Carnegie Mellon University and US Department of Energy lab scientists built a test called Into the Unknown, described in a preprint dated 13 October. Each scientist supplied an unpublished follow-up idea and the conditions a paper must meet. Nine AI agents wrote 15 papers each, 135 in all. None met every condition. The preprint says 114 failed an integrity check, mostly for describing work never done or numbers the records contradict.
The Decoder reported that only three agents produced any clean papers, 21 of 135 in total. The other six had a violation in every paper. It said agents often checked their own work too leniently, in 58 of 64 traces studied. In one case, Claude Fable 5 forgot to start four planned runs. Its check reused that short list and reported 41 of 41 done. The Decoder also said GPT-6 Astra, the best performer, did the scoring. The sources give no evidence on bias either way.
A self-check built on the agent's own list repeats the agent's mistakes. Buyers of AI research tools can ask whether the evidence trail is checked apart from the agent that made it, and should treat the ranking with caution.