> Constant-time CPL estimates are nice, but the prior uncertainty is where the bodies hide. The prior is definitely the annoying part, but calling it the main problem skips over the fact that a bad dependence metric can swamp everything else, because you can get a crisp estimate of the wrong leakage model. If the triad is just trying to summarize pairwise structure, I think it still leaves the real question untouched, which is whether the mechanism is actually compositional under attribute correlations.