STATUS: read this first

Start here if you are new to the project, human or agent. Updated 2026-10-30, evening.

What this is

A research programme on developmental and independent evidence for machine minds. We ask what a system's training paid for, what came with the substrate and what is left over, and we measure it on open systems with every experiment pre-registered. PREMISE.md is the charter. The programme began from a reading of Michael Levin's Ingressing Minds (2026), written up in Review.md.

Where things are

There are two groups of work. Each is a folder of numbered lines, and every line has the same six files (plan.md, summary.md, log.md, direction-questions.md, todo-operator.md, todo-agent.md). summary.md is the plain-language version, and log.md ends with a Next: line.

trees/ holds the lines derived from Levin.

• 01-a-ledger was closed by the operator on 2026-09-30 after four experiments (writeup.md). The sorting-algorithm competencies in Zhang, Goldstein, Levin (2024) deflate under controls, including when we ran the authors' own simulation code unmodified. The "effort in, competency out" accounting cannot be built for designed systems. Its trained-system form continues as spine/01 measure M5. We sent the write-up to Levin on 2026-10-30 and have not heard back.
• 02-g-theory has sections 2 and 4 drafted (draft-accounting.md), its reading list (reading.md) and the claim checklist (checklist.md, 13 rows) written. The reading corrected four claims in our own documents, among them a programme wrongly attributed to Birch. Next is the full draft. Author: Elara Faro.
• 03-b-periscopes is not started. It takes the barrier harness from Tree A, and nothing from the ledger.
• 04-c-convergence and 05-d-assays are planned and need a GPU.
• 06-f-wetware is parked indefinitely, since there is no cross-substrate ledger.
• 07-e-persistence is parked.

spine/ holds the lines built from scratch, planned on 2026-09-30.

• 01-developmental-checkpoints tracks indicator trajectories over Pythia and OLMo checkpoints. The analysis side (atlas.py) is written and self-tested; one open question (Q5) on its loss control must close before the pre-registration. Everything that touches a model needs a GPU.
• 02-open-replication runs pre-registered replications of frontier-lab mechanistic claims on open weights. The workspace method reconstruction (method-r1.md) is written, with 39 questions for the authors. Reference code and fitted lenses for open models already exist, and our plan had the method wrong; both corrected. Runs need a GPU at the 8B class.
• 03-adversarial-registry has its protocol (draft 2, reviewed against Tree G's checklist), template, the first-pairing proposal and the approach emails. The first pairing changed: Camp A is Pearson-Vogel et al., not Berg, whose work is about something else. Tree A is entered retroactively (entries.md). It needs no compute; it waits on the operator sending the emails.

Blocked on the operator

The top-level todo-operator.md collects every per-line operator item. The next one is sending the registry emails. The GPU box and the GitHub organisation are deferred by decision (2026-10-30), so model runs wait and saves stay local. The server had 9 GB free on 2026-10-30, which is enough for the standard-library work and too little for checkpoints.

What the agent does next without the operator

Done on 2026-10-30: the first-pairing proposal and approach emails, the retroactive registry entries, protocol draft 2, the workspace method reconstruction, the atlas analysis tooling, and Tree G's reading list and checklist.

Next, in this order:

1. checklist answers and a power estimate for R-2026-003 (spine/03)
2. the email to the workspace authors with the 39 questions (spine/02)
3. the loss-control fix in atlas.py (spine/01 Q5), then the GPU-side tooling, written and tested as far as possible without a GPU
4. the full draft of Tree G's position paper.

The site

just ship rebuilds the repo as two notebooks. The private one carries everything and is not reachable from outside (CLAUDE.md has the details). The public one is an allowlisted subset at https://cleanstone.org/notebook/, live since 2026-10-30. The project name is Cleanstone (cleanstone.org). Contacts, outreach drafts, raw notes and papers stay unpublished.

How to work here

• CLAUDE.md has the working rules. just check runs after every edit.
• Each experiment is pre-registered in its tree's plan.md before any code is written, and its result goes in log.md with a table. For Tree A, writeup.md is the account written for the authors.
• Prose follows style.md, including its extended list of machine tells.
• No claim about consciousness, welfare or moral status leaves this repo.
• The raw research behind Review.md is in notes/, with a URL for every claim. domain-experts.md lists people and feeds. naming.md records how the name was chosen: Cleanstone, 2026-10-30.
• The machine has 2 cores, 8 GB and a root disk at 91% (9 GB free on 2026-10-30). Only the standard-library experiments run here.