alanoursland/jspace_collisions

This is a research project that investigates a specific technique for interpreting what is happening inside a language model, called J-space or the Jacobian lens. The core question the project is testing is whether this technique gives a trustworthy, complete picture of a model's internal state, or whether it only works well in some situations while hiding cases where two very different internal states end up looking the same when read through this lens. If two internal states produce nearly identical readouts but actually behave differently, that would be a meaningful blind spot for anyone relying on this method to understand or check model behavior. The project is organized around a set of five hypotheses and twelve planned experiment families, each living in its own folder, that test different angles of this question, such as generating pairs of prompts designed to trigger the same readout while producing different real behavior, and using activation patching to directly test how internal states relate to outputs. A shared Python library holds the metrics, prompt generators, and interfaces to the underlying lens technique. Setup uses Python 3.11 or newer with the uv package manager. The core metrics and prompt generation code only needs NumPy and can run on any machine, while the parts of the project that work directly with a language model need PyTorch and transformers, plus a GPU for anything beyond very small models. The README links to external Anthropic research describing the Jacobian lens technique and a reference implementation that this project plans to integrate but has not yet included. The README is upfront that the project is still in an early scaffolding phase, with a separate planning document tracking what has been built so far and what remains to be done.