David Widmann

> My view is that it's essential to encode the order statically when it's half-integer That's the point that's not clear to me - why is it essential? To dispatch...

IMO the formula is not a completely convincing argument since it can be implemented with simple branches (and I still assume generally you don't want to use it anyway). I...

I don't remember the discussion 😅 Not completely clear to me from your comments what exactly the problem is but I'm happy with a PR that fixes current issues 🙂

I think, even though this PR might fix the issue (untested it seems) it's not the fix we want: The package already defines (too) many methods (see eg. https://github.com/JuliaDiff/ReverseDiff.jl/issues/226) and...

There's no multiplication of a `Diagonal` with a `Cholesky` matrix in the linked example? The model only multiplies a `Diagonal` with a `LowerTriangular` matrix. In any case, you could rewrite...

This approach was adopted by eg AbstractFFTs, TranscodingStreams, and (soon) Distributions. I think we should make Test a weak dependency to reduce the number of dependencies and to be nice...

> To my mind, a better implementation would be to switch the check on / off depending upon which kernel is employed. Yes, I agree, that seems to be the...

This issue was fixed by https://github.com/JuliaStats/StatsBase.jl/pull/499.

I'm on vacation, hence only some quick comments: > Allows us to define specific AD-rules, e.g. we can use Enzyme for logpdf-computations without affecting the broader community IMO this is...

> For example, would you be willing to introduce an rrule for Distributions.logpdf that uses Enzyme to implement it? I think in general that should neither be done in Distributions...