The concept of convergence in numerical methods could be useful here. In the AI case, it would have a large number of models ranging from high-energy physics to, say, social sciences, and it would run a number of models in the neighborhood of the one looking most suitable for a particular problem. It will check that the solution is not very sensitive to increase in resolution, i.e. to applying progressively more detailed models.

If it finds a situation where there is a decision gap between two neighboring models, it will make an effort to fill in the gaps in its understanding of the world before returning to solving a specific problem and validating convergence given its new worldview.