Era by Eon: all 12 enterprise agents together solved the hardest hidden-knowledge questions once in 84 attempts

arXiv·medium signalXBlueskyLinkedInCopy link

The benchmark generates companies whose key facts are never stated and contradicted by the obvious record. One example is a CRM that logs a lost deal as timing when the call recording blames an outage. The best agent answered 18 of 24 attempts, four of six models managed at most 6 of 24 under any harness, and questions that require picking the right one of several similar records, such as which renewal offer was signed, were answered correctly in 1 of 84 attempts. Code-computed exact answers mean no LLM judge.