Vectorscope
Tier: single-model scope (alpha). Object: a single open-weight model.
Vectorscope opens a single open-weight language model for forensic inspection. It reads weights, activations, attention, and token embeddings at every layer, turning the model from a black box into an object of empirical analysis. It is the anatomical instrument of the Vector Lab.
Why Vectorscope
Commercial embedding APIs return sentence-level composites from separately-trained embedding models, the output of a pipeline rather than a window into the model itself. That is fine for some applications and inadequate for critical work on how language models actually organise meaning. Vectorscope privileges open-weight models where every layer can be inspected: weights, attention matrices, hidden states, token embeddings, and the operations that compose them. The tool is designed to make the internal geometry of a model legible, on its own terms, at the resolution the question demands.
Operations
- Layer-by-layer inspection. Hidden states, residual stream contributions, and intermediate representations at each transformer layer. Follow a concept through the stack and observe how its position drifts.
- Attention analysis. Which heads activate on which inputs, and what structural patterns do they encode? Visualise attention matrices per head, per layer, per token.
- Token embedding exploration. The pre-contextual vocabulary, before any input has been read. Which tokens sit as neighbours, which as antipodes, and which have no neighbours at all?
- Signal degradation laboratory. Run the same input across precision regimes, FP32 through BF16, INT8, INT4, FP4, INT2, and observe how the signal compresses as the medium is quantised. The material substrate of a representation shapes what it can hold; Vectorscope measures the shaping.
Theoretical background
Vectorscope is the empirical instrument that follows from the essays What Is Vector Space? and Vector Theory. Where those essays characterise vector space as a new medium constituted by dimensionality, Vectorscope is the instrument that makes that space legible. The signal-degradation laboratory implements the methodological commitment of the Leverhulme Centre for Vector Media bid, treating precision regime as a material parameter that shapes meaning.
Stack
Next.js frontend for the interface and visualisations, FastAPI backend running PyTorch and Hugging Face Transformers for model loading and forward passes. Model weights are loaded locally; nothing is sent to third-party services. Works with any open-weight model Hugging Face can load.
Status
Alpha. The architecture is stable; individual operations are at varying levels of maturity. See the repository for the current state of each operation and its known limitations.
Siblings
Manifoldscope takes the output of a Vectorscope-inspected model and characterises it as a manifold in its own right. Manifold Atlas compares output geometries across models. Theoryscope extends the same critical posture from a model to a corpus of theoretical texts. LLMbench handles the prose surface.