Cleanstone is a small independent research group studying where the abilities of AI systems come from. For a given ability we ask what training paid for, what came with the architecture before any training, and when during training it arrived.

Researchers at NYU's Center for Mind, Ethics, and Policy and Eleos AI Research sort the evidence about machine minds into behavioural, internal and developmental kinds, and call the developmental kind "an important but underdeveloped source of insight". That is the kind we work on, using open models whose training checkpoints are public. The about page has the detail and the sources.

First experiments

The first experiments start from Michael Levin's Ingressing Minds (2026). Levin argues that patterns from a non-physical space ingress into physical systems, so that even a system as small as a sorting algorithm shows competencies its design never paid for, and he proposes to measure them as the gap between the effort put into a system and the competency it shows.

We reran the sorting-algorithm experiments behind it, using the authors' own code, with predictions registered before each run and the controls the original studies lacked. Then we tried to build the effort-against-competency accounting on toy systems whose answers are known in advance, to see whether the measure gives the right answer before pointing it at anything interesting.

How we work

Every experiment is pre-registered. Every result reports a random-substrate null and the strongest deflationary account we could construct, and is published whatever it shows. Cleanstone issues no verdicts on consciousness, welfare or moral status; it publishes measurements others can draw conclusions from.

The plans, logs and results are kept in the open in the notebook, as they happen.