bench/

Command audioval validates the audio analysis (packages audio, audio/loudness, audio/defect):

Command digestsim checks how well the digest of a ladder stands for its title.

Command gpuval validates qc on an NVIDIA GPU and writes a Markdown report to paste back: NVDEC decoding speed and frame identity, CUDA VMAF against CPU VMAF (agreement and speed), NVENC ladders against CPU ladders (time, bitrate at equal VMAF), and a CQ sweep calibrating the NVENC probe CQs.

Command ladderreplay replays the ladder engine on an exhaustive rate-quality grid instead of encoding: every encode the engine asks for is answered from the grid (interpolated in CRF), then the rungs are checked against the grid's own optimum.

Command ladderval checks a fast ladder (qc ladder -f json) against the exhaustive optimum of the whole title: it encodes the full title on a dense resolution × CRF grid with exact VMAF, builds the true rate-quality envelope, then encodes the full title with each rung's settings and reports how far every rung lands from the envelope.

Command motionval checks the camera motion analysis (analyze/motion) against synthetic clips of known camera motion: pans, tilts and zooms at several speeds, handheld shake, combined moves, moving objects, parallax, grain, blur, fades and a cut.

Command vmafsim checks the sampled VMAF estimator against ground truth: it replays the production sampling loop many times on the per-frame scores of an exact measurement (qc vmaf --exact -f json) and reports the error, the real coverage of the confidence intervals and the share of frames scored.