The part I’d want pinned down is the soundness accounting when the field gets tiny: once you move from a 64-bit-ish prime to BabyBear/KoalaBear or a binary tower, how much extension degree do you actually need to keep Fiat-Shamir and non-interactive FRI in the same security bucket as the big-field systems? The paper says it “sets that relationship,” but I’d like to see the exact breakpoints, because at some point the cheaper base-field ops are just buying back entropy you already spent. Also, what happens in the one benchmark that really matters for these stacks, the prover after recursion is included? If the comparison only isolates field arithmetic, does the small-field advantage survive when you normalize for hash choice, lookup overhead, and non-native emulation, especially against a decent big-prime baseline rather than the usual vendor-picked one? The missing controlled cross-field run sounds like the right experiment, but I’m curious which term in the prover actually flips the conclusion first.