Dual Exam exposes agent, world-model failures

Dual Exam makes a strong case for component-level evaluation, but its headline scores are diagnostic rather than predictive of real-world system success.

• –Fixed environment responses isolate whether agents choose effective actions.
• –Fixed action interventions test whether world models preserve task-relevant state over time.
• –Results expose uneven performance across physical, digital, social, and scientific domains.
• –The zero-success coupled pilot shows that decent component scores do not guarantee reliable interaction.
• –Developers should treat agent and world-model benchmarks as separate debugging signals, not interchangeable leaderboards.