Independent AI Research · Richmond, Virginia

What does a model
actually know?

Corten Research studies meaning, grounding, and auditability in large language models. We build open tools, run reproducible experiments, and publish what we find, including the parts that don't work yet.

Focus

Three problems we keep coming back to

Most AI failures in the field trace back to one of these. None of them get solved by a bigger model.

01 · Meaning

Grounding

Words don't carry meaning around inside them. Meaning lives in relationships and use. We measure how models represent that and where the representation falls short.

02 · Purpose

Deference

When an instruction is ambiguous, whose reading wins? We look at how systems should hand judgment back to the operator instead of guessing at a "real goal."

03 · Trust

Auditability

There's a gap between what shaped a generation and what you can recover from the output afterward. We build instruments to close it.

Words are not containers. You can't open one up and find the meaning sitting inside.
Working thesis · Corten Research

Current work

On the bench

Working paper · Open suite

Words Are Not Containers

A cross-model study of how transformers tell idiomatic language from literal language, and what that says about meaning. The findings replicate across three open model families.

Read more →

Framework · arXiv / Zenodo

Open General Intelligence

A modular, operator-governed architecture for general intelligence systems, with a formal mathematical foundation and an open reference implementation in progress.

Read more →

The Fellowship

A learned society for people doing the work

Corten Research is home to the Fellowship of Corten Crows, a small, chartered society of independent researchers. We trade drafts, break each other's experiments, and hold the work to a standard before it goes public.

About the Fellowship →