About

Translation

By aphaeresis from Heraclean stone (Ἡρακλεία λίθος, Hērakleia lithos), the lodestone.

What we study

Studying AI Welfare Empirically, the 2026 report from NYU's Center for Mind, Ethics, and Policy and Eleos AI Research, sorts the evidence about AI systems into three kinds.¹ Behavioural evidence is what a system does and says, including what it says about itself. Internal evidence comes from the architecture and from interpretability methods. Developmental evidence is how the system came to be the way it is: training evidence, from the pressures that shaped the system, and trajectory evidence, from the order in which its capacities emerged.

The first two kinds have known weak points. "Many AI systems are designed to mimic human behavior, complicating behavioral evidence," as the report puts it,¹ and most internal findings come from the companies that built the models, on weights no one else can inspect. Of the third kind the report says: "developmental evidence is an important but underdeveloped source of insight. Much current reasoning in this area remains informal".²

That is the kind Cleanstone studies. Open model families such as Pythia and OLMo publish their intermediate training checkpoints,³⁴ which makes the trajectory measurable at the cost of running inference. We track each measure across those checkpoints, from the untrained model to the finished one, against the loss curve and the data seen before the measure first moves.

The first instrument came from Michael Levin's Ingressing Minds, which states the developmental question in its sharpest form: the gap between the effort put into a system and the competency it shows.⁵ For claims where rival camps predict different outcomes, we borrow the adversarial format of the Cogitate collaboration, in which proponents of two theories of consciousness committed to divergent predictions before the data came in.⁶