About
A research institute.
Not a consulting firm, not a lab with a fundraising deck. One thread since 2016: build and open-source research tools, and let the record say whether they hold.
- 2016 Paper notes, RL study guide
- 2017 Reproductions that others ran
- 2018 Congruent AI, reproducibility
- 2020 Program synthesis, Bourbaki Space, Piplio
- 2026 Agentic engineering
Fig. 6 — The same question, changing tools. The long version is on the history page.
01
What's it
We build and open-source research tools: the framework, the workspace, the record around the model. Then we point them at problems whose answers can be checked.
Credentials and the so-called meritocracy do not define us. Our discipline and practice for doing research in science does. We aim to challenge existing standards through end-to-end results. We are driven by questions, with intuition and tangible outcomes that align with society's need to progress.
We are not a consulting firm. We don't take your money for big snake oil AI projects, and to this day we still have not raised any venture capital investment. Our researchers may independently be available for hire, especially if you follow our guidelines and are nice enough to them.
02
Why the harness
The problem did not go away. It changed scale.
We started in 2018 on research reproducibility in machine intelligence, when the failure mode was a human publishing a result that nobody could re-run. That was a slow problem with a natural ceiling, since people can only publish so fast.
The thing producing unverifiable output is now an agent, and it does not sleep. It writes code all night. Almost none of what governs whether that code is any good lives in the model: it lives in the loop you chose, the tools you granted, the check you ran, and whether you kept the record.
That is the harness, and the harness is the engineering. A better model makes a bad harness fail faster. So we build harnesses. Pacenote constructs the agent, Novalis gives it somewhere to work, staxtrace remembers what it did. The same machinery points at mathematics, where Lean 4 settles the question without asking anyone.
03
Goals
Our main goal is to build a sustainable path to independent and reproducible research.
A common theme runs through the research goals:
- Support independent research in general.
- Make the tooling for agentic engineering open, inspectable and local-first.
- Make checking a generated result cheap enough to be routine.
- Provide tailored education in machine intelligence.
- Use whatever tools the question needs: human intuition, computer programming, machine intelligence.
04
Who
We're a team of academics, nonconformists, school drop-outs and independent researchers from all walks of life.
Our interests are driven by curiosity; the discipline of research is what defines us.