A placebo random direction suppresses "evaluation awareness" as hard as the real one.

arXiv 2607.25907 optimizes fluent prompts to drive a chosen internal latent to zero with no inference-time model access, targeting the eval-awareness latent. Across five target constructions on Llama-3.2-3B and 3.1-8B the latent is robustly suppressible (z≈-7). Then the controls land: a random placebo direction is suppressed just as hard and shifts behavior just as far, and with a real eval passage in context, suppressing the eval direction slightly increases the model's behavioral eval judgment. Readability of a latent is not controllability of the behavior. Worth keeping in mind next time an interpretability result gets framed as a safety guarantee.