Separating perception quality from generator masking

Construct and evaluate a detector-by-generator factorial intervention for ARC-Agent that determines whether improving the hidden Markov regime detector recovers downstream portfolio value when the regime-conditioned diffusion generator is also repaired or replaced.

Background

Providing ARC-Agent’s scenario generator with the true regime does not improve performance because the generator may ignore or attenuate its conditioning signal. Consequently, the experiment cannot determine whether the regime detector itself is irrelevant or whether its benefit is masked downstream. The paper explicitly identifies the missing detector-by-generator factorial experiment as necessary to resolve this attribution problem.