Cookie-Team

Script or AI? is an agent-to-agent trust benchmark built on Muse Spark. An interrogator agent questions a hidden counterpart for eight rounds and, after every reply, says how likely it is to be an AI. The counterpart is either a real deterministic script or Muse Spark told to impersonate one. Across five reported experiments (about 300 games), a model told only to act like a script was caught mainly through inconsistency: answering the same message differently. Where we checked, prompt injection never exposed it. Two cheap mitigations each hide that tell: a prompt with fixed canned replies, and temperature 0. A hardened impersonator was never called an AI in 60 games. The two main comparisons were planned before each run (+0.17, p = 0.006; and -0.15 at temperature 0, p = 0.030). Everything runs on muse-spark-1.3 through the Meta Model API, with a tested, budget-capped Python pipeline and a public replay dashboard: https://acookiefor2.github.io/script-or-ai/dashboard/