FlashProve proving incurs a computational overhead of 72x to 94x the GPU time of unproven vLLM inference on Llama 4 Maverick