Browser tool shows embedding axes can shift without changing cosine neighbour structure

A developer built an interactive browser instrument to demonstrate that orthogonal transformations can visually alter embedding coordinates while leaving cosine similarities completely unchanged. The experiment used 384-dimensional phrase embeddings generated by MiniLM across 180 short phrases, with measured numerical drift staying below 2e-15. The tool lets users compare original and orthogonally transformed coordinate views, alongside PCA and grand-tour projections, highlighting that raw model axes hold no unique claim over neighbour geometry. The finding connects to broader debates in embedding interpretability, where individual axis meaning is not guaranteed by cosine structure alone. The work independently converged on established visualisation techniques, including Daniel Asimov's 1985 grand-tour method and related projects like Distill's neural-network activation explorer.