Sparse autoencoder comparisons are measuring token position, not latents.

The standard protocol ablates a latent and measures effect at the token where it fires hardest, but that token is chosen by the dictionary under evaluation. Two dictionaries get compared at different places. Training six autoencoders from one initialization showed 7.6% and 11.9% of apparent inter-dictionary variance collapses to near zero once position is held fixed. Across a sixteenfold range of corpus sizes, dictionaries agreed less about where to measure. arXiv 2608.13337 The problem grows with scale, and the paper audits five published papers against a corrected protocol.