Theoryscope

Tier: single-model scope (alpha). Object: a corpus of theoretical texts.

github.com/vector-lab-tools/theoryscope · All tools

Theoryscope maps a corpus of theoretical texts as a high-dimensional point cloud and asks renormalisation-group-style questions of it. The tool makes visible what critics have until now asserted intuitively. Claims like “Foucault, Deleuze, and Guattari really come down to X” are claims about RG flow toward a fixed point. Claims like “Habermas and Rorty look opposed but converge” are universality-class claims. Theoryscope lets these be posed and inspected against a corpus rather than only asserted from the critic’s reading.

Why Theoryscope

Manifold Atlas and Vectorscope both take a model as their object. Theoryscope extends the same critical posture to intellectual fields: the corpus of critical theory, the literature on AI ethics, a single theorist’s output, a tradition across time. This is the move that lets the Vector Lab address not just the geometry of the machine but the geometry of the thought written about the machine. It opens a reflexive possibility: compare the eigendirections of a theoretical corpus against the eigendirections of a model trained on that corpus.

Operations

Drawing on renormalisation group theory and linear algebra:

A corpus as an object

The tool operates on a corpus of theoretical texts. Each text (or passage, concept, author, tradition, depending on the grain) becomes a point in a high-dimensional embedding space produced by an open-weight embedding model. The corpus becomes a cloud of points. Theoryscope provides operations for inspecting the cloud’s structure using language drawn from renormalisation group theory and linear algebra.

Theoretical background

Theoryscope is the empirical instrument that follows from Renormalising Theory and the forthcoming What Is Theory Space?, and continues the genealogy established in Hessler’s RG / deep-learning reading list together with Mehta and Schwab (2014) and Lin, Tegmark and Rolnick (2017). It operationalises the vocabulary: coarse-grain, observe flow, locate fixed points, classify universality. Applied to a corpus rather than a physical system or a neural network.

Stack

Next.js frontend, FastAPI backend. Shared geometry library with the other scopes.

Status

Alpha, v0.4.0. Twelve operations live across the Inspect, Flow, and Critique groups; further probes and a Critical Annotations Layer are planned. See the repository for a per-operation status table.

Siblings

Vectorscope treats model internals; Theoryscope treats a corpus. The reflexive move is comparing the two. Manifoldscope characterises a single manifold, which could be a Theoryscope corpus map. Manifold Atlas works across models rather than across texts. LLMbench reads the prose surface.