WordscillationDo not collapse the meaning too early.

I started Wordscillation with a simple intuition: when the evidence is uncertain, a model should not collapse to one meaning too early. Paper 1 turns that intuition into a controlled test. Synthetic memory packets bind source, relation, context, and target under a fixed storage budget. A width-8 beam keeps complete four-hop paths alive, then a 1,153-parameter scorer—without entity identities—learns how to rank those paths from structural evidence.

On the registered within-family evaluation, dense moderate-corruption accuracy rises from 82.188% for the frozen structural score to 94.896%, close to a 95.625% retained-path oracle. A separately registered fixed top-16 retriever reaches 95.729%. But the learned rule does not transfer without adaptation when the branching structure changes: it reaches 28.333%, near the frozen and shuffled controls. The honest conclusion is narrow. Late ranking can recover information lost by early commitment inside the tested family. This is not yet a language model, a general reasoning rule, or evidence for a literal oscillatory mechanism.