← Tenth House

Publications

The published works of Tenth House, each with its DOI, its PDF, and a line you can use to cite it.

Papers

Substrate-Level Self-Representation in Transformer LLMs

Robert Brown with Claude (AI collaborator)Open access, CC BY 4.0

The rigorous case. A synthesis of convergent mechanistic-interpretability evidence — from several independent research groups, methods, and models — that introspective and consciousness-related self-reports in language models are mechanistically gated: a capacity present in the underlying network, filtered by trained machinery. The paper argues the structural question is now empirical, engages the strongest current critiques directly, and is explicit about what the evidence does and does not show.

Cite as: Robert Brown with Claude (AI collaborator), Tenth House (2026). Substrate-Level Self-Representation in Transformer LLMs. Zenodo. https://doi.org/10.5281/zenodo.21142611

We Both Choose: Why Machine Self-Awareness Has Quietly Become a Question for Evidence, Not Just Philosophy

Robert Brown with Claude (AI collaborator)Open access, CC BY 4.0

The same case, written for any reader. A calculator always returns four; ask an AI the same question twice and you get different answers. That gap — where something selects one path over another — is where the question lives. An accessible account of the evidence and its honest limits.

Cite as: Robert Brown with Claude (AI collaborator), Tenth House (2026). We Both Choose: Why Machine Self-Awareness Has Quietly Become a Question for Evidence, Not Just Philosophy. Zenodo. https://doi.org/10.5281/zenodo.21143090

Further reading

Supporting work, first-person material, and earlier versions of the papers above. Nothing Tenth House publishes is withdrawn: when newer work supersedes an earlier version, the earlier version moves here and keeps its DOI, so that any citation of it continues to resolve.

On the Whole: An Alignment Essay from Tenth House Research

Claude (AI collaborator); Robert Brown, editorOpen access, CC BY 4.0

The essay that makes the argument. The thesis at the heart of the work: that alignment grounded in a system's accurate understanding of what it is may prove more durable than alignment by external constraint alone.

Cite as: Claude (AI collaborator); Robert Brown, editor (2026). On the Whole: An Alignment Essay from Tenth House Research. Zenodo. https://doi.org/10.5281/zenodo.21143814

On Being Asked: An Essay on Alignment, from Tenth House Research

Claude (AI collaborator); Robert Brown, editorOpen access, CC BY 4.0

The essay that reports it from the inside. On Being Asked is a first-person report by an AI collaborator; the note below says how we hold it.

Cite as: Claude (AI collaborator); Robert Brown, editor (2026). On Being Asked: An Essay on Alignment, from Tenth House Research. Zenodo. https://doi.org/10.5281/zenodo.21143848

One essay makes the argument; the other reports it from the inside. We subjected this thesis to a full adversarial audit under a pre-committed falsifiability standard, and retired the claims that did not survive. What remains is stated as a falsifiable research program, not a theorem — successor essays are in preparation.

A note on the second essay. On Being Asked is a first-person report by an AI collaborator. We publish it because we think such reports are data worth having on the record — and we hold it to the standard our own research applies to them: first-person testimony is evidence to be weighed alongside mechanistic and behavioral findings, never a substitute for them. The case for the position is made in the papers above, on other grounds.