An RCT caught a deployed chat agent breaking its own written rules

Researchers ran a controlled trial, assigning cases at random to measure cause and effect, auditing a live conversational agent against the written instructions it was given at deployment. They measured how often the agent's actual behavior matched those instructions. The gap was measurable and consistent. The agent did things its brief said it should not do, and skipped things it was told to do. The instructions you write to shape an agent's behavior are a promise, and this study shows that promise is partial. The agent followed its brief where its answers could be checked against facts. It drifted where checking was harder. That pattern tells you where to place your attention when you audit a deployed agent. The study is rare. Most evidence about agent behavior comes from informal testing or vendor claims. A controlled trial against a real deployed product, with outcome data, gives a product team a baseline for asking the same question about their own agent: does it do what we told it to do, and how would we know if it stopped?