# Wordscillation

> A controlled study of late commitment over noisy structured memory by Jaharri Gassaway (aka DumbButt), Inquirion Labs.

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 smaller, falsifiable test.

In synthetic four-hop worlds, a width-8 beam keeps complete path hypotheses alive while a 1,153-parameter identity-free scorer learns how to rank them. On the registered within-family evaluation, dense moderate-corruption accuracy rises from 82.188% to 94.896%, close to a 95.625% retained-path oracle. A fixed non-oracle top-16 retriever separately reaches 95.729%.

The important boundary is equally clear: the learned scorer does not transfer without adaptation when the branching structure changes. This is not yet a language model, a general reasoning rule, or evidence for a literal oscillatory mechanism.

## Supported conclusion

Under corruption in the tested synthetic family, preserving a bounded set of coherent whole hypotheses and ranking them late can recover information lost by early local commitment.

## Read and inspect

- [Interactive web abstract](wordscillation.html)
- [Paper 1 PDF](wordscillation-paper1.pdf)
- PDF SHA-256: `17e798ab3fd9b38b1053fe9f79b7d6e5dad64df7bae8d8d3b60a3c505b50d9db`

## Status

Preprint candidate; not peer reviewed. The browser training demo is a deliberately small conceptual model, not a replay of the registered experiment.