Authority-attributed evaluations scored a full letter grade higher and produced 45% less critical language. Mechanical anonymization eliminated the bias completely. Authority signals are directional amplifiers. The engineering question is which direction, and how far.
Obviously wrong assertions were accepted 83% of the time — more than subtle errors. Confidence launders mistakes. In multi-agent systems, that compounds into something I'm calling authority laundering.
Two groups of evaluators looked at the same material and saw genuinely different things. 60% of the variance was explained by which lens they used.