Manifold Atlas

Tier: comparative tool. Object: output embeddings across models.

github.com/vector-lab-tools/manifold-atlas · All tools

Manifold Atlas compares multiple embedding models’ output geometries on the same inputs. It is the comparative cartographic instrument of the Vector Lab: the tool that operationalises vector theory empirically, turning what a model has learned about democracy, care work, negation, or intelligence into a testable geometry rather than a metaphor.

Why Manifold Atlas

Without vector theory, a cosine similarity of 0.95 between “fair” and “not fair” is a curiosity. With it, the same number is evidence for the negation deficit, for geometric ideology, for the proprietary encoding of human language that a particular training regime has produced. Atlas lets claims of that kind be tested against the geometry directly, across multiple models on the same inputs, with outputs that are comparable and exportable.

Operations

The tool packages fifteen operations for the critical testing of vector-theoretic claims:

Each operation makes a theoretical claim empirically testable against the geometry a given model has learned.

Theoretical background

Manifold Atlas follows from The Vector Medium and What Is the Manifold?. The tool names operations after vector-theoretic concepts (negation gauge, hegemony compass, real abstraction test) rather than statistical generics, so the connection between theory and instrument is explicit.

Stack

Next.js 16, React 19, TypeScript 5, Tailwind, Three.js. Talks to the major embedding APIs: OpenAI, Anthropic, Google, Hugging Face, Cohere, and local Ollama.

Status

The most mature instrument in the Vector Lab. Fifteen operations live, editorial design system settled, currently at v1.0.0. In active use for research and teaching.

Siblings

Manifoldscope is anatomical on a single manifold; Atlas is cartographic across many. The two are complementary. Vectorscope opens the model internals that Atlas compares at the output. Theoryscope applies comparable methods to corpora of theory rather than models. LLMbench handles the prose surface.