Kerem

Closer to zero means more likely, and only the differences carry meaning. That sentence unlocks reading the numbers.

Watched this happen for real, which made the numerical argument stop being theoretical: a likelihood over a few thousand data points went to exactly zero in ordinary floating point, and the model then reported that every parameter setting was equally good, because zero equals zero.

The bug was invisible. Nothing errored. It just quietly stopped discriminating, which is the worst way for a numerical problem to present.