== DeepReview-targeted self-tests (gates: leader silence, mandatory delivery spec, IDR cadence) ==
[reference] score = 1.0   (expect 1.0)
[naive_try] score = 0.0   (expect 0.0)
[neg A: audible leader audio] score = 0.0   (expect 0.0, leader_silent_4s gate)
[neg B: shifted/non-IDR GOP] score = 0.0   (expect 0.0, gop cadence / IDR gate)
[neg C: MP3 44.1 kHz mono audio] score = 0.0   (expect 0.0, aac/48k/stereo + bitrate gates)

ALL TARGETED SELF-TESTS PASSED: reference 1.0, naive 0.0, all three violations rejected at 0.0
