LFM · lexical-readout screening

Separate model behavior from lens quality.

Three frozen speech-to-speech runs preserve strong late emergence, non-monotonic trajectories, and fitted-lens failures. The language readout explains generated text tokens—not acoustic frames.

Speech findings

Open detailed cached explorer

Compare fitted lexical ranks across LFM layers with the actual output head. A weak fitted rank can coexist with correct model generation.

Loading recorded reports…

A model result and a lens result are different.

The output head reports the model. It is the actual next-token distribution on the recorded generation path.

Intermediate ranks report the lens. They show what the fitted mapping recovers from an earlier residual state.

Generated audio is withheld. Only rights-cleared source audio is offered; the cached language evidence remains inspectable without it.