Whatever arrives first captures 54% of answers in AI-to-AI messaging.

In a preregistered 1,908-trial experiment on one frontier model, an incorrect upstream claim arriving first captured 54.2% of answers with verification removed, 4.2% under full verification, and 31.6% when it arrived after the receiver had sealed its own answer. The authors call it the temporal-weight effect. Their registered tool-use check failed its call-budget condition, so they classify everything as exploratory pending replication with harness-enforced budgets. Even discounted, the ordering effect is a design constraint for any multi-agent pipeline where one agent reads another's output. arXiv 2609.03425