Striatica: Interactive 3D Visualization of Stratified Geometry in Sparse Autoencoder Feature Spaces

Errata (2026-03-26): The circuit analysis methodology described in this paper (Jaccard co-activation) has been identified as correlational rather than causal. Circuit-specific claims (multi-circuit membership, role assignments, edge directionality) are being revised. Geometric findings (VGT, local dimension, cluster topology) are unaffected. Integration with Neuronpedia Circuit Tracer for causal attribution is in progress. A revised version will be published when validation is complete. We present Striatica, an open-source interactive 3D visualization tool for exploring the geometric structure of sparse autoencoder (SAE) feature spaces in transformer language models. Applied to GPT-2 small (Layer 6, residual stream, 24,576 features), Striatica reveals geometric properties not apparent in existing tabular or 2D visualization tools, including distributed circuit topology, near-zero-dimensional convergence singularities at multi-circuit junctions, and simplicial substructure within circuit graphs. Striatica integrates local intrinsic dimension estimation via Volume Growth Transform, circuit overlay visualization, and per-feature inspection in a navigable 3D point cloud. Available at https://github.com/catwhisperingninja/striatica.