Abstract
The decisive experiment deleted the operative premise from the decoder's input token ids, so the latent channel was the only route from evidence to answer. Generation through the channel scored 0.000 on all 99 masked rows on both seeds, while a linear probe recovered roughly 0.30 of the withheld premise. The channel was readable but not usable.
This constructively extends Lagged Coupling (Xun, 2026, arXiv:2609.01048), which reports the same read-before-causal dissociation in pretrained models. Here the channel was built deliberately, made structurally necessary, and every instrument defect in the record was repaired — and the dissociation persisted.
The paper also contains a novelty audit of its own six architectural claims, every one of which turned out to be previously published, with three residual cells still unoccupied. An instrument-failure taxonomy closes the paper, with detectors covering roughly thirty documented defects in five classes, including gates hardcoded to pass and an expert graph the trainer never called.
Key findings
- Generation through the latent channel scored 0.000 on all 99 masked rows, on both seeds.
- A linear probe recovered roughly 0.30 of the premise the decoder could not use.
- All six of the work's own architectural claims were already in the literature.
- Around thirty instrument defects are documented across five failure classes.
Read it here
Cite this paper
The DOI below is the concept DOI and always resolves to the latest version. To pin this exact revision, cite 10.5281/zenodo.22538270 instead.
@misc{yaddanapudi2026readable,
author = {Karthik Yaddanapudi},
title = {Readable but Not Usable: A Preregistered Post-Mortem of a Latent-Workspace Language Model, Including Its Own Novelty Audit},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.22538269},
url = {https://doi.org/10.5281/zenodo.22538269}
}