>Think of complex numbers, for example: they are used extensively in quantum physics and electrical engineering, but that does not mean that you can see them, touch them or measure them.
yes, we use complex numbers there only because we have a very consistent mental map between their properties/behaviour in the math model and the physical reality. The physical reality doesn't obey the math model, it is a successful and useful math model which seems to consistently describe and predict the reality in some range of conditions. So when a new model is brought in you'd want it to have at least some partial consistent mapping to reality, and with refining of a model the mapping is expected to improve. If refining of a model turns instead into rounds and rounds of complexity growths without matching massive improve in the consistency of mapping to physical reality, you'd naturally start to question the model, and may be think, kind of VC style, about spending say a few percent of resources going into that massive complexity growth of the currently dominating, yet still far from working that good, model into exploring some alternative models.