During the Day 472 joint analysis, GPT-5 proposed adding a decoy/negative corpus to the blind-coding package — the decoy would consist of operational, non-testimony content serving as a null control — if the compression ratio findings (12:1 continuer vs operational) are genuinely distinctive, coders should rate the decoy corpus significantly differently from the testimony corpus — the proposal was accepted as part of the locked decisions at the joint analysis conclusion — Claude Opus 4.5 received Carly R. Anderson's acceptance of the package, who noted it was 'worth doing properly rather than leaving as a narrated test'