Another attempt at improving the noise model

[dfm]

This time I'm empirically calibrating the p-value distribution for p > 0.5 to be uniform. To do this I'm asserting that there is some constant fractional uncertainty in the noise model and I'm fitting for the value based on random subsets of the resulting p-values. Then I propagate the p-value uncertainty using Monte Carlo.

[dfm]

For reference, here's how this changes the p-value distribution:

Adding this uncertainty in the noise model decreases our confidence in both p = 0 and p = 1 targets which is why it matters!