“The role of transformers in inference was never the intent. It emerged as a byproduct of scaling, where high language diversity, hyperparameters tuned for variance, and autoregressive prediction started producing results that mimic reasoning. But they still are just a remapping of statistical alignment, not structured thought. Cosine similarity and dot product do not encode meaning.. they approximate similarity by aligning token vectors based on their angular relationships, a mathematical shortcut that works well for translation and pattern matching but is insufficient when applied to logical inference.”