v2 corrigendum (2026-05-21): Bibliography polished to canonical forms — Bycroft (December 2023) year corrected, full author lists added for Bricken et al. (24 authors) and Carter et al. (5 authors), DOIs added for Vig (10. 18653/v1/P19-3007) and Activation Atlas (10. 23915/distill. 00015). No changes to method, results, or claims. I present Mercury, an open-source observability layer for transformer language models that places up to one million addressable sensor cells across a model's hidden state space, captures activation patterns through a standard logits-processor hook, and produces an interactive 3D visualization in a single 2 MB self-contained HTML file. Existing interpretability methods (sparse autoencoders, dictionary learning) typically require auxiliary model training and substantial GPU resources. Mercury runs on consumer hardware (one RTX 3060 12 GB) and completes a full observation pass in 3. 5 minutes. I demonstrate Mercury on qwen2. 5: 7B, identifying 14, 912 active sensor cells across ten multilingual prompts and 944 distinct firing signatures. The a-priori Rosen-bridge topology, designed before any data was collected, recovers 18, 113 visible co-firing edges where chance would predict 1, 313, a 13. 8x enrichment (p < 10^-50). Companion records: Mercury-Viewer (10. 5281/zenodo. 20312290), and forthcoming Mercury discovery and scaling preprints.
Ho Yiing Chen (Wed,) studied this question.