Separate generation conditioning from decoder reference context

When a voice-cloning option changes both generation conditioning and waveform-decoder context, improved audio does not identify which change caused it. Hold the generated codec trajectory fixed while changing only decoder context, then separately swap the conditioning example while preserving the speaker embedding. This distinguishes upstream register selection from downstream waveform effects without treating a multi-factor option as a clean ablation. A shorter reference can be useful, but validate its matching transcript, varied wording, termination and concurrent requests; pitch alone does not establish perceived speaker identity. Applies when: Reference changes appear to fix inconsistent cloned speech. Keywords: voice cloning conditioning decoder context codec trajectory speaker embedding ablation