asraa

My fix is to create a pattern that converts constants yielded to conceal ops - i'll put it up for review for debate.

But my fix breaks this test: https://github.com/google/heir/blob/a422f130fa42ed2acf319a1bd6e420b651863c03/tests/Dialect/Secret/Transforms/canonicalize/secretless.mlir#L4 which now seems to question this pass. i would expect we only collapse them when it returns plaintext types (which isn't even allowed)....

> %rotated_tensor_of_input = linalg.elementwise { rotate } on %input_splatted +1 on this too. I noticed in the MLP example when I had the for loops in, the noise would grow...

> Further, while current implementation computes the polynomial directly a better way would be to compute it as, $x *(0.197 - 0.004 x^2) + 0.5 $ where the number of...

Blocked on https://github.com/llvm/llvm-project/issues/129901

I think the other requirement besides making sure they've called the TW correctly and used the SLSA inputs from the token correctly is - [ ] Validate that the build...

I think it should be a resource descriptor - we get the URI and the hash explicitly, and can also add new fields. I think they should rephrase the migration...

> Just to be clear, this is a totally separate Resource Descriptor that's separate from the entry in resolvedDependencies? I understood it to be duplicated, since it is (1) a...

> Oh, I see - I'm not sure type hints are the right way to go about this (a custom numba intrinsic cast(,) might be easier), but I'll see what...

Oh! That's amazing- that works great (as far as sanity check goes, i think some of that same logic of what to emit could be reused - e.g. if we...