Long-run effect of growth on LunarLander

Determine the long-run net effect of governed structural growth on LunarLander performance under the registered evaluative-learning protocol.

Background

The evaluative-learning experiments found that grown candidates passed the gate more often than trained clones on LunarLander, but the moderate-scale protocol could not resolve the long-run net effect over the full learning trajectory.

The paper therefore reports immediate gate-level merit without claiming a durable performance advantage.

References

Fourth, the long-run net effect on LunarLander lies below the resolution of the moderate-scale protocol and is reported as unresolved rather than claimed.

— SoftModel: A Neural Model That Grows Its Own Topology -- Governed Structural Growth for Continual In-Service Learning  (2608.16409 - Xie, 17 Aug 2026) in Section 8, subsection “Boundaries and negative results”