Research plan: pricing the free lunch

Note 2026-09-30. Tree A ran and closed on the first day (see trees/01-a-ledger/writeup.md). The stance in section 0 and the trees still stand as written. Tree A's instrument is reduced to the trained-system form (spine/01), and Tree F is parked. For the current picture read PREMISE.md and research-directions.md. What follows is the original plan, unedited in substance.

Draft 2026-09-30. The plan depends on the verdict in Review.md, which is that Levin's Ingressing Minds wraps a productive research heuristic in an unfalsifiable ontology. We adopt the heuristic and drop the ontology. Levin says one thing is measurable and has never measured it: the delta between the effort put into an embodiment and the competency that comes out. We call that delta the residual, and we target it. The think tank's job is to build the ledger that prices it, in software first, and to publish whatever the outcome.

We first survey the whole space in seven research trees, chosen to be orthogonal, and then lay out a staged programme with a recommended entry point. We expect to run one tree in earnest and keep two as side projects.

0. Stance and ground rules

We are agnostic about ontology. We do not claim that patterns are non-physical, and we do not claim that they are physical. Our claim is that "competency not paid for by training, design or selection" is either a measurable quantity or nothing, and we intend to find out which. If the residual is always zero once inductive bias is priced, the free lunch is bias, and we will have the first rigorous ledger to show it. If the residual is positive under a pre-registered accounting, the result is bigger, and a physicalist has to explain it. Both outcomes are results. That is why this is a research programme and why it is not advocacy.

Metrics come first. Every claim of an "unexpected competency" must survive four checks:

• (a) metric substitution (Schaeffer et al. 2023), where we replace a discontinuous or mismatched metric with a continuous one and see whether the effect survives;
• (b) a matched-effort baseline;
• (c) a random-substrate null. Random networks, reservoirs and random programs already have usable structure, and the Pigozzi GRN and Ertle fractal results can both be explained this way;
• (d) pre-registration of which competencies would count as surprising.

The deflation goes in the same paper. The model is BFF 2026, where Agüera y Arcas's own team published the simpler explanation of their own Computational Life headline. Every positive result we ship comes with the strongest mechanical explanation we could find.

Self-report is never primary evidence. For LLMs trained on the consciousness literature, Birch's gaming problem and Suleyman's "hall of mirrors" objection are correct. We use base models and open weights where we can. Prompts never mention minds, goals, consciousness or the hypothesis, and assays run in channels the model was not trained to narrate. Levin prescribes the same thing himself (§6.5: probe "outside the forced language interface").

Replication is adversarial. Under the Cogitate rule the people who dislike the hypothesis run the replication. We build that in from Stage 1. Dodig-Crnkovic, Jaeger, Sweet, Schaeffer's group and Brbić's group are the natural adversaries (see domain-experts.md).

We watch our vocabulary. "Residual", "unpriced competency", "periscope" and "persuadability" are fine. "Ingression", "Platonic", "soul" and "non-physical" appear only when we quote Levin. The record from 2026 shows the field rewarding mechanistic results and punishing vocabulary.

1. Translating Levin's logic into AI terms

Two rows do not translate. A trained LLM has no analogue of "no evolutionary history", because training is its history. The periscope logic therefore applies cleanly only to untrained or minimally trained systems. And "effort" has no natural unit, so the accounting rule of the ledger is itself an object of research (Tree G).

2. The seven research trees

Each tree is a self-contained line with its own hypotheses, kill conditions and natural collaborators. The trees are orthogonal, in that a result in one does not settle another.

Tree A: The Ledger (free-lunch accounting)

Question. Can "effort in" and "competency out" be put in commensurable units, so that the residual is a number? Once inductive bias is priced, is that number ever positive?

Hypotheses.

• A1 (deflationary): under any accounting that puts the substrate's inductive bias on the effort side, the residual is zero within noise for every system tested.
• A2 (Levin): some minimal systems have a positive residual under a pre-registered accounting.
• A3 (structural): the residual is positive only under accountings that leave out some identifiable term. The deliverable is a catalogue of the omitted terms.

Method. We start with description length (the compressed size of rule, substrate and data) as effort, and task-relevant bits under perturbation as competency. We calibrate on systems where we know the answer: a lookup table, a random reservoir on a task it cannot do, a hand-designed controller. Then the sorting algorithms, the maze agent with the fractal prior, the GRN ODE models and small trained networks all go through the same ledger.

Kill conditions. A2 dies if no system shows a positive residual under the pre-registered accounting. A1 dies if one does and the result survives adversarial replication. The tree dies if no accounting rule can be made to give the right answer on the calibration cases. That would mean "effort" is not a coherent quantity, which is a finding too.

Prior art to beat. The Wolpert–Macready no-free-lunch theorems; Kolmogorov/MDL; Chis-Ciure & Levin's "search efficiency" metric; the literature on scaling laws.

Tree B: Periscopes (systems with no training history)

Question. What competencies appear in untrained or minimally specified systems? Which of them survive metric substitution and a matched random null?

Hypotheses.

• B1: every "unexpected competency" reported in the sorting paper reproduces, and each is explained either by metric mismatch (delayed gratification) or by an efficiency race (clustering).
• B2: pre-registered barrier assays on a zoo of untrained systems (random-init transformers, reservoirs, NCAs, Lenia, BFF soups, cellular automata) find, in at least one system and under continuous metrics, goal-directedness that a straightforward reading of the rules does not predict.
• B3 (Levin's own untested prediction): chimeric arrays of two "species" cluster by algotype even when their efficiency is matched. In that case the efficiency-race explanation is incomplete.

Method. We reimplement the sorting paper with per-cell scheduling under control. The original uses uncontrolled Python threads. We add the Schaeffer substitution, scoring with the algorithm's own local objective, and check whether "delayed gratification" vanishes. We then generalise the barrier assay into a harness that takes any dynamical system, with the criteria for "surprise" pre-registered.

Kill conditions. B2 dies if nothing in the zoo survives the null and the substitution. The periscope logic would then add nothing to the catalogue ALife already has. If B1 dies, meaning the competencies survive substitution, we have the first real signal in the whole programme.

Tree C: Discovered or built (convergence and universality)

Question. Independently trained models converge on the same representational structure. Is that structure discovered, present before training in the data or in the world? Or is it built, an artefact of shared data, shared architecture and scale?

Hypotheses.

• C1 (the Platonic reading, made falsifiable): models trained on disjoint data with disjoint architectures converge on relational structure faster than a null based on data overlap predicts.
• C2 (the Aristotelian reading, Brbić): after calibrating for scale, shared data statistics plus simplicity bias explain the convergence, and nothing is left.
• C3 (the Sheldrake discriminator): convergence between later and earlier models depends on whether the earlier models' outputs are in the later corpus. If it does, the "pattern gets easier once it has occurred" effect comes from the data and is not free of history.

Method. We train small models on disjoint corpora bounded by date: text from before 2022 with no model outputs, and synthetic corpora with controlled structure. Relational convergence is measured with the calibrated tools of 2026 (You et al.; Gröger/Brbić), with vec2vec translatability, and with SAE feature overlap. Data overlap, architecture and scale are varied independently.

Kill conditions. C1 dies if convergence tracks data overlap and nothing else. C3 settles the Sheldrake-vs-Levin question for AI by experiment.

Collaborators. Isola, Cheung, Brbić, Jha and Morris. Of all the trees, this one is the most publishable at mainstream ML venues.

Tree D: Outside the language interface (behavioural agency assays on LLMs)

Question. What kind of goal-directedness, if any, does a language model show in channels it was not trained to narrate? Where does it sit on the persuadability spectrum?

Hypotheses.

• D1: goal-directedness measured in activation space is present, and it can be dissociated from what the model says about itself. The measures are return to a setpoint after perturbation and navigation round barriers in latent tasks.
• D2: for a given model the persuadability ranking is stable across tasks, and it shifts predictably with scale and post-training. The ranking orders the intervention classes (rewiring, gradient, reward, instruction, argument) by how cheaply each changes behaviour.
• D3: once the consciousness vocabulary is taken out of prompts, self-reports about internal state do not correlate with the D1 measures. This is what the gaming problem predicts.

Method. We adapt Zhang & Levin's "Language Game", which freezes the model's dynamics inside an RL policy with linear I/O, and Han/Chalmers/Izmailov's method for the functional-welfare axis. Concept injection is the perturbation and recovery is the readout. We use base models and open weights, and the prompt never mentions the hypothesis.

Kill conditions. D1 dies if next-token statistics under a matched null fully explain the "goal-directedness" in activation space. If D3 dies and self-reports do track internal measures, the welfare field would find that interesting.

Collaborators. Eleos, CMEP, Reciprocal Research, Lindsey's team. Here the think tank meets the model-welfare field, and it must be scrupulous.

Tree E: Patterns as agents (cross-substrate persistence)

Question. Levin's most radical claim is that at the top of the spectrum the agent is the pattern and the body is not. Is there anything measurable there?

Hypotheses.

• E1: some attractor patterns recur across models that share no post-training. Candidates are the "spiritual bliss attractor" documented in Claude system cards, persona attractors and specific failure modes. If they recur, they are properties of the data or of language and belong to no one model.
• E2: such patterns have dynamics of their own. They propagate between models through outputs, and between models and humans through text, with fidelity and mutation rates that can be measured (memetics, done properly).
• E3: a pattern's persistence can be manipulated independently of the substrate. One would steer it out of one model and watch it reappear in another trained on the first model's outputs.

Method. Surveys across models with open weights, controlled chains of distillation, and tracking of specific patterns' lineage through corpora.

Kill conditions. E1 dies if every recurring pattern traces to shared post-training data. E2 and E3 are high-risk and should wait until E1 has been tried.

Warning. This is the tree that reads as Sheldrake. Run it last and quietly, and never describe a result as a pattern "wanting" anything.

Tree F: Wetware comparison

Question. Run the same barrier assay on cultured neurons (Cortical Labs CL1) and on a model with matched task performance. Do they give the same residual?

Hypotheses.

• F1: on matched tasks neurons show a larger unpriced residual than silicon. This is Seth's prediction from biological naturalism, made quantitative.
• F2: once the training signal is matched, the residual is the same (substrate independence).

Method. We would need a partner with a CL1 and a shared assay protocol. The DishBrain Pong and Doom paradigms already have the shape of a barrier assay.

Kill conditions. Either hypothesis can die. Without a partner the tree dies.

Cost. The highest of any tree. Do not start before Tree A has a working ledger, because without one the residual cannot be compared across substrates.

Tree G: Theory (what a positive residual would even mean)

Question. If a residual is ever positive, what follows, and what would a physicalist say?

Work items.

• Formalise the accounting. Relate effort measured as description length to the NFL theorems and to the "possible/impossible transformations" of constructor theory. Constructor theory is the one physics formalism that treats "what a substrate can be made to do" as primary, and Levin does not cite it.
• State Kim's causal-exclusion objection in the terms of the ledger, and say what a positive residual does and does not show about non-physical causation. We give our answer in advance. It shows nothing about that. It shows the accounting left out a term, and the task becomes finding the term.
• Reconcile the programme with Dodig-Crnkovic's Cognitive Platonism and with Birch's centrist programme. (Corrected 2026-10-30: Birch's version 1 has no biology-independent programme; that reading came from a secondary post. See trees/02-g-theory/reading.md.)
• Write the Sheldrake discriminators (history-dependence, manipulable variable, attributed agency) as testable statements for AI, and apply them to every claim we make.

Deliverable. One position paper, written early, so that no reviewer can say we do not know what we are claiming.

3. Staged programme

Stage 0: Reproduce and deflate (months 0–3; one person, a laptop)

Everything else is built on this stage. It is cheap and fast, and either outcome is useful.

Experiment 0.1 does the sorting paper properly. We reimplement Zhang, Goldstein and Levin with controlled scheduling and pre-register three predictions. (a) Delayed gratification appears with zero frozen cells and in classical sorts, as their own tables predict. (b) It vanishes when each cell is scored on its own local objective (the Schaeffer substitution). (c) Clustering strength tracks the efficiency gap between algotypes and vanishes when efficiencies are matched, which extends their own hypothesis. We ship whatever comes out, and either way it introduces the group. If (b) and (c) hold, we have the first published deflation and the tools that produced it. If either fails, we have the first real signal.

Experiment 0.2 gives the fractal prior a proper null. We reproduce Ertle, Levin and Scheutz with three nulls: structured noise with a matched spectrum, patterns that are random but fixed, and shuffled fractals. We pre-register that any input with 1/f structure will do as well as a Halley fractal. If a specific mathematical object beats noise with the same spectrum, that is a signal for Tree A.

Experiment 0.3 builds the calibration set for the ledger. It has five toy systems with known answers (a lookup table, a random reservoir on an impossible task, a hand-coded controller, a trained MLP, a random MLP) and a first accounting rule. The goal is a rule that says "zero" where we know the answer is zero. It makes no scientific claim. It is the instrument.

Deliverables. One or two short papers, a reusable harness for barrier assays, and a draft of the Tree G position paper written in parallel. The gate: if 0.1 and 0.2 both deflate cleanly and 0.3 cannot produce a coherent accounting, we consider stopping. The programme would then have shown the heuristic to be empty in the cleanest cases, and that is still worth publishing.

Stage 1: The periscope zoo and the ledger v1 (months 3–9; two people, modest compute)

• The barrier-assay harness is applied to a pre-registered zoo: random-init transformers (no training), echo-state reservoirs, NCAs, Lenia creatures, BFF program soups, elementary cellular automata and the GRN ODE models. We use continuous metrics only, and we write down before running which competencies would count as surprising.
• Ledger v1 runs over the same zoo. This gives a first map of "which accounting rule makes the residual vanish for which system".
• We send the first invitation for adversarial replication. Sweet will say yes; Jaeger and Dodig-Crnkovic are worth trying.
• Gate: either at least one system has a residual that survives the substitution and the null, or we get a clean negative across the zoo and publish the ledger. In both cases Tree B and Tree A get one paper each.

Stage 2: Discovered or built (months 9–18; needs training compute)

• Tree C in full. Small models with disjoint corpora and disjoint architectures, calibrated relational convergence, vec2vec translatability, and the test on corpora bounded by date for history-dependence (C3).
• Tree A extended to trained models. Does the residual change with training, and how does it relate to the exponent of the scaling law?
• This stage is the most likely to produce a mainstream ML publication. It is also where the Sheldrake-vs-Levin question gets settled for AI.

Stage 3: Outside the language interface (months 12–24; open frontier weights)

• Tree D. Goal-directedness in activation space and the persuadability ranking, on open models. The self-report dissociation (D3) is the result that connects us to the welfare field.
• Present at AMCS Models of Consciousness (the 2026 meeting is 12–16 October in Copenhagen, so aim for 2027), at Eleos and at ALIFE.
• Gate: if D1 survives its null, the group has a measurement that the welfare field wants and lacks.

Stage 4: Wetware and the radical claim (year 2+; partners)

• Tree F with a partner who has a CL1, and only if the ledger works.
• Tree E, restricted to E1 on open weights. It waits until Stages 1–3 have built enough credibility for the group to survive the reaction.

4. The outcome space

The programme will end in one of four places. We decide now what each one means, so that nobody can move the goalposts later.

Some secondary outcomes are worth having whatever happens. They are a reusable harness for barrier assays on arbitrary dynamical systems, a method for ranking the persuadability of models, the first empirical test of history-dependence in representational convergence, and a clear statement of what the class of "developmental evidence" (CMEP/Eleos 2026) can and cannot show.

5. Questions the think tank might work on beyond the trees

We collect these so they are not lost. Some are not worth pursuing.

• Is there a measure of the cognitive light cone for models, the largest spatio-temporal goal a system pursues? Does it scale with anything we can price?
• Do post-training methods shrink or enlarge the residual? (Instruction tuning read as "forcing the language interface".)
• Is the "spiritual bliss attractor" a property of language, of the training distribution, or of one lineage of models? (Tree E1, on a small scale.)
• What is the minimal system with non-trivial persuadability, one whose behaviour changes more cheaply under instruction than under rewiring?
• Can the Sheldrake discriminators become a checklist that any "emergent-mind" claim, from any lab, has to pass? (Tree G. It is cheap, and it may be the most cited thing we write.)
• Does Levin's prediction about planarian chimeras have a software analogue with a real criterion for the outcome, and would a positive result there mean anything?
• Levin writes "we may be fishing in regions of Platonic space we have never explored". Without the ontology this becomes "models may instantiate behavioural patterns not present in any human". Can that be tested? Probably through Tree C, by looking for patterns in model-to-model translation that have no antecedent in human text.

6. Choosing a tree

We recommend entering through Stage 0 and committing to Tree A, with Tree B as its instrument, for five reasons.

1. Of everything in the paper, only this is measurable, publicly wanted by Levin, and not yet done by anyone.
2. It is cheap to start, and it gives a result within a quarter whichever way it goes.
3. It is neutral about ontology by construction. That keeps the group out of the advocacy bucket and lets it work with Levin's lab and with his critics.
4. It builds the instrument every other tree needs, the ledger. Trees C, D and F all compare residuals.
5. In the worst case we get a clean deflation, and that is still a methods paper with an audience.

Tree C is the strongest second choice and has the clearest path to a mainstream ML venue. Start it as soon as there is compute for training. Tree D has the money and the attention, and for that reason we should not enter it before the group has a record of deflating its own results. Tree E is not for year one.

7. What is needed to start

• One person and a laptop for Stage 0. This box can run Experiments 0.1 and 0.3. Experiment 0.2 needs a small GPU for the deep-Q agent.
• A venue for pre-registration (OSF) and a public repo from day one.
• A one-page statement of stance, an edited version of Section 0 of this plan, posted before any result. Then the "why are you doing this" question already has an answer on record.
• Two conversations. One is with Levin's lab, Hazan or Hartl first, since Levin himself has said he wants this done. The other is with a critic. Dodig-Crnkovic is the right first call, because she keeps the programme and rejects the realm, and so do we.
• A name (naming.md).