How to read the corpus probabilities

When you see a case split into "46% mundane/natural, 25% indeterminable, 9% open non-human…", that is not a hunch dressed up as a number. It is a distribution: each incident case splits 100% among the same six explanations, which exclude one another and cover every possibility. Because every case uses the same explanations, the numbers are comparable —one can coherently say which explanation accounts for more cases.

But comparable is not the same as true. These splits are structured analytical judgments, not measured frequencies: the model says which explanation is most coherent with each case's analysis, not which is objectively correct. That is why the "indeterminable" category exists —the share the available evidence does not let us assign—. It is the model's honesty valve: not everything can be resolved.

The archive total adds those splits into the expected number of cases per explanation. Being an expectation (linear), it holds even if the cases are correlated. When you doubt a number, don't look at the isolated decimal: look at the case's full split and how much lands in "indeterminable". That is where what we still don't know is on display.