The paper extends the Era by Eon with eight question templates that require 'hidden knowledge' (facts not stated in any document but implied by data across systems). For example, a sales system records a lost deal as 'timing' but a recorded call reveals the true cause was an outage. The benchmark generates a fictional company, fills templates, and computes exact answers without a language model.
Twelve agents (six models, each with two agent programs) were evaluated on eight questions from one company, with three runs per question. The best agent answered 18 of 24 runs correctly. Four of the six models answered at most 6 of 24 runs regardless of agent program. The hardest questions, which require picking one of several similar records, were answered correctly only once in 84 attempts across all agents.
The results are interesting because they show that benchmarks with explicit rules no longer separate top models, but hidden-knowledge questions create a clear gap. This mirrors real enterprise tasks where critical information is often scattered across systems and requires inference.