nx_voiceprint_gate.nx

nx_voiceprint_gate.nx -- REFEREE for VOICE-CLONE-001 rung 1 (voiceprint). Synthesizes three speakers with the SOURCE-FILTER model (glottal impulse train -> 3 cascaded 2-pole formant resonators, r=0.9, fs=8 kHz): A : vocal tract /a/-like, F0 120 Hz (period 67) A2 : SAME tract as A, F0 125 Hz (period 64) <- same speaker, different pitch B : vocal tract /i/-like, F0 182 Hz (period 44) <- different speaker PITCH-INVARIANCE + DISCRIMINATION (the identity floor under cloning): d_self = dist(A, A2) must be SMALL (same tract, different pitch) d_cross = dist(A, B) must be LARGE (different tract) assert d_cross > 4 * d_self -> the print captures WHO, not the momentary pitch. F0: A,A2 print ~120-130 Hz; B prints ~165-195 Hz (the source path works too). TAMPER (neg-control): a SCRAMBLED voiceprint (PARCOR order reversed) must be FAR from the original (> 4*d_self) -- a corrupted vocal tract is NOT the speaker, so the metric must reject it (no false "still you"). Every measured value PRINTED. Evidence -> stdout + knowledge/status/ voiceprint_gate.log. Exit 0 GREEN / 1 RED. Sovereign x86_64 throughout. license_tier: ORIGINAL