Research directions

Written 2026-09-30. We had two questions. Which of the seven trees in research-plan.md should we approach first, and why? And if we started from scratch, is there a better AI-consciousness programme than the one Levin's paper suggests?

1. Order of the trees

The folders under trees/ are numbered in this order.

1. A, the Ledger. We entered it first and closed it the same day (2026-09-30). The sorting result deflated under controls, and the accounting failed calibration for designed systems. It survives as spine/01 measure M5.
2. G, theory. It is cheap and early. It is also the artefact most likely to be cited by people who never read the rest, because it produces a checklist that any emergent-mind claim has to pass. We wrote it in parallel with A, since its accounting sections constrain A's definitions.
3. B, periscopes. This is A's instrument generalised to a zoo of untrained systems. It is the calibration end of the same axis and is not meant as a destination in itself.
4. C, discovered or built. The strongest second research line. It has a clear path to a mainstream ML venue. It also turns the Sheldrake-versus-Levin discriminator (does a pattern get easier once it has occurred?) into a corpus experiment bounded by date. It needs compute.
5. D, outside the language interface. The attention and the money are here. For that reason we will not enter it until the group has deflated its own results at least once.
6. F, wetware. It needs a partner and an ethics review. Year two.
7. E, patterns as agents. It reads as Sheldrake. Year two or never.

Stage 0 of research-plan.md covers only A and B.

2. Starting from scratch: what would be better

The map in notes/ai-consciousness-field.md shows who held which area as of 2026-09.

• theory-derived indicator checklists: Butlin, Long, Eleos, the TiCS author group
• mechanistic evidence on frontier models: Anthropic (introspection, emotion vectors, global workspace), all on closed weights
• self-report studies: Anthropic system cards, Reciprocal Research, DeepMind
• welfare and moral status philosophy: Sebo, Birch, Chalmers, Schwitzgebel
• expert and public surveys: Cambridge Digital Minds
• neuroscience adversarial collaborations: Templeton ARC, Cogitate
• wetware: Cortical Labs, FinalSpark, Johns Hopkins

Few people work on the three areas below, and the three fit together.

2a. Developmental trajectories on open training checkpoints

Pythia and OLMo publish their full training data and hundreds of intermediate checkpoints. For no consciousness indicator or welfare measure has anyone asked when in training it appeared, what data came before it, or whether it tracks a capability or arrives on its own. Every mechanistic result so far is a snapshot of a finished closed model. This is the class of "developmental evidence" that CMEP and Eleos named in July 2026 and described as "an important but underdeveloped source of insight". The work needs inference and no training, so a rented GPU and the public checkpoints are enough. Levin's ledger is one instrument on this axis, with the residual taken as a function of training step.

A concrete first study would run the Anthropic introspection protocol (concept injection, reimplemented) and the NYU functional-welfare axis at twenty Pythia checkpoints. We would plot detection rate and axis strength against step and against the loss curve, then test whether either arrives with a phase transition and what data the model saw just before it.

2b. Independent replication on open weights

CMEP and Eleos wrote that the field needs research "independent of AI companies". The three most-cited mechanistic results of 2025-2026 (introspection, emotion vectors, global workspace) exist only on Anthropic models, and only Anthropic has checked them. A replication of the first by Lederman and Mahowald (March 2026) has already changed how it is read. Suppose a small group pre-registers each frontier-lab claim and replicates it on open weights against untrained-model nulls. Within a year everyone in the field would be reading its reports. The compute cost is low, and the work builds the record of deflations that Tree D needs.

2c. Adversarial pre-registration for AI consciousness claims

In Cogitate's format two theory camps pre-register their predictions and a neutral team runs the experiment. Each camp commits to its predictions before the data come in. We know of no one running the format for AI systems. A small institution could. It would need a registry, a protocol, two willing camps and open models to run on. IIT says no to digital computers, and GWT-via-interpretability says something is there. The work is institutional, and it is what would give the group's name some weight.

2d. Two further candidates, less certain

• Stream and unity in long-running agents. LLM calls have no persistent state. Eleos's 2026 commentary says the workspace result shows "signs of the unification necessary for a stream without being convinced one exists". A stream would form, if anywhere, in agentic systems with memory and loops, and nobody is studying the question. It is harder to make rigorous. A scoping note would be enough for now.
• Substrate. IIT's claim that digital computers cannot be conscious rests on physical cause-effect structure. Neuromorphic hardware would be the test bed, and nobody uses it for this. It is expensive and depends on a partner, so we park it.

3. Same institution, or a change of focus

We keep the same institution and reframe the spine around one thesis:

developmental and independent evidence for machine minds, on open systems,
pre-registered, with deflation published alongside.

On that thesis 2a is the spine. 2b builds credibility, and 2c is what the institution produces. Levin's ledger (Tree A) is one instrument on the developmental axis. Trees B and C stay where they were, at the calibration and convergence ends of the same axis. Tree D becomes a special case of 2b. Trees E and F stay parked.

The group is therefore founded on the gap the whole field agrees exists, developmental and independent evidence, and Ingressing Minds is where we started. Levin's residual is the first instrument built for that gap because it is the cheapest one to build and calibrate.

Stage 0 stays as written. Experiment 0.1 is still the right first week of work. 2a and 2b run on the harness it produces and on its calibration discipline.

We recommend keeping the Stage 0 plan. The three lines above have their own folders under spine/, planned 2026-09-30 with the same layout as trees/, and PREMISE.md describes the institution built on them.