-
Eval harness v10 (canonical): all fixes baked in — sf.read tuple unpacks, proc.save_audio write path, demo-exact generate pattern
-
Eval harness v3: demo-exact generate pattern (return_tensors=pt, voice paths, tokenizer pass-through, max_new_tokens=None)
-
Eval harness v2: clean rewrite — no walrus helpers, ref embedding computed before arm loop, trackio imported
-
Eval harness v1: regenerate eval set with stock + SOAR heads, Whisper WER + resemblyzer SIM-O + drift, push to eval repo
