And the issue with that is that they're using PSNR and SSIM as quality metrics, which are useful for images, but haven't been designed for video, so they cannot

And the issue with that is that they're using PSNR and SSIM as quality metrics, which are useful for images, but haven't been designed for video, so they cannot take into account temporal factors and how motion masks distortion, etc. I think for AV1 they're using VMAF too, which is an actual video quality metric (https://github.com/Netflix/vmaf https://github.com/Netflix/vmaf).