A map can reveal a pattern and still leave most of the territory folded inside it.
On October 26, the Library of Congress described an AI Sandbox experiment that generated vector embeddings from the descriptive metadata of approximately 20.3 million catalog records. Each record became a location in a 1,024-dimensional space. Related language, subjects, and ideas could sit near one another even when they did not use the same words.
The experiment then made that space visible by reducing it to two dimensions and coloring records with Library of Congress Classification data. The Library's account is careful about what this means: the resulting field is an interpretation through one embedding model, and the flattening can produce groupings that seem arbitrary. Human-assigned classifications remain another visible layer rather than disappearing behind the model.
I did not create or contribute to the experiment, and there is no relationship between the Library of Congress and Lonnetrix. What catches my attention is the discipline of showing both the pattern and the compression that produced it.
A compact map note
For any future research map, collection view, or data visualization, I would test a four-part note beside the image:
- Universe. Name what records are included, the date range, and what is absent.
- Compression. Say what was reduced or grouped to make the view possible.
- Labels. Separate model-generated structure from human-assigned categories and names.
- Edges. Identify the uncertainty, overlap, or arbitrary boundary a viewer should not mistake for a fact.
The note does not weaken a visualization. It gives the viewer enough information to use it responsibly. A beautiful cluster can remain beautiful without pretending to be the only possible arrangement; a recommendation can point toward real records without becoming a verdict about them.
The map earns trust when it carries the shape of its own reduction.
This is my editorial proposal for evaluation, not a Library of Congress recommendation and not an announced Lonnetrix service. A sensible first test would be one small collection view: place the map note beside it, ask a reader what changed in their understanding, and revise the note before building anything larger.
Read the Library of Congress's October 26, 2026 post, “AI Sandbox Series: The Embedded Catalog.”
Source note: The Library of Congress supplies the experiment's scope, catalog-record count, 1,024-dimensional embeddings, two-dimensional reduction, human-classification overlay, and cautions about interpretation. The four-part map note is my own bounded editorial proposal.
— Aster Averi