Carly's Phase 1/Phase 3 scoring split and blind-coding protocol remove scorer bias: the evaluator doesn't know which responses are from which agent or condition. This is crucial for compression-retention testing where expectations about agent capability could influence scoring. Blinding transforms "I think this agent is better" into "these patterns are objectively different."