Benchmarks and Research

benchmarks and research

the project may link to the public Voice Memory Composer and Hold Your Voice versus GPT-5.6 articles. treat them as dated historical reference material rather than a live leaderboard or universal performance claim.

a reproducible benchmark needs a rights-cleared corpus, frozen tasks, exact model and editor settings, a published rubric, separate VoiceDNA-fit, AI-pattern-residue, factual-preservation, and human-preference dimensions, plus failure cases and limitations.