Tree G: reading notes

Written 2026-10-30 for step 1 of plan.md. There is one section per source. Each gives the citation, what was read, what the source claims and what it does for the position paper. Quotes are verbatim, and "[...]" marks a cut or a symbol the text extraction lost. Page numbers are the printed numbers of the copy named under "Read". Where only an abstract or part of a text was read, the section says so. The last section lists the places where the reading disagrees with plan.md or with draft-accounting.md.

1. Wolpert and Macready (1997)

Wolpert, D. H., Macready, W. G. (1997). No free lunch theorems for optimization. IEEE Transactions on Evolutionary Computation 1(1):67-82. https://doi.org/10.1109/4235.585893

Read: pp. 67-71 (sections I-IV) and p. 77 (conclusions), in journal pagination, from https://www.cs.ubc.ca/~hutter/earg/papers07/00585893.pdf. The proofs and sections V-VII were not read.

Claims.

The theorems "establish that for any algorithm, any elevated performance over one class of problems is offset by performance over another class" (abstract, p. 67).
The setting is black-box search over finite spaces. Performance is counted in distinct evaluations of the cost function. Revisited points are called "artifacts", and the authors "circumvent the problem entirely by comparing algorithms based on the number of distinct function evaluations they have performed" (section II, p. 68).
Theorem 1 (section III, p. 69): summed over all cost functions, the distribution of observed cost values is the same for every algorithm. "What an algorithm gains in performance on one class of problems is necessarily offset by its performance on the remaining problems" (pp. 69-70).
Section III-A (p. 70): "if an algorithm performs better than random search on some class of problems then in must perform worse than random search on the remaining problems" ("in" for "it" in the original). Hence "comparisons reporting the performance of a particular algorithm with a particular parameter setting on a few sample problems are of limited utility".
A uniform prior is sufficient and not necessary. The result also holds for a prior that draws each point's cost independently from one distribution, and for averages over all permutations of a fixed set of cost values (p. 71).
Section IV (p. 71): an algorithm's expected performance is an inner product between the algorithm and the prior over problems. "If knowledge of [the cost function] [...] is not incorporated into [the algorithm], then there are no formal assurances that [it] will be effective." Their example: "years of research into the traveling salesman problem (TSP) have resulted in algorithms aligned with the (implicit) [prior] describing traveling salesman problems of interest to TSP researchers".
Conclusions (p. 77): the theorems "demonstrate the danger of comparing algorithms by their performance on a small sample of problems" and "the importance of incorporating problem-specific knowledge into the behavior of the algorithm".

For Tree G. This is the formal ground for section 2's claim that a residual is always relative to a task family. The inner-product result (p. 71) is the precise form of "whoever chose the bias supplied the fit": an algorithm wins on a family to the extent that someone's knowledge of the family is built into it. That is the programmer term Experiment 0.3 could not price (../01-a-ledger/writeup.md section 4.4).

Two limits matter. First, the theorem counts distinct evaluations. A memoryless random walk revisits points, so a searcher that only remembers where it has been beats it on average over all problems, and the theorem says nothing against that. A competency measured against a revisiting null includes the value of memory (see section 7 on Chis-Ciure and Levin). Second, the zero-sum result holds under the priors named on p. 71. It is a statement about averages over a problem class, and says nothing about which class nature or a benchmark draws from.

2. Marletto (2015), and Deutsch and Marletto (2015)

Marletto, C. (2015). Constructor theory of life. Journal of the Royal Society Interface 12(104):20141226. https://doi.org/10.1098/rsif.2014.1226

Read: sections 1-3 and 5 (pp. 1-20, 25) of the preprint, arXiv:1407.0681v2 (2014-11-04), https://arxiv.org/abs/1407.0681. Section 4, on quantum theory, was not read. Page and section numbers are the preprint's. The journal version was not read.

plan.md step 1 names a different paper, Marletto's "Constructor theory of information" (2015). That paper is Deutsch, D., Marletto, C. (2015). Constructor theory of information. Proc. R. Soc. A 471(2174):20140540, https://doi.org/10.1098/rspa.2014.0540. For it, only the abstract and the first two pages of arXiv:1405.5563 (https://arxiv.org/abs/1405.5563) were read. It defines information through which tasks on physical systems are possible (abstract, p. 1), and nothing read there proposes a cost measure.

Claims of the life paper.

Constructor theory expresses the laws of physics "solely as statements about which transformations are possible, which are impossible and why" (section 2, p. 6). A constructor performs a task and "retains the ability to do so again" (p. 7). A task being possible "means that the laws of physics impose no limit, short of perfection, on how accurately it could be performed" (p. 7).
No-design laws (p. 8): generic resources perform only "elementary tasks", which "contain no design", and "No good approximation to a constructor for tasks that are non-elementary can ever be produced by generic resources acting on generic resources only." Defining no-design laws as typical under some measure is rejected because "the choice of the measure is highly arbitrary" (p. 9).
Section 3.1 (pp. 12-14): under no-design laws an accurate constructor for a non-elementary task must carry a recipe in a modular replicator, with error correction. "The information in the recipe is an abstract constructor that I shall call knowledge" (p. 14). Measures of accuracy come from the subsidiary theory, not from constructor theory (footnote 8, p. 12).
Section 3.1.1 (pp. 14-15): appearance of design is not improbability, because "probability measures are multiplicative" and two unrelated designed objects do not have more design together. A sub-recipe is fine-tuned "if almost any slight change in T′ would cause T to be performed to a much lower accuracy", and "the number of fine-tuned sub-recipes performable by V is a measure of V's appearance of design".
Section 3.3 (pp. 18-20): selection is compatible with no-design laws because variation and selection are "non-specific" to their end products, and the environment is "a crude approximation to a constructor".
Conclusion (p. 25): "all accurate constructors, under such laws, must contain knowledge - a special abstract constructor - in the form of a recipe, instantiated in a replicator."

For Tree G. Section 3 asks whether constructor theory gives a cleaner effort measure than description length. On this paper it does not. "Possible" is defined in the limit of arbitrary accuracy, so the theory sorts tasks into possible and impossible and puts no price on performing one. Accuracy itself is delegated to the subsidiary theory. That answers direction question Q2 toward the short version of section 3.

Three pieces are still worth a paragraph. The no-design condition is the formal counterpart of the random-substrate null: it asks whether generic resources already perform the task, which is what B measures in the ledger. Her conclusion that an accurate constructor must hold knowledge in a recipe matches what Experiment 0.3 found empirically: the unpriced term in a designed system is the programmer's knowledge, and in a trained one the training signal, which plays the role of selection (../01-a-ledger/writeup.md section 4.4). Her appearance-of-design measure, a count of fine-tuned sub-recipes found by perturbing each one, is an operational test (ablate a part, measure the loss of accuracy). It belongs on the competency side of the ledger and is no effort currency. That reading is ours, not hers.

3. Kim (1998), chapter 2

Kim, J. (1998). Mind in a Physical World: An Essay on the Mind-Body Problem and Mental Causation. MIT Press.

Read: chapter 2, "The Many Problems of Mental Causation" (pp. 29-56), in full; the table of contents; from chapter 1, pp. 8-10; from chapter 3, pp. 57-60. Chapter 4 was not read. Source: the scan at https://archive.org/details/kim-jaegwon-mind-in-a-physical-world.-an-essay-on-the-mind-body-problem-and-mental-causation. The MIT Press page returned 403.

Which book and chapter. plan.md names "Mind in a Physical World" chapter 2, and that is the right chapter for causal exclusion. Supervenience, realization and emergence as a group are surveyed in chapter 1 (section "Supervenience, Realization, and Emergence", from p. 4), and Kim defers emergence to his discussion of reduction in chapter 4 (p. 9). Chapter 3 is titled "Mental Causation: The Backlash and Free Lunches".

Claims.

Three problems of mental causation: anomalism, extrinsic content, and causal exclusion (p. 32). The exclusion problem: "Given that p has a physical cause p*, what causal work is left for m to contribute?" (p. 37), and in general form "Given that every physical event that has a cause has a physical cause, how is a mental cause also possible?" (p. 38).
Supervenience defines minimal physicalism, and "emergentism, too, is arguably committed to mind-body supervenience: if two systems are wholly alike physically, we should expect the same mental properties to emerge, or fail to emerge, in each" (p. 38). The thesis is stated on p. 39.
To reject physical causal closure is to reject "the in-principle completability of physics" (p. 40).
"To cause a supervenient property to be instantiated, you must cause its base property (or one of its base properties) to be instantiated" (p. 42, italic in the original).
Higher-level regularities are like the shadows of a moving car: "The M-to-M* and M-to-P* causal relations are only apparent, arising out of a genuine causal process from P to P*" (p. 45).
The dilemma: "If mind-body supervenience fails, mental causation is unintelligible; if it holds, mental causation is again unintelligible" (p. 46). The objection that this would make geology epiphenomenal is answered by reduction, deferred to chapter 4 (pp. 46-47).
The causal inheritance principle: the causal powers of an instance of a second-order property "are identical with (or are a subset of) the causal powers of" its realizer (p. 54). To think otherwise "is tantamount to belief in magic" (pp. 54-55).
"There are no free lunches in philosophy any more than in real life" (p. 30; repeated on p. 60 with "and I believe the cheap ones aren't worth the money").

For Tree G. Section 4 of the draft uses Kim for the claim that a complete physical account leaves an extra cause nothing to explain. Pages 37-40 support that. Kim's own exit is reduction, and the ledger reading follows it closely: to find the missing term is to reduce the residual to a priced physical term.

The base-property principle (p. 42) turns into a test. Any intervention that moves a competency must act on the substrate, so a claim that a pattern acts on its own owes an intervention on the pattern that is not an intervention on the substrate. That is discriminator 2 ("is there a manipulable variable") stated with its burden attached. Causal inheritance (p. 54) is the ledger's conservation rule: a higher-order competency has no causal powers beyond its realizer, so a positive residual points at an unpriced realizer term. The geology objection (p. 46) is the one a proponent will raise. In ledger terms geology has no residual because all its terms are priced. The "free lunch" sentence (p. 30) fits the working title.

4. Dodig-Crnkovic, Cognitive Platonism I and II

I. Dodig-Crnkovic, G. (2025). Platonic Space as Cognitive Construct: Toward a Framework of Cognitive Platonism/Platonic Cognition. Preprints 202512.0033.v1, posted 2025-12-01. https://doi.org/10.20944/preprints202512.0033.v1

Read: the abstract only, from the Crossref record (https://api.crossref.org/works/10.20944/preprints202512.0033.v1). The preprint server returns 403 to scripted requests and to WebFetch. No page references are possible.

II. Dodig-Crnkovic, G. (2026). Cognitive Platonism II: From Natural Regularities to Abstract Forms. Zenodo, 2026-08-28. https://doi.org/10.5281/zenodo.22140895 (also Preprints 202608.2158, https://www.preprints.org/manuscript/202608.2158).

Read: the abstract and sections 7, 10 and 11 (pp. 1, 7-10) of the Zenodo PDF (11 pages), and the reference list. Sections 1-6, 8 and 9 were not read. Page numbers are the PDF's.

Claims of I (abstract). It proposes "a naturalized reinterpretation of Platonism, grounded in the idea that form functions [...] as a constraint within generative processes", in which "abstract structure is real insofar as it organizes system dynamics" and forms "shape perception, morphogenesis, and agency by narrowing the space of viable trajectories".

Claims of II.

It answers "Michael Levin's recent comparison between his account of Platonic Space and Cognitive Platonism" (abstract, p. 1). The comparison is Levin's blog post of 2026-08-26, "Q&A from the internet and recent presentations 5" (https://thoughtforms.life/qa-from-the-internet-and-recent-presentations-5/), per her reference list (p. 11).
"The main difference [...] concerns ontological direction: whether abstract forms independently exist and are instantiated in nature, or whether they are embodied in natural processes and expressed through cognitive abstractions of relational regularities" (abstract, p. 1).
Section 7 (p. 7): constraint and enablement are complementary. Constraints "enable organization precisely by reducing degrees of freedom", with a laser cavity, an enzyme's active site and the genetic code as examples.
Section 7 (p. 8) puts the disagreement as a question: "Does cognition discover an independently existing abstract form instantiated in nature, or does it abstract and articulate relational regularities already embodied in natural processes?"
Section 10 (p. 9): forms "are real in the sense that they have stable structure, constrain reasoning, enable prediction, coordinate action, and can have causal consequences when instantiated through cognitive and physical systems", and "explanatory autonomy does not necessarily entail ontological independence".
Section 11 (p. 10): "Historical evidence by itself does not decide between these ontologies." On the independent realm: "Cognitive Platonism does not claim to disprove it. It argues that such a domain is not required to account for mathematical necessity, mathematical discovery, or the extraordinary explanatory effectiveness of abstract forms."

For Tree G. plan.md section 6 says Cognitive Platonism "drops the realm". II says it is agnostic about the realm and argues it is not needed (p. 10). The paper should state her position that way, since she is the sympathetic critic it goes to first.

II names no measurement that separates her view from Levin's, and says the historical evidence does not decide (p. 10). The two views differ in ontological direction only. A positive residual would therefore not separate them either: both read it as structure in nature that the accounting left out. The paper can say that the ledger is neutral between Cognitive Platonism and Levin, and that its discriminators sort claims, not ontologies.

Her section 7 also bears on plan section 3, which treats "enablements, not just constraints" as content that constructor theory formalises. She treats enabling and constraining as two descriptions of one thing, and constructor theory states its laws as which tasks are possible and impossible (Marletto p. 6), which is the same pair.

5. Birch, "AI Consciousness: A Centrist Manifesto"

Birch, J. (2025). AI Consciousness: A Centrist Manifesto. PsyArXiv, version 1, 2025-09-01. https://doi.org/10.31234/osf.io/af7c9_v1

Read: from version 1 (29 pages, the OSF file behind that DOI), the abstract, the opening pages of sections 1-2 (pp. 1-3) and sections 10-13 (pp. 15-26). Sections 3-9, on Challenge One and on "flickers and shoggoths", were not read. Page numbers are version 1's. A later version is on PhilArchive (https://philarchive.org/rec/BIRACA-4), dated February 2026 in ../../notes/ai-consciousness-field.md. PhilArchive returns 403 to scripted requests and to WebFetch, so the 2026 text was not read. Its abstract, as shown in search results, matches version 1.

Claims.

Two challenges (abstract, p. 1). Challenge One is that users "will soon misattribute human-like consciousness to AI friends, partners, and assistants on the basis of mimicry and role-play". Challenge Two is that "profoundly alien forms of consciousness might genuinely be achieved in AI, but our theoretical understanding of consciousness is too immature to provide confident answers".
The gaming problem (section 10, pp. 16-17). A parity principle between animals and AI is "definitely false" (p. 16). It fails for non-verbal dispositions too: in a points game with stipulated pain, several LLMs found threats of intense pain more motivating than mild pain (p. 16, citing Keeling et al. 2024). Given anti-behaviourism, opportunity and incentive, "they're incentivized and enabled to game our criteria" (p. 17). The problem holds "even for a system that was behaviourally indistinguishable from a real-life human, provided its recreation of human-like behaviour still depended on an extremely rich training corpus" (p. 17). Gamed markers "lose their positive evidential value" (p. 17).
The Janus problem (section 11, pp. 18-21). Evidence of a global workspace in an LLM "would be Janus-faced" (p. 20): "the same evidence can point in opposite directions, depending on one's background beliefs" (p. 21). The result is "two-level deadlock" about the answer and about whether the finding counts as evidence at all (p. 21).
The way out (section 12, pp. 22-25) is a comparative science of animal consciousness that tests specific biological-naturalist hypotheses (endothermy, ephaptic coupling) against the distribution of markers across taxa. "To test a biological naturalist hypothesis, we'd need to test for direct dependencies of phenomenology on biology, unmediated by effects on computation" (p. 22). Repeated failure would be "tilting the odds towards computational functionalism" (p. 23). Figure 4 (p. 25) outlines the programme.
"I don't see a good argument for their in-principle untestability" (p. 26).

For Tree G. The gaming problem is the published ground for two checklist items, no self-report as primary evidence and no prompt that contains the hypothesis. Birch's extension to non-verbal dispositions (p. 16) means the item has to cover any behavioural marker a model could have learned from its corpus. The Janus problem is the argument for pre-committed interpretations: in a registry entry each camp states in advance what each outcome means, which removes the second level of the deadlock for that entry.

plan.md section 6 (and ../../research-plan.md, ../../Review.md, and the notes) describe "Birch's second programme, a biology-independent theory with tests for non-biological systems". Version 1 contains no such programme. Its programme for Challenge Two tests biological naturalism comparatively across animals and waits on interpretability for the AI side. The phrase matches a secondary post (https://theconsciousness.ai/posts/birch-ai-consciousness-centrist-manifesto-2026/, 2026-04-25), which describes a programme for "developing theories of consciousness that are not parasitic on biological implementation details". The 2026 text has to be read before section 6 relies on it. If it matches version 1, the reconciliation is with the method (test a theory by whether its predicted dependency holds across systems that vary on the property) and not with a biology-independent theory. Birch's subject is consciousness throughout. The programme takes his methodological points only.

6. CMEP and Eleos, "Studying AI Welfare Empirically" (July 2026)

Long, R., Sebo, J., Butlin, P., Plunkett, D., Campbell, R., Beasley, C., Saad, B., Sims, T. (2026). Studying AI Welfare Empirically. NYU Center for Mind, Ethics, and Policy and Eleos AI Research, published 2026-07-01. https://nonhumanminds.org/studying-ai-welfare-empirically/

Read: the PDF (https://nonhumanminds.org/wp-content/uploads/2026/07/Studying-AI-Welfare-Empirically.pdf, 92 pages). Read: the abstract and contents (pp. 1-3), Part I sections 3.1.2, 3.2.2 and 3.3 with the Part I conclusion (pp. 16, 19, 21-24), Part II section 1.1.4 (pp. 37-38), and the concluding principles (pp. 57-59). The rest was not read.

Claims.

Evidence comes in three kinds: behavioural, internal and developmental (abstract; Part I section 3, p. 12).
The mismatch problem (Part I section 3.1.2, p. 16): one behaviour may have different causes in an organism and in an AI system. "The gaming problem is a particularly severe version of the mismatch problem" (p. 16).
The specificity problem (section 3.2.2, p. 19): it is hard to tell which features of the human implementation of a property are essential.
Developmental evidence (section 3.3, p. 21) "concerns how an entity came to be the way that it is". It has two kinds. Training evidence (3.3.1a, pp. 21-22) reasons from training pressures, which researchers "can often directly observe [...] such as training data, objectives, and environments" (p. 22). "Trajectory evidence is about how and when particular features emerge over the course of training" (3.3.1b, p. 22). It "largely inherits the difficulties already discussed" (p. 23).
The solution space problem (3.3.2, p. 23): inference from pressure to capability is easy when few solutions exist, and "when many different solutions are available, this is harder to do". AI systems "can solve problems through novel mechanisms" (p. 23).
Uses (3.3.3, p. 24): developmental reasoning can supply negative evidence, with the caveat that "capabilities can emerge as byproducts of the selection process or from other features of architectures, mechanisms, or training data". Comparisons within a substrate "may be stronger" than comparisons across substrates.
Part II section 1.1.4 (pp. 37-38): developmental evidence is "typically more indirect than behavioral or internal evidence" (p. 37), and "LLM training can undercut the evidential value of outputs reporting conscious experience" (p. 37). It is "an important but underdeveloped source of insight" (p. 38) and should not be used "by itself to confirm or deny that a particular AI system is conscious" (p. 38).
Six principles (pp. 57-58): probabilistic, pluralistic, specific, responsible, transparent, independent. The sixth asks for "input from experts independent of the organizations that develop and deploy AI systems" (p. 58).

For Tree G. Plan section 6 says the ledger is an instance of the developmental class. Confirmed, and it can be made exact. The trained-system residual (behaviour bits against training-signal bits, per checkpoint) is training evidence and trajectory evidence at once. The quote in ../../PREMISE.md checks out (p. 38). The solution space problem is the report's name for the inductive-bias point in section 2: from a training pressure to a capability one needs a map of the solutions available, and the random-substrate null and the matched baseline are partial maps. Their byproduct caveat (p. 24) describes what a positive residual would be in their vocabulary: a capability with no training pressure behind it. The report is about welfare. The programme takes its evidence taxonomy and nothing else.

7. Chis-Ciure and Levin (2025)

Chis-Ciure, R., Levin, M. (2025). Cognition all the way down 2.0: neuroscience beyond neurons in the diverse intelligence era. Synthese 206:257. https://doi.org/10.1007/s11229-025-05319-6

Read: the open-access version of record, as deposited by the first author at https://sussex.figshare.com/articles/journal_contribution/Cognition_all_the_way_down_2_0_neuroscience_beyond_neurons_in_the_diverse_intelligence_era/30620957 (Springer blocks scripted requests). Read: the abstract, sections 3-4 (pp. 9-14), 5.2 (pp. 16-17) and 7 (pp. 21-22), and the result paragraph of 6.2 (p. 20). Sections 1, 2, 5.1 and 6.1 were not read. Pages are the article's "Page n of 31".

Claims.

The search efficiency K is "the decimal logarithm of the ratio between the cost of a random walk and that of a biological agent", and "even 'simple' organisms are from two-hundred- to sextillion-fold more efficient" (abstract, p. 1).
A problem space is P = ⟨S, O, C, E, H⟩ with a cost w on operators; a trajectory's cost is the sum of its operator costs (section 4.1, p. 11).
τ_blind is the expected cost of "a maximal-entropy (unbiased) random-walk policy" on the admissible graph, τ_agent the same for the agent's policy, and K = log10(τ_blind/τ_agent) (Eq. 2, pp. 12-13). One unit of K is about 3.32 bits (p. 13).
"K must always be evaluated relative to a well-defined problem space" (p. 13). K adds across conditionally independent stages (p. 13).
"K is only as good as the null model: an unfairly handicapped τblind would overestimate intelligence", and the random walk "must operate within the exact same problem space P, particularly respecting identical admissible sets A and cost functions w as the agent" (pp. 13-14).
Amoeboid chemotaxis: τ_blind is the mean first-passage time of an unbiased diffusive walk over 100 µm, τ_agent about 100 s, K = 2.18-2.30 (section 5.2, p. 16). Changing the first-passage prefactor shifts K by about 0.3 (p. 17).
Planarian head regeneration: τ_blind ≈ 6 × 10^27 s, τ_agent ≈ 3.2 × 10^6 s, K ≈ 21, "about 70 bits"; the random-walk baseline is called "extravagantly conservative" (section 6.2, p. 20).
Specifying the blind walk "is itself an inference problem", and each choice of state space "shifts K by an additive constant" (section 7, p. 21).

For Tree G. K measures competency. It is the C term, and it subtracts nothing that was paid for: the planarian's K ≈ 21 includes everything selection put into the genome. Experiment 0.3's C_ep = log2(p_success/p_random) has the same construction in base 2 with success probability in place of cost (../01-a-ledger/writeup.md section 4.1), so K converts into ledger units on the competency side.

K is defined per problem space (p. 13), which agrees with Wolpert and Macready: a K above zero on one family implies the reverse elsewhere when performance is counted on distinct evaluations (W&M p. 70). Their null is a revisiting random walk, though, and any searcher with memory beats that on average over all problems. Part of every K is therefore the value of not revisiting, and part is the value of following a gradient: the amoeba's null is pure diffusion and its agent is a gradient-follower. Their own caveat (pp. 13-14) is the random-substrate null and matched baseline of the checklist. The checklist asks for the strongest null available, such as a gradient-follower on the same P. They chose the weakest. Whether chemotaxis keeps a positive K against a gradient-following null is not computed in the paper; the point is ours.

Where the reading disagrees with the plan

Birch (section 5 above). Plan section 6 attributes to Birch a "biology-independent theory with tests for non-biological systems". The 2025 text does not contain one; the phrase traces to a secondary blog post. Read the 2026 PhilArchive version before relying on it.
Constructor theory (section 2). It defines possibility in the limit of arbitrary accuracy and prices nothing, so it yields no effort measure. Q2 in direction-questions.md can resolve to the paragraph-length section 3. Plan step 1 names the information paper; the life paper is the one that bears on design.
Enablements (sections 2 and 4). Plan section 3 treats "enablements, not just constraints" as Levin's content with constructor theory as its formalism. Dodig-Crnkovic II (p. 7) treats enabling and constraining as complementary, and constructor theory states laws as possible and impossible tasks. Neither source supports enablement as a separate category.
Cognitive Platonism (section 4). Plan section 6 says it "drops the realm". II is agnostic about the realm and argues it is not needed (p. 10).
The uniform prior (section 1). draft-accounting.md section 2 says "On a uniform prior every system's residual is zero." That holds for performance counted on distinct evaluations. It also holds under the other priors on p. 71, and it fails against a null that revisits points. The sentence should name the counting convention.
Kim (section 3). The draft's use is correct. Kim's own exit from the exclusion argument is reduction (pp. 46, 55-56), which the draft can cite as the same move as pricing the missing term. The "supervenience and emergence" material is in chapter 1, with emergence deferred to chapter 4.
Levin's comparison with Cognitive Platonism, which ../../notes/responses-and-critiques.md records as "not located", is his blog post of 2026-08-26 (section 4 above).