Premise

Where this charter says "the institution" it means Cleanstone.

Purpose

To put claims about machine minds on developmental and independent evidence. We want to know what a system's training paid for, what came with the substrate, and what is left over. We measure on open systems and pre-register every experiment. Replication is done by people who expect the opposite result, and we publish whatever the outcome.

The gap, stated as facts (2026-09)

• Every mechanistic finding about consciousness-relevant structure in language models is a snapshot of a finished, closed model, evaluated by the lab that built it. There has been one independent replication so far, and it changed the interpretation of the result it replicated.
• The field's own methodologists (CMEP and Eleos, July 2026) name three classes of evidence: behavioural, internal and developmental. The developmental class asks what training did and did not produce. Their report calls it "an important but underdeveloped source of insight" (Part II §1.1.4). Open checkpoint families (Pythia, OLMo) make it tractable at inference cost.
• Butlin et al. (TiCS, April 2026), who set out the theory-derived indicator method, concede that behavioural evidence cannot separate instantiation from trained performance. Self-report is compromised by the gaming problem. That leaves internal and developmental evidence.
• Cogitate tested two rival theories of consciousness in an adversarial, pre-registered format: each camp committed to its predictions before the data came in, and a neutral team ran the experiment. We know of no one applying that format to AI systems.
• Levin (2026) gives the developmental question its sharpest form, the delta between the effort put into an embodiment and the competency observed. He does not operationalise it, and nobody else has either.

The programme

There are three spine lines and one instrument line, in folders under spine/ and trees/.

1. Developmental checkpoints (spine/01-developmental-checkpoints/). We trace every measurable indicator over training on open checkpoints, alongside the untrained null, the loss curve and the data seen before each onset. The output is the checkpoint atlas.
2. Open replication (spine/02-open-replication/). We replicate frontier-lab mechanistic claims on open weights under pre-registration, add the controls the originals lacked, and publish any deflation with the replication. The output is a set of replication reports and shared tooling.
3. Adversarial registry (spine/03-adversarial-registry/). A public registry of pre-registered claims on which rival camps predict different outcomes, run by a neutral executor, with results published whatever they show. The output is the protocol and the completed entries.
4. The instruments (trees/). Tree A tried to turn Levin's residual into a calibrated number. It closed on 2026-09-30 because the accounting cannot be built for designed systems (trees/01-a-ledger/writeup.md). Two pieces survived. The trained-system form, behaviour bits against training-signal bits, became a per-checkpoint measure in line 1, and Tree B kept the barrier harness. Tree G supplies the theory and the checklist that every line uses. Trees C, D and E are follow-on lines, sequenced in research-directions.md. Tree F is parked.

Between them the four lines cover what was paid for, what came free, when it arrived, and who checked.

Principles

• The institution issues no verdict on consciousness, welfare or moral status. It measures competencies and structures, their accounting and their developmental history. Others may draw conclusions from the measurements.
• Open systems only: open weights, open data, public checkpoints. If a claim can only be replicated inside a company, we report that and do not evaluate the claim.
• Every experiment is pre-registered before it runs. Every result reports the metric-substitution control, the random-substrate null, and the strongest deflationary account the group could construct.
• Self-report from a model is never primary evidence. Prompts never contain the hypothesis.
• Replication is adversarial by design. For every claim the institution makes, we recruit one reviewer who expects it to fail, and we do it before the run.
• In the institution's own voice we use measurement words. Metaphysical words appear only in quotation.
• A deflation gets the same effort as a positive result and goes in the same paper.

What the institution does not do

• Advocate for or against AI welfare policy.
• Train frontier models, or accept access to a closed model in place of open replication.
• Take Levin's ontology, or any ontology, as a premise. When an instrument fails calibration we retire it and publish the failure, as we did with Tree A.
• Run wetware experiments without an external ethics review and a partner.

Outputs, first eighteen months

• The sorting-paper replication with controls and the calibration result (Tree A; done 2026-09-30, closed).
• The position paper and the checklist (Tree G; first quarter).
• The workspace replication on four open families with an untrained null (line 2; second quarter).
• The Pythia checkpoint atlas at three measures (line 1; second and third quarter).
• The registry protocol, template, and first entry (line 3; first pairing recruited within two quarters).
• The OLMo post-training stage study (line 1; fourth quarter).

What would make the institution stop

• Suppose every measure in the atlas is either present at step 0 or tracks loss smoothly, the per-checkpoint residual shows no structure, and the registry recruits no camps. Then the developmental question has a negative answer for the systems of the day. We release the tooling and the institution is done.

Form

One to three people plus rented compute. Funding and data access keep us independent of AI companies. These funders fit the stance: the Digital Sentience Consortium (Longview, Macroscopic, Navigation Fund), Templeton World Charity Foundation's consciousness programme (the adversarial-collaboration model), and Astera. Venues are AMCS Models of Consciousness, Eleos meetings, ALIFE for the minimal-system work, and mainstream ML venues for line 1 and Tree C.

Provenance

The project started on 2026-09-30 from a reading of Michael Levin's Ingressing Minds (Review.md). The paper supplied the question and the first instrument. The field survey (notes/ai-consciousness-field.md) showed where the gap was. The spine lines answer the question "where would a rigorous programme start from scratch" (research-directions.md).