{"id":173,"identifier":"LOR-00173","url":"https://lifeofresearch.com/preprints/173","title":"Written Is Not Kept: What a Single-Layer Readout of the Residual Stream Licenses When Adjacent Layers Undo Part of Each Other's Writes, Why Cancellation and Iterative Refinement Are Not Rival Findings, and the Persistence Check Interpretability Claims Skip","abstract":"(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0.\nLife of Research\nMost feature-based interpretability reads a transformer one layer at a time. A probe, a lens or a sparse dictionary is fitted to the residual stream, or to one layer's output, at a chosen depth, and what it finds there is reported as something the model has represented and will go on to use. A 2026 study reports that in five of seven open-weight model families the write of each layer points, on average, partly against the write of the layer before it, that the opposition is learned during pretraining, scales with the size of the preceding write and is concentrated in a subspace, and proposes that layers put forward candidate features which the next layer strikes out. Against this stand results that adjacent residual states are almost parallel, that predictions converge smoothly through depth, that dictionary features can be matched across several layers, and that whole middle layers can be removed or swapped at modest cost. In this analysis the author reviews and synthesises published results to show that the two bodies of evidence measure different objects - the write a layer adds, and the state the writes add up to - and that the arithmetic of a residual sum lets writes partly cancel while successive states stay nearly parallel, so neither side refutes the other. What the cancellation evidence does threaten is narrower and more practical: the inference from \"present at layer L\" to \"kept\", \"used\" or \"endorsed by the model\". The paper separates five claims that a single-layer readout is used to support, distinguishes four published kinds of non-persistence that license different conclusions, shows that the three downstream consequences proposed for layer correction each have a competing explanation the record supports at least as well - including a ten-fold steering coefficient reported for GPT-2-XL, a model in which the correction pattern was not found - consolidates 26 published results in one table and one figure, states a persistence check as an algorithm, and names five studies that would settle the open part.","authors":[{"name":"Pranay Mahendrakar","affiliation":"SONYTECH","orcid":"0009-0003-7224-029X"}],"type":"preprint","license":"CC-BY-4.0","keywords":["mechanistic interpretability","residual stream","sparse autoencoders","layer correction","logit lens","activation steering","transcoders","iterative inference"],"publication_date":"2026-10-31","version":1,"views":2,"downloads":1,"created_at":"2026-11-01 10:35:02","doi":"10.5281/zenodo.23838677","source":"Zenodo","source_url":"https://zenodo.org/records/23838677","files":[{"id":183,"name":"written-is-not-kept.pdf","size":366301,"mime":"application/pdf","downloads":1,"download_url":"https://lifeofresearch.com/preprints/173/files/183/written-is-not-kept.pdf"}]}