falsifiable) predictions, and then test those predictions. The string theory revisions, however, seem to fall out of conflicts between the theory and existing observations (like, say, that the dimensionality of our physical space is 3, or that the cosmological constant is non-positive). For example, general relativity could make Newtonian mechanics fall out as a low-mass, low-velocity special case, and solved the already-known theoretical problem of the constant speed of light, but there a lot of conceivable models that could solve that. To be accepted, its predictions needed to be tested (starting with gravitational lensing in the 1919, and going through higher-precision tests later). This is related to the problem of overfitting in machine learning. You can get a system to be very very familiar with your training data, so it can predict the things it has already seen. However, until you have validated it on data it wasn't trained on, you don't know if the model reflects any of the underlying properties of the system it's observing.