One may not reasonably be able to assign probabilities to different metaphysical realities. But one could potentially assign probabilities to the output of a method of evaluating whether something is true, e.g. apply a method to some claims and see how many you get right.
If your method is "believe everything you see", you may misclassify dreams, illusions or other mental conjurations.
If you refine this to exclude dreams and verify what you see, your misclassifications would go down, and the method would have a higher accuracy percentage (even if you can't assign a probability to a particular dream being true).
There are also multiple probabilities to consider, which can be based on some combination of true positives, true negatives, false positives, and false negatives.
The problem of figuring out whether even one belief (beyond "I am") is true may not be trivial, considering the unprovability of the existence of external reality and other minds. This may e.g. lead to some form or variant of coherentism, and considering whether a particular model, when applied consistently, produces internally consistent results.
If you cannot independently verify whether a claim is true, in some way, then that claim cannot serve as support for a method. If a method produced no independently verified claims, then, under this view, it cannot be deemed "good", and it certainly cannot be deemed "better" than any method that has produced independently verified claims (even if unreliably, although if it also hasn't produced disproven claims, it's not necessarily "worse" either - we just wouldn't have enough information to trust it).