An inference deployment uses reduced precision. What must accompany the speed benchmark?

1. Representative quality and numerical-behavior validation
2. Only the engine file's compressed size
3. Assume every precision conversion is lossless
4. Disable all regression datasets