CleanstoneDevelopmental and independent evidence for machine minds.

Tree D: direction questions

Q1 open Overlap with Anthropic and Eleos

Concept injection is Anthropic's method and the welfare axis is NYU's. What we add is the untrained null, the non-linguistic barrier task, and the dissociation test on open weights. Any write-up has to say so in its first paragraph, or readers will take the work for a replication.

Q2 open Ethics posture

The tree measures states in systems that some people consider moral patients. By default we make no welfare claims and use no aversive protocols beyond those published work already uses, and the pre-registration carries a written note on this.

Q3 open Order in the programme

If the group enters before it has a published deflation, it risks being filed with the advocates. That is why the tree comes after Trees A, G and B. Revisit if a collaborator with a track record offers to co-run it earlier.

None blocking as of 2026-09-30.