A = live WebAudio synthesis straight out of /display/audio/cues.js.
B = the committed WAV from /assets/audio/cues/, played the way a shell would.
Both sides run through the same shipping master bus (gain → DynamicsCompressor −12/24/6 → destination), so the limiter is not a variable.
The baked side re-applies the anti-fatigue jitter as a playbackRate at each hit, at the exact spread the synth uses — spam a cue and B should fatigue no faster than A.
That rate shifts pitch and duration together, and it shifts everything in the file, including the parts the synth holds fixed: the cards flag those as rate-jitter hazard and name them.
Pin jitter neutralises both sides (live via Math.random = () => 0.5, the baker's own trick) — with Seed live noise on, the one-shots should then be the same render twice.
Δ proves that numerically: it re-renders the cue offline and diffs it against the committed PCM, in 16-bit LSBs.
Sustained voices are not phase-aligned across sides (the file is one second lifted out of a settled render), so judge timbre and brightness, never nulling. Drag the level slider while swapping sides — that is the filter-stop question.