Masking Stale Observations Helps Search Agents -- Until It Doesn't: A Regime Map and Its Mechanism

Masking stale retrieved observations during multi-step search improves accuracy only under specific conditions: it peaks when a strong retriever is paired with a mid-capacity model, plateaus under weak retrievers, and collapses as model capacity saturates. The effect follows an asymmetric inverted-U curve across backbones ranging from 4B to 284B parameters tested on offline and live-web benchmarks. Practitioners should treat observation masking as retriever- and model-capacity-dependent, not a universal efficiency win.