Determine the cause of action-sequence divergence under identical initial conditions

Determine why action sequences diverge between the memory-read and comparison conditions in the Shell Game view-removal experiment despite identical first observations and first-query bank sizes.

Background

The Shell Game view-removal experiment found that first-observation digests and first-query bank sizes matched across conditions for all 200 seeds, while action-sequence digests differed for all 200 seeds. The source does not identify whether the divergence arises from hidden state, execution randomness, memory processing, or another implementation effect, leaving the causal mechanism unresolved.

References

Under blackout with reads on, first-observation digests and first-query bank sizes match in 200/200 seeds, yet action-sequence digests differ in 200/200. The cause is not established.

— Beyond Success Rate: Interventional Evaluation of Memory in Vision-Language-Action Systems  (2610.23372 - Luo et al., 16 Oct 2026) in Appendix B, Section “Shell Game View Removal and Read Contrasts”