AVGC diagnostics (hypotheses 1 and 3; diagnostic only) Deep models, frame accuracy % by stimulus shift (ms; negative = stimulus fed earlier): shift_ms -93.75 -62.50 -31.25 0.00 31.25 62.50 93.75 held model AVGC bmms3 52.58 60.58 66.72 65.22 57.14 49.14 46.10 directnet_sw2 53.87 63.06 69.12 66.42 56.03 46.51 43.76 DTU directnet_sw2 52.08 59.21 68.18 72.35 70.03 62.61 55.54 shift_ms -93.75 -62.50 -31.25 0.00 31.25 62.50 93.75 held model AVGC bmms3 0.066 0.273 0.429 0.391 0.182 -0.020 -0.099 directnet_sw2 0.094 0.339 0.504 0.428 0.156 -0.084 -0.158 DTU directnet_sw2 0.047 0.230 0.470 0.589 0.520 0.318 0.137 Linear decoder (trained on the other datasets), window accuracy % by envelope shift: shift_ms -125.00 -93.75 -62.50 -31.25 0.00 31.25 62.50 93.75 125.00 held window_s AVGC 5.0 50.39 54.96 60.79 63.23 59.53 51.72 45.59 44.32 46.81 10.0 50.34 57.14 65.00 67.90 63.36 52.94 43.88 41.77 45.37 DTU 5.0 47.84 50.62 56.74 63.19 65.21 60.53 53.10 48.45 48.18 10.0 46.83 50.96 59.70 68.32 70.39 64.36 54.34 47.60 47.74 Linear ceiling inside AVGC (window accuracy %, mean over subjects); R2 references DTU: XDS 70.5 / LOSO 68.3 / SS 73.1 (10 s): window_s 5.0 10.0 regime LOSO 59.34 62.58 SS 63.88 68.90 XDS 59.53 63.36