nucleation

Curation log — September 2026

143 min read

Graph View

47 entries. All books.

Contents

Seeded from the anthology’s out-of-scope clusters

2026-09-25 19:40:46 · curation

The anthology’s reading-time triage (its curation entry of 2026-09-25) set aside 34 works, about 14.4 hours of reading, that no anthology topic could hold. The owner kept that record’s topic decision and made this repository the catchall for what it leaves out. This entry records how the vocabulary was seeded, so the first filings can test it.

The seed

Each of the nine topics is one of the triage’s clusters, renamed to say what the cluster is about rather than how it was found:

topicthe cluster it came fromworksreading time
quantum-foundationsquantum foundations7~2 h
mathematicsthe maths neighbours of that cluster (spectral sets, Betti Lagrangians, topoi)3~40 min
network-sciencenetwork science32.4 h
complex-systemsemergence and information decomposition51.35 h
probabilistic-modelingBayesian nonparametrics21.0 h
neurosciencethe cortical-representations paper115 min
natural-sciencesplanetary, toxicology, concussion3~4 h
operations-researchvehicle routing and facility location131 min
society-and-governancelaw, politics, security82.0 h

What the seed does not settle

  • Filing depth. Nothing is filed yet. A NOTE’s Read and Skimmed statuses already say how deeply a work was read. Each filing batch should say which it did, here.
  • Non-papers. Eight of the 34 are court filings, doctrine, reports or opinion pieces. A LIT can carry a url:. Whether this record wants them is the owner’s call, and it should be made once, in writing.
  • Boundary cases. Four works may fit anthology topics after all: Language Design as Information Renormalization (signal-structure), the cortical-representations paper (concept-geometry), Topos and Stacks of Deep Neural Networks (analysis-and-evaluation) and AgentSociety (deployment-and-society). Each should be placed in one record or the other, never in both.

Seeding the out-of-scope reading, and the order to read it in

2026-09-25 20:42:31 · curation

The seed entry of this afternoon left three questions open, and the owner answered all three: file all of it, including the works that are not papers; seed from abstracts and skims, then triage the result, because the deeper readings still have to happen; and when in doubt, the anthology. This entry records what the seeding did and the order the deeper readings should take.

29 works filed here, 5 to the anthology. 26 carry a Skimmed NOTE, 3 carry none. Every LIT is Deferred (“Unreviewed”), with a status note saying it was seeded and not read in full.

How it was done, and how far to trust it

Four agents each fetched one cluster’s sources and wrote a dossier per work: the bibliography, a paraphrased abstract, skim bullets each naming the section or page it came from, what a deeper reading should check, a priority, and what could and could not be reached. A script turned each dossier into a LIT and, where the work was actually skimmed, a NOTE. The LIT’s “Standing in the record” keeps the priority, the checks and the access record, so the triage below can be audited from the documents rather than from this entry.

Depth was downgraded where the skim was not a skim. The concussion review (LIT-029) and Bourrat on individuality (LIT-005) were reached only through a fetch tool that returns a model-written summary; the publishers served a bot check. A summary of the text is not a skim of it, so both are filed at abstract depth, with no NOTE. So is the wasp-venom proteomics paper (LIT-026), where only the abstract was reachable.

Identities corrected on the way. The network-dimension paper the triage called Unveiling the Dimensionality of Networks of Networks is filed under the title it took in v2 and in Phys. Rev. E 114, L022301. The arXiv hash quant-ph/9601020 is Fuchs’s 1995 PhD thesis. The two unnamed PDFs are Bottke et al. on the Eulalia impact shower and Rungsa et al. 2016 in Toxicon. The hearing transcript (LIT-012) matches the CourtListener docket for D. Minn. 0:26-cv-00603 by case number and judge; the Drive copy has not been checked against a docket entry, a transcript holds no ruling, and whether a written order followed is unverified.

The seven that are not papers. The ISW report, JP 3-12, Drezner’s essay and Gelman’s post carry url: sources. The three court documents carry descriptive titles marked [descriptive title], because a filing has a caption, not a title.

The five that went to the anthology

The seed named four boundary cases, and the owner’s rule sent all four there. A fifth joined them: Rigor with Machine Learning from Field Theory to the Poincaré Conjecture, triaged under “other sciences”, whose transferable claim is methodological — let a learned model search and an exact procedure verify what it finds. That is an analysis-and-evaluation claim, so it is in doubt, so it is the anthology’s. Topos and Stacks went under model-architecture and concept-geometry rather than the analysis-and-evaluation the seed guessed: it is about what an architecture lets information and semantics do, not about how to measure anything.

That leaves neuroscience with no work under it as the primary topic, only as the concussion review’s second tag. The word stays; it was blurbed for a cluster the reading feed keeps producing.

Reading time is not the priority here, and why

The two largest totals in the whole feed are the Eulalia paper (11,500 s) and the network-dimension letter (6,080 s), and their dossiers rank them low and medium. That is not a contradiction: time measures the owner’s attention, and the priority below measures what a close reading would add to this record.

Corrected the same evening. This section first said the long totals meant those works had “probably been read”, and that a deeper-reading pass should begin by asking the owner which NOTEs to mark Read. That misreads the status. A NOTE records a reading this record made — Read is “the full text, closely enough to state its assumptions and results exactly” — and it is earned by doing that reading, not by the owner’s attention in a browser tab. The owner’s reply: the status is an invitation to read the thing. So the deeper readings below are this record’s to perform, in this order, and a Skimmed NOTE is replaced by a Read one when the reading is done.

The order for deeper readings

First — read in full. Short, foundational, and each unlocks others.

  1. Spekkens, Quasi-quantization (LIT-007, 1,615 s). The framing for the two later Spekkens-group papers. Read it before Why interference phenomena… (LIT-019) and the structure theorem (LIT-003), whose skims are enough once this is in hand. §II–III, the technical core, is what the skim missed.
  2. Abramsky & Brandenburger, the sheaf-theoretic structure of non-locality (LIT-016). The core theorems are short. Corrected by the reading (NOTE-016): this entry first called it “this record’s one bridge to the anthology”, with Topos and Stacks building on it and sheaf neural networks descending from it. Neither holds. Belfiore & Bennequin cite it once, in a related-work list, and use nothing from it; the sheaf-network papers checked do not cite it. It is first on its own merits, as the founding and fully proved paper of its programme.
  3. The U.S. v. Google remedies opinion (LIT-006, 1,040 s over four sessions). The skim did not reach the findings of fact on generative AI (pp. 17–46 — the entry first said 17–36): how a federal court understood LLMs, grounding and the GenAI market. That is the part worth reading.
  4. Mediano et al. and Varley & Hoel, side by side (LIT-025, LIT-027, ~2,500 s combined). Two definitions of emergence from the same theme issue. A deeper reading is a comparison, and it should take in Rosas et al. 2020, the primary behind the first. Neither is the anthology’s capability-thresholds sense of “emergence”, and whoever reads them should write that down.

Second — targeted sections, not the whole work.

  • Forman–Ricci curvature (LIT-020): the curvature formula and flow definitions, §2.3 and §4.
  • Fuchs’s thesis (LIT-018): chapters 2–3.
  • Nested HDP (LIT-002): the SVI derivation in §4.
  • The network-dimension letter (LIT-028): the Supplementary Material derivation, after asking whether it was already read.
  • JP 3-12 (LIT-015): the definitions and the authorities chapter.
  • The Rhode Island Hospital subpoena order (LIT-024): whether it was appealed, and how it interacts with the Texas enforcement order.
  • Lindgren & Liukkonen (LIT-001): the linearization assumptions, against Nelson’s stochastic mechanics.

Third — the skim is enough unless an interest makes it otherwise. The remaining sixteen: Eulalia, wasp venom, vehicle routing, the emergence theme-issue introduction, Betti Lagrangians, spectral sets, Ricci-flow community detection, quantum thermodynamics, Pitman–Yor topic models, the two Spekkens-group follow-ups named above, the concussion review, Bourrat, the ISW report, the hearing transcript, Drezner and Gelman. Several of their dossiers say the same thing: the value is in the primary sources the piece cites, not in a second pass over the piece.

None of this promotes anything. A work leaves Deferred when somebody has read it closely, and here nobody has yet.

The first five close readings, and what the seed got wrong

2026-09-25 21:30:54 · curation

The first tier of the reading order in 2026/09/25/204231 is done. The owner’s correction set this pass in motion: Read is not a question for the owner about their browser history. It is a status the record earns by reading the work, and it is an invitation to do that. Five works were read in full, each by one reader working from the full text and the seeded skim. Each skim NOTE became a Read NOTE, at version 2, with a claims table and a section listing what the full text contradicted. All five LITs left Deferred and are now Active.

LITNOTEworkwhat the reading changed
LIT-007NOTE-013Spekkens, Quasi-quantizationThe continuous-variable equivalence is with the quadrature subtheory, which sits strictly inside the Gaussian one, and not with “stabilizer/Gaussian”. Table II’s list of classically explainable phenomena is asserted, not derived.
LIT-016NOTE-016Abramsky & BrandenburgerThe paper is not a bridge to the machine-learning work. The seed and this record’s reading order both said it was.
LIT-006NOTE-004U.S. v. Google remedies opinionThe data remedy is far narrower than the skim said. The GenAI findings run to p. 46. The opinion holds the opposite of “search data is a moat for LLMs” (p. 156).
LIT-025NOTE-010Mediano et al., Greater than the partsIt is a restatement of Rosas et al. 2020 with no new theorem. The summary had inherited the review’s overstatement that it “extends” that paper.
LIT-027NOTE-020Varley & Hoel, Emergence as the conversion of informationThe headline conversion reverses for redundancy-leaning deterministic systems, and the Gaussian effect is 0.005–0.07. There are internal errors: a swapped figure caption, and a reference to a figure that does not exist.

The pattern

Every one of the five skims carried at least one claim the full text does not support. Every one of those errors made the paper look more connected, more general or more conclusive than it is. That is the drift the anthology’s DP-010 warns about: the most citable sentence is often the least supported. It also held on this record’s own most-cited sentence. This record called Abramsky & Brandenburger “this record’s one bridge to the anthology”, with Topos and Stacks building on it and sheaf neural networks descending from it. Belfiore & Bennequin cite it once, in a related-work list, and use nothing from it. The two sheaf-network papers checked do not cite it at all. The claim was corrected in place, in this record’s reading order and in the anthology’s boundary entry, and each correction is marked as one.

The lesson for the seeding step is simple. A skim is honest about its depth, but it is not honest about connections. A connection named from a skim is a guess, and it should be written as one.

What the reading turned up to file

Rosas et al. 2020, Reconciling emergences (PLoS Comput. Biol.), is the primary source behind LIT-025. The review restates its theorem, and every proof and assumption lives there. It is not held. It is the next filing in this cluster, and LIT-025 cannot be cited for the theorem until it is.

Every work read but two, and what the readings did to the seed

2026-09-25 22:10:08 · curation

The reading order of 2026/09/25/204231 is done. 27 of the 29 works now have a Read NOTE. Each is written from the full text, with a claims table and a section listing what the reading contradicted in the seed. The two left over are the Toxicon venom proteomics paper (LIT-026) and the concussion review (LIT-029). Neither is open access. Every route to a full text failed: PMC and Europe PMC, the publishers, OpenAlex, Semantic Scholar and ResearchGate. Both stay Deferred at abstract depth. Reading either needs a PDF supplied by hand.

Where the readings placed the works

standingcountworks
Active22the rest
Proposed3Bourrat (LIT-005), the Eulalia impact shower (LIT-009), Spectral Networks (LIT-017)
Rejected1Lindgren & Liukkonen (LIT-001)
Deferred2the two unobtainable works

Bourrat. The measure it proposes is never computed on any system. Its claim that the measure equals the observational one fails on a two-bit counterexample: take S′ = S ⊕ E with uniform inputs. Then A = 1 bit, and the interventional measure is 0.

The Eulalia paper. It is a plausibility model. Every lunar date scales from one sample-dated crater. Its five fitted breakup ages span 785–965 Myr, not the 800–900 its prose says.

Spectral Networks. It is unrefereed, and several theorem statements are broader than their proofs.

Lindgren & Liukkonen. It puts the imaginary unit in twice by hand: through √g instead of √(−g), and through a complex noise variance chosen to force linearity.

The pattern, held across all 27

Every skim carried at least one claim that the full text does not support. Every one of those errors pointed in the same direction: toward the work being more connected, more general or more conclusive than it is. The first five readings (2026/09/25/213054) found this. Twenty-two more did not break it. The corrections include:

  • Theorems stated more broadly than proved. Ni et al.’s Theorem 4.1. Casals & Nho’s Theorem 3.
  • Printed equations that do not do what the text says.
    • Paisley et al.’s SVI update re-adds the prior at every step.
    • Weber et al.’s workflow silently uses the flow form the paper had just rejected.
    • Grimaldi et al.’s finite-size prefactors have the wrong power of N. A numerical check confirmed this.
  • Headline numbers the paper’s own tables contradict.
    • The Eulalia age range.
    • The waste-routing GA’s “consistently” better routes.
    • The ~30% optimality gaps the skim left out.
  • Attributions that were syntheses. The “Russian military definition” in the ISW report is Clark’s composite, not a quotation.

Several readers also did their own arithmetic, and those checks are part of the record:

  • a comb-network simulation;
  • a linear-programming check of a lemma;
  • a Kesten–Stigum detectability calculation, which puts the stochastic block model’s limit near 0.78, where Ni et al. read collapse as absence;
  • ontic-state counts for the structure theorem, which match the GPT dimension exactly.

Where a reader went beyond the paper, its NOTE says so.

Filing questions the readings raised

  • Rosas et al. 2020 is still the primary source behind LIT-025, and it is not held.
  • Levitsky & Way, The Path to American Authoritarianism (Foreign Affairs, Feb 2025) is the primary source for most of Drezner’s post (LIT-022), and it is not held.
  • JP 3-12 has a December 2022 revision. Secondary reports say it is restricted. If it is ever obtained, LIT-015 becomes Superseded by it.
  • The Rhode Island order (LIT-024) is on appeal in the First Circuit, with argument on 2026-10-07. The Texas enforcement order is on appeal in the Fifth. The record should be re-checked after either rules.

Sixty-five more, seeded, and the order to read them in

2026-09-25 22:45:41 · curation

What the survey was for. The owner asked for a pass over the anthology’s curation record to find readings it had skipped. The purpose was to pick out early-triage candidates for this record: work the anthology set aside because it was out of scope. The survey scanned all 239 anthology curation entries, every ADR, issue #290 and the papers feed.

What it found: 70 candidates.

  • One named decline. The Work Capacity of Channels with Memory (LIT-041) is the only scope decline the anthology named by title that neither record held.
  • Fourteen unassigned works. These have five or more reading sessions and are the clusters the 09/25 triage named only as a group.
  • Fifty-five works below the triage’s ten-minute line. None had been given a disposition.

The owner asked for all 70. Five went to the anthology as boundary works, and the other 65 are filed here. Each is Deferred, seeded from its abstract and a skim, and was filed against the vocabulary as ADR-003 enlarged it.

Depth.

  • 60 carry a Skimmed NOTE.
  • 5 are at abstract depth, because their full text was paywalled or blocked:
    • the Dual EC post
    • The Homological Nature of Entropy
    • Dynamics of collectives with opinionated agents
    • the polycrystalline-water video
    • Kupiainen’s renormalization-group comment

First: read in full, in these pairs

  1. Bourrat’s coarse-graining account of individuality (LIT-049), alongside his Adding causality… (LIT-005, already read). It is the same author’s next move. Corrected by the reading: it is the same author’s earlier move — received July 2022, published August 2023, and cited by LIT-005 — so it cannot repair LIT-005’s defects; it holds the observational idea LIT-005 later mis-equates with its interventional measure.
  2. The Work Capacity of Channels with Memory (LIT-041). This is the anthology’s one named decline. The skim found an information-thermodynamics theory: when actions affect what an agent later perceives, the agents that extract the most energy must trade prediction against forgetting. It carries no ML practice, which supports the decline. It went to Physical Review X in 2026, but the DOI Crossref lists has placeholder metadata, so it was left off until it can be confirmed.
  3. The linguistics triad: Haspelmath on word segmentation (LIT-063), Bruening (LIT-071) and Krauska & Lau (LIT-031) against lexicalism. Haspelmath is filed under its 2011 first appearance, not the 2017 reprint the survey found. It is the one work here with a plausible bearing on tokenization. That bearing is the seed reader’s inference, not the paper’s claim, so the work stays here unless a reading finds otherwise.
  4. The quantum-foundations anchors:
    • the PBR theorem (LIT-062), which the graphical-calculus paper (LIT-054) and the knowledge-and-reality paper (LIT-090) position themselves against;
    • Barandes’s stochastic-quantum correspondence (LIT-086), read against What is quantum in quantum randomness? (LIT-077). Both reach the Born rule through the same unistochastic matrices.

Second: medium priority

  • Information theory:
    • Shannon invariants (LIT-042)
    • the I-MMSE relation (LIT-058)
    • Talagrand Meets Talagrand (LIT-064). The seed reader called it the most ML-relevant work in the batch: it bounds softmax-weighted Gaussian averages. It is a probability paper, and it should be re-checked as a possible anthology boundary work when it is read.
  • Networks:
    • higher-order networks (LIT-091)
    • surface optimization in physical networks (LIT-069)
  • Collective intelligence and cooperation:
    • ants against humans (LIT-075)
    • the computational foundations of collective intelligence (LIT-046)
    • the ergodicity solution (LIT-035)
    • morality as cooperation (LIT-047)
    • grand narratives of inequality (LIT-087)
  • Foundations:
  • Security and influence operations:
    • DISINFOX (LIT-083)
    • the EEAS FIMI report (LIT-084)
    • Fridman on the Russian perspective (LIT-088), which pairs with the held ISW report (LIT-023)

Third: the skim is enough

The remaining 38 are rated low. They include:

  • the blog posts, commentary and news items the survey itself suggested dropping;
  • the court and county records;
  • the heavy mathematics a non-specialist cannot check.

They are filed because the owner asked for all of it. Each note says where its value lies.

Filing notes the seeding raised

  • Vocabulary pressure:
    • philosophy-of-science (structural realism) and a security/cryptography word (DISINFOX, Dual EC) each have one or two works. ADR-003 holds such words back until a second or third work arrives. They now have that.
    • The anthology side recorded two gaps of its own: a scientific-domain-model word and a behavioural-modelling word.
  • The nLab two-level-type-theory page (LIT-040) stands in for a real paper: Annenkov, Capriotti, Kraus & Sattler, arXiv 1705.03307. That paper is the better source.
  • The Law Dork post (LIT-052) should give way to the order it reports: DHS v. D.V.D., No. 24A1153, 3 July 2025.
  • Mean-field-type games with Rosenblatt noise (LIT-050) makes an ML claim (“super-diffusion transformers”) that its pages do not support with any experiment. It must not be cited as evidence.
  • The Kerr County transcript (LIT-078). Its link to the 2025 flood is the seed reader’s inference, not something the document says.

The survey’s first and second tiers, read

2026-09-26 01:32:39 · curation

This pass works through the reading order of 2026/09/25/224541. All ten first-tier works and all 21 medium-priority works now have a Read NOTE, each written from the full text by one reader.

The owner also added philosophy-of-science (ADR-004) before the pass began.

Where the readings placed them

standingworks
Active24
Proposed6: Bourrat’s coarse-graining account (LIT-049), Krauska & Lau (LIT-031), DISINFOX (LIT-083), The Computational Foundations of Collective Intelligence (LIT-046), morality-as-cooperation (LIT-047), surface optimisation (LIT-069)
Rejected1: The Relativity of Causal Knowledge (LIT-038)

The Rejected paper. Its Theorem 1 is unsound as proved and false in general: a mixture of pushforwards is not the pushforward of the averaged map. The reader gives a counterexample and a simulation. Its other two theorems are vacuous.

What the readings changed

Framing this record got wrong.

  • The reading order called Bourrat’s coarse-graining paper his “next move” after LIT-005. It is the earlier paper: LIT-005 cites it. The reading order carries the correction.
  • The seed’s “tokenization” bearing for Haspelmath is extrapolation. The full text says nothing about subword segmentation, so it is not filed.

Theorems stated beyond their proofs.

  • Talagrand’s Theorem 3 constant: its proof gives only cσ/√2.
  • The GPT review defers its headline tensor-product theorems to other papers.
  • Riehl & Shulman’s semantics is explicitly a sketch.
  • Baudot & Bennequin’s quantum case is the opposite of the seed’s claim: there, entropy is a coboundary.
  • The Work Capacity paper prints its capacity formula with the sign flipped from its own proof, and never exhibits the prediction–forgetting trade-off its title names.

Empirical headlines the methods do not carry.

  • Morality-as-cooperation: the test is weakly diagnostic. The blind coder’s κ is .58, and the codebook defines behaviours in virtue terms.
  • The inequality paper: its governance index codes architecture, the outcome it predicts.
  • Surface optimisation: the SI’s trifurcation merge partly makes the trifurcations it counts.
  • The ant–human puzzle: no significance test anywhere.

Candidates for the anthology, flagged and not moved

Four readings say a work may belong in the Anthology of the SOTA. Each is the owner’s call:

  • Consciousness in AI (LIT-056). It assesses properties of AI systems. Suggested for analysis-and-evaluation.
  • NUTS (LIT-085). It carries a real practice. No anthology topic fits; a Bayesian-computation word would also suit SGLD there.
  • The I-MMSE relation (LIT-058). It is the uncited root of the anthology’s “diffusion likelihood is integrated denoising error” accounts.
  • Talagrand Meets Talagrand (LIT-064). A weak case: exact Gaussian-score relations for softmax sharpness against sequence length.

Still to read

The 38 works rated low in the reading order, and the two works still unobtainable: the concussion review and the venom proteomics paper. The low tier’s own notes say the skim is enough for most of them. A pass over them should start by asking which of them anything in either record now cites.

The low tier, read: every reachable work in the record now has a Read NOTE

2026-09-26 01:56:28 · curation

This entry covers the rest of the reading order of 2026/09/25/224541: the 34 works it rated low, where the skim was judged enough. The owner asked for them to be read anyway.

30 were read in full. Four could not be reached:

  • the Dual EC post (LIT-043), which is paywalled;
  • Dynamics of collectives with opinionated agents (LIT-070), which is closed access with no copy anywhere;
  • the AlphaPhoenix video (LIT-074), for which no transcript or write-up is reachable;
  • Kupiainen’s Rigorous renormalization group (LIT-039), which is paywalled.

Those four stay Deferred at their seed depth. So do the two works from the first seeding that were never obtainable: the venom proteomics paper and the concussion review.

Every other work in the record now has a Read NOTE. The record holds 94 LITs:

standingLITs
Active70
Proposed13
Rejected4
Superseded1
Deferred6, all unobtainable

What “the skim is enough” turned out to mean

The low tier was rated low because its seeds said the skim captured the work. In most cases the skim was not enough. Full readings corrected a seed claim in nearly every one:

  • Local separators (LIT-068) is Rejected. Its critique of modularity is factually wrong by its own formula, and its method is never defined.
  • Rosenblatt mean-field games (LIT-050) is Rejected. Its headline, that fBM surrogates give suboptimal control, is false for the linear-quadratic problems it solves. Its “super-diffusion transformer” has no transformer.
  • The Broadview Six transcript (LIT-065). The 42 pages the seed thought were lost were only redacted in one filing and filed whole in another. The reading covers all 77 pages, and the caption is United States v. Rabbitt et al.
  • The Mothership Vortex (LIT-060). Its headline numbers cannot be reproduced from its own replication repository.
  • Causal claims in economics (LIT-079). Between versions, its validation F1 against the same benchmark fell from 0.99 to about 0.12.
  • The Airbnb paper (LIT-094). Its abstract’s “more than 2% misused” is 0.16% in its body.

Superseded, and the rule it tested

Five readers proposed Superseded. The status needs a successor held in the record, so it was applied only where that could be met.

  • The nLab two-level-type-theory page (LIT-040). It is superseded by Annenkov, Capriotti, Kraus & Sattler. That paper was filed for the purpose from its reader’s partial reading. It is Active and still owes a Read NOTE.

  • Four other works stay Active, each with a status note naming the primary source that is not held:

    • the Vance post (the D.C. Circuit docket)
    • the Orosz post (Indeed Hiring Lab’s series)
    • the NEWSCASTER article (iSIGHT’s disclosure)
    • the Law Dork post (the D.V.D. order)

    So are Drezner and Grangier–Auffèves, from earlier passes. Together these are a list of primaries to file. When each is filed, its secondary becomes Superseded.

What is left

  • A full reading of the two-level type theory paper.
  • The six unobtainable works, which need PDFs supplied by hand.
  • The primaries named above.
  • The owner’s decisions on four possible anthology moves:
    • Consciousness in AI
    • NUTS
    • the I-MMSE relation
    • Talagrand Meets Talagrand

The papers-feed swept for philosophy: 119 works seeded

2026-09-26 05:58:20 · curation · record

The owner asked for the papers-feed tracker’s snapshot to be checked for more philosophy to file, after Nagel (LIT-096). The owner also decided that everything from the sweep stays here, and asked for anthology candidates to be flagged with a tag. That tag is anthology-candidate, from ADR-005.

The funnel

stagecount
papers in the snapshot (gh-store-snapshot.json)7,800
scored as possibly philosophy217
kept by hand, with duplicate captures merged132
already held (the SEP entry on structural realism)1
seed dossiers written, by six agents131
dropped as not philosophy3
merged as the same work under another capture9
filed, all Deferred119

95 were skimmed from a full text and carry a Skimmed NOTE. 24 are seeded from the abstract alone.

How the score worked:

  • Journal names: Synthese, Mind, Philosophy Compass, Biology & Philosophy, Minds and Machines and the like.
  • Source URLs: PhilPapers, PhilArchive, PhilSci-Archive, the SEP and JSTOR.
  • Keywords: philosophy words in the title and the first 1,500 characters of the abstract.

Philosophical Transactions of the Royal Society matches the journal filter and is a science journal. Its hits were removed by hand.

The score misses philosophy that carries no signal: an arXiv paper in a physics category, say, that never uses a philosophy word. A second pass over the low scorers would find more. Nobody has made that pass.

Three drops:

  • a political-science paper on propaganda technologies;
  • a mathematics manuscript on Connes’ rigidity conjecture, hiding behind an unresolved PhilPapers code;
  • a JSTOR chapter on Hero of Alexandria’s automata.

Traps

  • PhilPapers, PhilArchive and Taylor & Francis are behind Cloudflare from here. The tracker had captured about twenty of them as “Just a moment…” or as a bare hash, with no title.
    • The agents identified them from the PhilPapers code (an author-surname + title stem) and web search.
    • They read the texts through Wayback copies (web.archive.org/web/2026id_/…), author pages, PhilSci-Archive with a plain curl user agent, and rd.springer.com.
  • Duplicates only show once titles are resolved. Seven of the nine merges were an unresolved PhilPapers capture of a work already in the batch under its journal URL: Wallace, Vetter, Floridi, Drayson, Šekrst, the distributional-semantics chapter, and the epistemic-injustice paper. So deduplicate after the seed, by DOI, arXiv id and title, never only before it. Reading time is pooled across the captures.
  • The dossiers cross-referenced each other by batch id (p078). The filing script copied those into 150 files as-is. They were rewritten to the record’s codes before commit.

The vocabulary

Two words were added by ADR-006:

  • epistemology, used by fourteen works. One had no topic at all, and the rest had the nearest wrong one.
  • logic, used by two.

A philosophy-of-mind word was considered and not added. The mental-causation papers are honestly metaphysics.

Two works are flagged anthology-candidate:

  • Borges and AI (Bottou & Schölkopf);
  • Terziyan et al., on ethics carried across generations of AI.

The borderline cases the dossiers kept here are:

  • Freeborn’s A Model of Understanding in Deep Learning Systems;
  • the three philosophy-of-unsupervised-learning papers;
  • Mallory’s LLMs Are Stochastic Measuring Devices.

Dates

  • Pettigrew’s two “in preparation” drafts (Knowledge is not always more valuable than mere true belief and What is purely epistemic normativity?) carry the tracker’s capture date, 2025-10-29, as an upper bound. The PhilArchive record, which would give the upload date, was unreachable.
  • Weyl carries 1949, the English edition, with no month. Two more have a year only. Each access: line says which.

Reading order

These are the high-priority deeper readings, as the dossiers ranked them:

  1. The AI-consciousness cluster:

    • Birch’s Centrist Manifesto;
    • Seth’s Conscious AI and biological naturalism;
    • Schwitzgebel’s AI and Consciousness;
    • Goldstein & Lederman’s What Does ChatGPT Want?.

    These sit beside Butlin et al. (LIT-056) and Nagel (LIT-096).

  2. Structural realism: Wallace’s Stating Structural Realism, Morganti, and Chakravartty. They sit beside the SEP entry (LIT-045).

  3. Philosophy of biology:

    • Baedke et al. on reciprocal causation, to be read with Hazelwood’s reply;
    • the SEP’s Biological Individuals.

  4. Evidence: Douglas on inductive risk, and Hartmann & Trpin on coherence-based explanatory power.

A pass over the mereology tag: four readings, eight retags

2026-09-26 06:37:03 · curation · record

The owner asked for a pass over the mereology tag. It held six works: two read and four skimmed seeds from the morning’s philosophy sweep. All four seeds were read in full, and each is now Active:

workwhat the reading found
Varzi, SEP Mereology (LIT-124)the formal reference: M ⊂ MM ⊂ EM ⊂ GEM, with the exact axioms. GEM is reached from M by exactly three pairings of a supplementation principle and a sum principle.
Carroll, Reality as a Vector in Hilbert Space (LIT-123)a programme, not a result: parts are factorisations chosen afterwards, to suit the dynamics
Register, The Depth of the Body (LIT-138)a moral “body part” relation, bounded below by depth, applied to BCIs hosted in the cloud
Meincke, One or two? (LIT-142)process ontology keeps the foster as a part, and makes the count indeterminate

What the readings corrected

Each skim overstated its paper, in the same direction: a proposal was reported as a result.

  • The SEP entry was summarised as calling the partial-order core “uncontroversial”. §2.1 says the opposite: all three core axioms have serious dissenters, and Antisymmetry “can hardly be regarded as constitutive”.
  • Carroll was credited with a uniqueness result for the quasi-classical factorisation. He does not claim one. The only uniqueness claim is for local factorisations, and it is borrowed from Cotler et al. Carroll’s own criterion needs inputs beyond the spectrum, which he admits are “somewhat cheating”. Fields were reported as “recovered”. The essay only hopes to recover them.
  • Carroll’s finiteness figure, e^(e^123), is what the paper prints (p. 5). It does not follow from the paper’s own premise, S = πr²/G, which gives a single exponential. Quote the order of magnitude, never the figure.
  • Register was summarised as resting his ethics on the real-definition metaphysics and dissolving the ownership conflict. He says the principles do not depend on that ontology, and the conflict is narrowed, not dissolved.
  • Meincke was summarised as offering a rival to the Parthood View. She keeps parthood and rejects only its pairing with substance ontology.

Eight works that were about parts and wholes and did not say so

The tag’s blurb reads “parts and wholes — composition, individuation of systems and collectives”. The test was whether someone browsing that topic would expect the work. Eight held works pass, and now carry the tag after their existing topics:

  • Composition of wholes: LIT-025 (ΦID causal emergence) and LIT-027 (emergence as the conversion of information).
  • Constitution: LIT-097 (The Extended Mind).
  • Individuation:
    • LIT-157 (Bourrat, reproducees and reproducers)
    • LIT-167 (Baedke, The Organism)
    • LIT-168 (SEP, Biological Individuals)
    • LIT-188 (Griesemer & Shavit, scaffolding individuality)
  • A collective as a candidate subject: LIT-159 (Schwitzgebel, the United States as conscious).

Considered and not tagged: Philosophy of Microbiology, collective intelligence and the ant/human puzzle study. They discuss collectives, but not as a question about parts and wholes.

The tag is wider than its reference

The SEP entry excludes two senses of “mereology”:

  • the systems-theory and ecological sense (its note 1);
  • quantum part–whole formalisms (its note 18).

The tag’s blurb names both on purpose: “individuation of systems and collectives”. So LIT-124 is the background for the philosophical works on the tag (LIT-142, LIT-138 and the individuality cluster). It is only loosely the background for LIT-123, LIT-025 and LIT-027, which use “part” for a tensor factor or a coarse-grained variable. A reader should not expect the SEP’s axioms to hold of those.

Still open on the tag

  • The two Bourrat papers (LIT-005, LIT-049) remain Proposed. Each would be settled by a worked model: an explicit computation of its individuality measure on a model of an evolutionary transition. Another reading cannot supply that.
  • Six of the newly tagged works are still Deferred: the four individuality works, The Extended Mind and Schwitzgebel. They are the natural next readings for this tag.

The mereology tag’s newly tagged works, read

2026-09-26 06:50:21 · curation · record

The earlier pass over the tag (the previous entry) added eight works whose subject is parts and wholes. Two of them, LIT-025 and LIT-027, had already been read. The other six are now read in full:

workstatuswhat the reading found
SEP, Biological Individuals (LIT-168)ActiveWilson & Barker’s own nine-region framework: organism / some parts / some groups, crossed with living agent / Darwinian individual. It never defines “part” and gives no rule for which parts or groups count.
Bourrat, Reproducees, reproducers, and Darwinian individuals (LIT-157)Activenested statuses, with material overlap necessary for reproducers but not reproducees. Its measures are recipes, never computed.
Baedke, The Organism (LIT-167)Activethe clearest statement of persisters vs overcomers. Separation from the environment is LIT-110’s model, re-presented.
Griesemer & Shavit, Scaffolding individuality (LIT-188)Proposeda sketch. Its non-circularity is one sentence in the abstract, and its third stage presupposes the group.
Schwitzgebel, If materialism is true… (LIT-159)Activea whole of conscious parts can be conscious; it rejects anti-nesting, including IIT’s exclusion postulate
Clark & Chalmers, The Extended Mind (LIT-097)Activethe step from coupling to parthood rests on assumed functionalism and a proposal about the word “belief”. “Part” is never given a mereological sense.

What the readings corrected

The skims overstated in the same direction as the first pass found: a proposal was reported as a result, or a sufficient condition as a list of paradigm features.

  • Clark & Chalmers’ conditions on a coupled resource were summarised as sufficient (“provided…”). The paper says it gives no “categorical answers”.
  • Schwitzgebel’s §6 does not test the Dretske, Dennett, Humphrey and Tononi theories. It rebuts four objections, Chalmers’s the most seriously.
  • Bourrat’s dossier had reproductive causal control evolving through scaffolding. The paper (p. 21) says scaffolded collectives get control for free, and that what must evolve is autonomy.
  • The SEP entry’s framework was credited to Pepper & Herron, who only pose the question.

Where the works disagree

  • Genes. Bourrat (LIT-157) makes genes, like hearts, mere reproducees. The SEP entry (LIT-168) places genes as Darwinian individuals. The record holds both. Neither is retired.
  • Bourrat against his own earlier papers (LIT-005, LIT-049):
    • Inherited: nothing is computed; the environment is handled by conditioning; and the control measure is biased toward fine grains.
    • Fixed: autonomy (associational) and control (interventional) are kept apart.
    • Scale rule: “smallest object at maximal MI” runs opposite to LIT-049’s coarsest grain.

A defect found and fixed along the way

The script that writes a Read NOTE rewrites the LIT’s summary. It kept a citation head only when the head ended in ”).” or ”>.”. For “Clark et al. (1998), DOI-10.1093/analys/58.1.7.” it fell back to splitting at the first ”. ”, and kept only “Clark et al.”. That dropped the year and the identifier.

All ten were restored from each document’s own fields. The script now matches “Author (YEAR), identifier.” first.

  • Check that finds it: any summary whose first lines carry no “(YYYY” lost its head.
  • Left alone: LIT-074, whose unverified head is deliberate.

Still open on the tag

  • Three works are Proposed: LIT-005, LIT-049 and now LIT-188. Each waits on a worked model, not on another reading.
  • The rest of the tag is read.

A pass over the agency tag: sixteen readings, two retags

2026-09-26 14:21:08 · curation · record

The owner asked for a pass over the agency tag, after the one over mereology. The tag held 18 works:

  • 3 already read: LIT-041, and LIT-167 and LIT-168 from the mereology pass.
  • 12 skimmed seeds.
  • 3 seeded from the abstract only.

Fifteen were sent for close reading. So was LIT-211, newly tagged. Two could not be read.

workstatuswhat the reading found
Quillien, Luck and Intentional Action (LIT-132)Activea verbal causal account with no formal model, as the skim had implied it had. Causal judgment carries about half the luck effect, and every mediation analysis was exploratory.
Pernu, Mental causation via neuroprosthetics? (LIT-147)Activefootnote 1 limits the argument to difference-making with proportionality
Tomasello, The moral psychology of obligation (LIT-125)Activeopen after all (the MPI-EVA author copy). Joint commitment, not collaboration, carries the load. The Response retreats to explaining the sense of obligation.
Pettigrew, Consent and the Formation of Preferences (LIT-202)Activeone criterion decides every case: the room each mechanism leaves for free agency. Enoch 2020 is rejected, not “examined next”.
Bruckner, Welfare Subjects and Autopoiesis (LIT-120)Active”autopoiesis” is really adaptive self-maintenance. His key case, the self-maintaining car, produces none of its parts.
SEP, The Moral Status of Animals (LIT-146)Activeevery account of agency it surveys depends on experience, so none transfers to AI without further argument
Nahas & Sachs, naturalizing teleology (LIT-192)Activeread from the accepted manuscript. McShea’s field theory, which the skim left out, is a main target.
Goldstein & Lederman, What Does ChatGPT Want? (LIT-206)Activethe abstract’s “strong case” is argued only “presumptively” in the body. Its primary topic moved to agency.
Roberts, Talkative AI (LIT-207)Activethe anti-realism is narrow: only states that depend on a body. Fictionalism is permitted, not required.
Walsh, Kaepernick as Confucian Sage (LIT-209)Active”Confucian incivility” is Kim’s 2011 concept. Walsh’s contribution is extending it to the body.
Safron et al., the world models editorial (LIT-195)Activea map of the theme issue, 17 contributions, not the 12 the skim listed. Its three senses of “world model” are conditional readings, not a criterion.
Bender et al., What Lives? (LIT-211)Activeworth reading as a dataset. Its robustness figures are unadjusted: adjusted Rand is 0.22–0.38 and silhouette 0.079. The “continuous landscape” is built into the method.
Elshatlawy, Rickles & Arsiwalla, observers (LIT-133)Rejectedits one “theorem” is trivial, its complexity bound is misproved, and its quantum claims are asserted
Terziyan et al., ethical evolution (LIT-182)Rejectedsays it “demonstrates” what it never implements, and its central principle contradicts itself (pp. 9, 13, 18)
Enoch, Agency, Shmagency (LIT-117)Deferrednot read: no full text reachable. A Wayback snapshot of an old Semantic Scholar PDF exists; the status note gives its address.
Sharman, AI and free will (LIT-197)Deferrednot read: closed access, and no copy anywhere

The agency tag held on every work read.

What changed besides the readings

  • anthology-candidate removed from LIT-182. The seed flagged it as ML practice. The reading found no method, data or empirical claim in it.

  • Two works newly tagged agency:

    • LIT-046, Pilgrim et al. on collective intelligence: a collective that senses, decides and acts.
    • LIT-211, What Lives?, on the authors’ own manual coding: 17 of its 68 definitions are coded autonomy, agency or goal-directedness.

    My first rationale for LIT-211 counted its “Cognitive Autonomy” cluster as two thirds of the definitions. That fraction included the thermodynamic cluster, and the reading caught it. The rationale was corrected before commit.

  • Considered and not tagged:

    • Bourrat’s reproducers (LIT-157): the autonomy there is reproductive.
    • Seth (LIT-135) and Vieira (LIT-126): consciousness and moral patiency.
    • The ant puzzle study (LIT-075): collective memory.
    • The reciprocal-causation pair (LIT-110, LIT-156): causal structure.

For the next pass

  • Two held documents misdescribe the editorial. LIT-131 (Freeborn) and NOTE-123 call the editorial’s three senses of “world model” a “three-tier criterion”. It is not one, and both should be corrected when LIT-131 is read.
  • The ethics of obligation now has a thread: Tomasello (LIT-125) grounds obligation’s force in Korsgaard 1996, and Enoch (LIT-117) is the standard attack on that grounding. Reading Enoch would join them.
  • Reading notes record local download paths. Many of this record’s notes name the scratch path a download was saved to (raw4/…, scratchpad/…), 116 files on main alone. Those paths mean nothing outside the session that made them. Cleaning them up is a separate change.

A pass over the metaphysics tag: thirty-two readings, two new topics

2026-09-26 19:01:09 · curation · record

The owner asked for a pass over metaphysics, after mereology and agency. The tag held 45 works: 16 read, 27 skimmed seeds and 2 seeded from the abstract. The 29 unread went to seven close readers, grouped by cluster. The coverage audit added four works to the tag, and the three of them that were unread went to an eighth reader.

Outcome

  • Read in full: 30.
    • Active: 27.
    • Proposed: 1. Allori, Wave-functionalism (LIT-200). Its claim to bypass PBR is asserted in one footnote. It is also inconsistent about whether ψ is nomological.
    • Rejected: 2.
      • Karpenko (LIT-215). It misreads Floridi three ways. It says an AdS/CFT boundary theory is two-dimensional, and that a quantum field operator obeys the Schrödinger equation. Its abstract’s “it is proven” has no argument behind it.
      • The pregeometry paper by Arsiwalla, Elshatlawy & Rickles (LIT-102). Nothing in it is constructed; the “link” it draws is a summary of other programmes. The same group’s observers paper (LIT-133) was Rejected in the agency pass.
  • Not obtainable: 2. Pernu 2015 (LIT-144) and Carson’s Element (LIT-104). Both stay Deferred, with the attempts in their status notes. Pernu’s open 2013 dissertation (HELDA) covers the same argument at an earlier stage.

The skims again overstated in the same direction: a programme or proposal was reported as a result. The most consequential cases:

  • Chakravartty (LIT-196) says outright that nothing in the chapter tells against eliminative OSR. The skim had it collapsing the view into eliminativism.
  • Gao’s 194-page monograph (LIT-121) gives three routes to ψ-reality. All three rest on one premise that his own fn. 10 undercuts.
  • Holik et al. (LIT-107) do not derive the Indistinguishability Principle. They move it into a postulate, and two printed proofs contain errors.
  • Barrett et al. on IIT (LIT-185) concede that IIT 4.0 does say “high Φ = more consciousness”. Their multi-dimensional measure is a proposed revision of the theory.
  • The SEP Personal Identity and Ethics (LIT-130) shows the metaphysics constraining the ethics but rarely settling it. The current entry, Dufner 2025, also replaced Shoemaker’s at the same URL.
  • De Haro & Butterfield (LIT-199) say AdS/CFT is an exception only to duality being compatible solely with epistemic emergence. It is not an exception to “dualities are not indicative of emergence”.
  • Hazelwood (LIT-156) targets Laland and colleagues, not LIT-110. LIT-110’s own standing note, which said otherwise, is corrected.
  • Rizi (LIT-150) and McKenzie (LIT-141) both lean on LLM “emergent abilities”.
    • Rizi does so in one unhedged sentence. Its example is arithmetic, the case Schaeffer et al. (ANTH-LIT-471) show to be a metric artefact.
    • McKenzie examines the question (§21.2), and still calls Wei et al. (ANTH-LIT-470) “concrete and rigorous”.

Tags

  • Two topics added (ADR-007):

    • philosophy-of-mathematics (philosophy group): Carson, Weyl, Leitgeb & Ladyman, Corfield.
    • religion: the SEP Religion and Science entry, where it is primary, and the epistemic-injustice paper.

  • Primary topic moved to metaphysics, on each reader’s recommendation:

    Dorato’s reader called either defensible, so it is left.

  • logic added to Holik et al. (LIT-107).

  • Four works newly tagged metaphysics:

    Considered and not tagged:

    • Baedke et al. (LIT-110): its boundary is epistemic.
    • Wigner (LIT-180): physics, not a philosophical question.
    • Quillien (LIT-132): an empirical study of causal judgment.
    • Bourrat’s coarse-graining paper (LIT-049): borderline.

Process

  • Every applied reading is checked for local paths (the cleanup of #23). Four new phrasings reached the record from this pass and were caught.
  • A reader sent the owner’s email address to Unpaywall’s API. ADR-008 now forbids it, and every brief carries the rule.
  • Holik et al.’s NOTE used [[x]]_z notation, which luria reads as wiki links: 284 lint violations. It is now in code spans.

A pass over the identity tag: five readings, five retags, five summaries repaired

2026-09-26 19:20:31 · curation · record

The owner asked for a pass over identity after the metaphysics pass. Its blurb: “what makes something the same thing over time or across descriptions — individuality, persistence, personal identity”. The tag held 18 works. 13 were already read, eleven of them by the mereology, agency and metaphysics passes. The coverage audit added five, one of them unread.

Readings

workstatuswhat the reading found
Lloyd, SEP Units and Levels of Selection (LIT-160)Activeread it as one side’s case, not a survey. It is Lloyd’s own four-role framework, in the first person. It criticises Damuth & Heisler (LIT-201), Bourrat’s individuality papers, and applying the transitions model to holobionts.
O’Malley & Parke, SEP Philosophy of Microbiology (LIT-210)Activemore cautious than its skim. Host–microbiome units of selection are “still disputed”, and claims about the self are “vague”.
Corfield, Duality as a category-theoretic concept (LIT-115)Activethat predicate logic is ill-suited to sameness questions is expected, not argued. The positive programme is only announced. Its primary topic moved to philosophy-of-mathematics.
Browning & Veit (LIT-116)Activeonly as an illustration of the relational self. Its printed title is The Desperate Search for Meaning in Life, and the LIT title is corrected.
Archer & Robb, Ethics of Parasocial Relationships (LIT-172)Activeargues fans’ duties only, despite promising duties for both sides. Its wrong presupposes a person on the far side, so it does not transfer to AI companions.
Robb, Archer & Dennis (eds.), Philosophy of Fame and Celebrity (LIT-137)Deferrednot read: closed access

Tags

Five works newly tagged identity:

  • LIT-188 (Scaffolding individuality), on individuality.
  • LIT-045 (SEP Structural Realism), because quantum non-individuality motivates ontic structural realism.
  • The three duality works (LIT-115, LIT-199, LIT-113). Each asks when two formulations describe the same world, which is identity across descriptions.

Two readers flagged a question about the tag’s reach:

  • Archer & Robb concerns narrative or practical personal identity. Fans who freeze a celebrity’s persona are the subject.
  • Browning & Veit concerns individuality as a boundary with others.

The blurb’s “personal identity” covers both. Neither concerns numerical identity. The tags stay, and this entry records the reading so that a later narrowing of the word has the cases in front of it.

A defect found and repaired

The audit found LIT-045’s summary saying its first sentence twice. The cause was the first version of the script that applies readings. It took a summary’s citation to end at the first ”).” in the text, so part of the seed summary could be kept in front of the reading’s summary.

  • Checked: all 142 summaries written from full readings, against their original seed text in git, following each file back through its renaming from a temporary code.

  • Affected: five.

    Six others share wording with their seed deliberately, and were left alone.

  • Repair: in each of the five, exactly the text before the reading’s own summary (as its NOTE gives it) is removed. Each LIT’s history records the repair.

  • Script: fixed earlier, in the mereology pass, when the related “et al.” defect was found.

A pass over the consciousness tag: eleven readings, and The Weirdness of the World

2026-09-26 19:58:08 · curation · record

The owner asked for a pass over consciousness after identity. The tag held 21 works: 10 read, 10 skimmed seeds and 1 seeded from its abstract. All 11 unread were read in full, by four readers. Schwitzgebel’s AI and Consciousness had a reader of its own, since it is book-length.

Midway through, the owner asked for Schwitzgebel’s The Weirdness of the World (Princeton, 2024). It is filed as LIT-216 and read in full from the author’s final manuscript (June 2022, 349 pp.).

Readings

  • Active: 8.
    • Birch’s Centrist Manifesto (LIT-111). It now also carries its PsyArXiv DOI, which the reading found.
    • Seth on biological naturalism (LIT-135). Read from his OSF accepted manuscript; the BBS Author’s Response was reached only as an abstract.
    • Schwitzgebel’s AI and Consciousness (LIT-191).
    • Schwitzgebel & Sinnott-Armstrong’s review (LIT-166).
    • Frankish (LIT-103).
    • Kleiner & Ludwig (LIT-143).
    • Veit (LIT-105), now reached in full.
    • The Weirdness of the World (LIT-216).
  • Proposed: 2. Winkielman, Ziembowicz & Nowak (LIT-190); Terrill & Veit (LIT-193).
  • Rejected: 2.
    • Vieira (LIT-126). The abstract promises engagement it never delivers, and its venue is unverified.
    • Šekrst (LIT-212). Its “WikiBERT” was an extractive QA model asked a different question, and its GPT tests report no settings.

What the pass established

  • Birch and Seth answer each other. Birch’s v4 names Seth, and its §12 overlaps his BBS commentary on Seth. They disagree on four things:

    • whether the illusion that an AI is conscious can be blocked at the level of belief (Birch: yes; Seth: perhaps not, like Müller-Lyer);
    • how likely consciousness is in current LLMs;
    • whether a link between consciousness and life is testable;
    • what a global workspace found in an LLM would show.

    They agree on Nagel’s definition (LIT-096) and on treating LLMs as role-players.

  • Schwitzgebel’s arguments are layered. The Weirdness of the World collects the earlier arguments. LIT-159 is its chapter 2 in manuscript and chapter 3 in print, revised: the IIT footnote is much expanded and now engages the expander-graph case that LIT-185 defends against. AI and Consciousness (LIT-191) continues its chapters 10 and 11. LIT-159 now says so.

  • Kleiner & Ludwig’s published version drops a result. The finding that metric spaces are not structures of experience is only in arXiv v1 §4. Cite the preprint for it.

  • Veit now rejects plant consciousness. That conflicts with the paper he co-wrote, LIT-193.

Tags

  • Coverage: the audit found no untagged work whose subject is consciousness. The ΦID review (LIT-025) and Borges and AI (LIT-178) mention it in passing.

  • Secondary tags added on the readers’ recommendation:

    Weaker suggestions were not taken.

  • Citation fix: a bare ADR-009 in a NOTE, meaning the anthology’s, is now ANTH-ADR-009.

Ladyman and Ross on the record: ontic structural realism and rainforest realism

2026-09-26 20:23:45 · curation · record

The owner asked for James Ladyman and Don Ross’s accounts of ontic structural realism (OSR) and rainforest realism to be on the record, ideally from their book Every Thing Must Go: Metaphysics Naturalized (OUP 2007, with Spurrett and Collier).

What could not be obtained

Only legitimate sources were used: publishers, repositories, the authors’ own pages and the Wayback Machine. Shadow libraries were ruled out.

  • Every Thing Must Go. Not open access, and OUP returned 403. No free sample was found. An archive.org upload labelled “Public Domain” is an unauthorised copy and was not used.
  • Ladyman 1998, “What is structural realism?” (SHPS), the founding OSR paper. Paywalled, with no repository copy.
  • Ross 2000, “Rainforest Realism: a Dennettian theory of existence” (in Dennett’s Philosophy, MIT). Closed access.
  • Berenstain & Ladyman 2012, “Ontic Structural Realism and Modality”. On PhilArchive, which blocked this session, as did the Wayback Machine.

What was read instead

Four freely available works, each read in full and Active:

workwhat it contributes
Ladyman 2001, Science, Metaphysics and Structural Realism (LIT-217)the earliest free statement of OSR by its founder
Ladyman, Patterns All the Way Up (2024 preprint, MIT 2026) (LIT-219)the core of rainforest realism, and his current second thoughts
Ross 2008, Ontic Structural Realism and Economics (LIT-218)OSR carried beyond physics
Dennett 1991, Real Patterns (LIT-220)the criterion that rainforest realism generalises. Read because Ross 2000 was unreachable.

What each supplies:

  • Ladyman 2001:
    • the quantum non-individuality argument, framed as an underdetermination between individuals and non-individuals, not a proof;
    • entanglement against Humean supervenience;
    • theories describing “the modal structure of the phenomena”;
    • a non-eliminative line: “the phenomena have structure but they are not structure”.
  • Patterns All the Way Up:
    • “to be is to be a real pattern”;
    • scale-relative, effective ontology;
    • no single hierarchy of levels;
    • agnosticism, not denial, about a fundamental level (fn. 9);
    • the Principle of Naturalistic Closure quoted from the book (p. 37), then called “problematic” by Ladyman himself.
  • Ross 2008: individuals are “book-keeping devices”, not denied. That is a deflationary third position, which Chakravartty’s dilemma (LIT-196) does not address.
  • Dennett 1991: a pattern is real iff some description beats the bit map, “whether or not anyone can concoct it”. Rival real patterns can coexist.

How far the two ideas are covered

  • OSR: covered by the new works together with LIT-045 (the SEP entry, by Ladyman) and LIT-196. The explicit eliminative/non-eliminative choice still rests on those two held notes: no new text by the authors stakes it out.
  • Rainforest realism: covered for real patterns, scale relativity, the rejection of a fundamental level of individuals, and the PNC.

Gaps:

  • The Primacy of Physics Constraint is stated in none of the four.
  • The book’s own real-patterns definition (Ross 2000 and Every Thing Must Go), which adds projectibility to Dennett’s compression, is referred to but not quoted.

The book, Ross 2000, or Berenstain & Ladyman would close both. Berenstain & Ladyman is the one to try first, from a session that can reach PhilArchive.

Corrections the readings raise

These are recorded here, not written into the older notes, because each is about what the new sources say.

  • NOTE-038 and NOTE-129 credit the projectibility and perspective clauses to “real patterns” in general. Dennett’s own criterion is plain lossless compression.
  • NOTE-129 has Ladyman and Ross “deny” a fundamental level. Patterns All the Way Up fn. 9 says they are agnostic about one. That may be a later retreat, or a misreading; the book would settle which.
  • The record’s two accounts of real patterns disagree on whether the fundamental level is itself a pattern: Wallace (LIT-151) against Ladyman & Lorenzetti (LIT-184). Patterns All the Way Up states the disagreement.

The Yoneda lemma, and whether it supports ontic structural realism

2026-09-26 20:52:02 · curation · record

The owner asked for an account of the Yoneda lemma, connected if possible to ontic structural realism (OSR). Three works were filed and read in full, all Active:

workwhy
nLab, Yoneda lemma (LIT-221)the lemma stated and proved, with the embedding as fully faithful and the corollary y(c) ≅ y(d) ⇔ c ≅ d, in a self-contained unit that can be read whole
Bain 2013, Category-theoretic structure and radical ontic structural realism (LIT-223)the main category-theoretic argument for radical (eliminative) OSR
Lam & Wüthrich 2015, No categorial support for radical ontic structural realism (LIT-222)the reply. It corrects Bain in the record’s relation graph.

The finding: no source in the core literature names Yoneda

Neither Bain nor Lam & Wüthrich mentions the Yoneda lemma, representable functors or universal properties. Lal & Teh (arXiv:1404.3049), searched in full text, does not either. Eva 2016 (EJPS) could not be obtained from a legitimate free source.

  • Bain stops one step short. He quotes Awodey on objects “determined … by their position in a structure”, and defines products by probing them from outside, but never takes the step to the lemma.
  • Lam & Wüthrich use two of its consequences without the name. Generalised elements X → A exist in every category, and morphisms “only distinguish isomorphism classes of structures”.

Two works that do name the lemma were found and not filed:

  • Adlam 2025 (arXiv:2507.20313). One sentence: Yoneda “implies that objects may emerge from the ‘morphisms’”. The source it cites, Yokura 2017, was read in full and says nothing of the kind.
  • Hendren, Grasso, Ellia & Tononi 2026 (arXiv:2608.24928). A box calls Yoneda OSR’s “mathematical centerpiece”. That is an assertion, and the paper is about qualia, not physics.

What the lemma does and does not support

The lemma says an object is determined up to isomorphism by its hom-functor, its relations to every other object. That supports objects as positions in a structure: the non-eliminative OSR of Ladyman 2001 (LIT-217), and Lam & Wüthrich’s “balanced” OSR. It does not support radical OSR, for three reasons:

  1. The determination presupposes the category’s objects. The hom-functor is indexed by them, and the conclusion is an isomorphism between two of them.
  2. It needs the composition structure, not merely which hom-sets exist. The nLab page’s own counterexample has two non-isomorphic objects whose hom-sets are in bijection at every object.
  3. It fixes objects only up to isomorphism. That is the categorical twin of the limit on weak discernibility in Leitgeb & Ladyman (LIT-154).

So Bain’s inference to eliminating objects is overstated. His reliance on elements as arrows out of a terminal object also fails once generalised elements are admitted.

Errors the readings found in the works

  • nLab: its finite counterexample breaks the identity law as written, and its HoTT proof has small errors. It gives only the presheaf form. Nat(Hom(A,−),F) ≅ F(A) follows by passing to C
  • Bain: treats sections of a sheaf of Einstein algebras as maximal ideals, and mis-defines a maximal ideal.
  • Both papers: say Hilb has no products. Direct sums are products.
  • Lam & Wüthrich: mis-define a topos, and say non-commutative algebras have no maximal ideals.

The epistemology tag, read through

2026-09-26 21:45:33 · curation · record

A pass over the epistemology tag, the same shape as the earlier tag passes: audit the tag, then read every unread work on it in full.

The audit

Seventeen works carried the tag; three had already been read (LIT-187, LIT-191, LIT-216). Five untagged works were about knowledge, evidence or trust and now carry it as a secondary tag:

workwhy
LIT-014Gelman on “reckless disregard for the truth”: testimony and trust in experts
LIT-131an account of understanding in deep learning systems
LIT-161what replication evidence can and cannot justify
LIT-162whether clustering is epistemically superior to classification
LIT-213rigour versus evidential diversity as warrant for singular causal claims

Readers also moved two primary topics, and added three tags:

  • LIT-158 (Desmond) now leads with epistemology: it is about when distrust of expert testimony is rational. It gains social-science, since its argument rests on the sociology of the professions.
  • LIT-163 (Wenmackers’ thesis) now leads with epistemology and gains philosophy-of-mathematics, which the thesis names as the home of its non-Archimedean probability chapter.
  • LIT-175 (Worrall on evidence-based medicine) gains natural-sciences.

The readings

Eleven works were read in full, and all are Active: LIT-118, LIT-131, LIT-139, LIT-158, LIT-161, LIT-162, LIT-163, LIT-165, LIT-175, LIT-189 and LIT-213. Two readings were of author versions rather than the version of record: LIT-118 (the 2015 Durham working paper) and LIT-161 (the PSA2018 manuscript). Each NOTE says so.

The corrections to the seeded summaries ran one way: the seeds overstated.

  • Desmond (LIT-158) argues that distrust can be valid reasoning. He calls it a “how-possibly” argument, not the showing that distrust is justified which the abstract implies.
  • Douglas (LIT-139) does not say every internal choice carries inductive risk.
  • Worrall (LIT-175) calls selection bias “a cast-iron argument for randomization”; the seed had listed it among the arguments that fail.
  • Wenmackers’ “under 2%” inconsistency figure (LIT-163) is a per-agent, single-update, numerical result for two and three atomic sentences. At the population level it reaches 16.9%.
  • Hartmann & Trpin’s measure (LIT-165) escapes Roche & Sober by two examples, not by a theorem.

One status was overridden. The reader rated LIT-163 Superseded: each chapter has a later journal version. None of them is on this record, and Superseded has to name a successor that is. The thesis also holds two things nothing else does: the Ch. 1 survey and the Ch. 4 synthesis. So it is Active, and its status note lists the four journal versions as the better citations for the other chapters.

Six works that could not be reached

These stay Deferred, and each status note records what was tried:

  • LIT-119: a closed Lexington chapter.
  • LIT-129 and LIT-203: Taylor & Francis refused the requests.
  • LIT-171, LIT-181 and LIT-186, all Pettigrew. These are open on PhilArchive, but PhilArchive’s bot wall refused every request from here. The Wayback Machine reset its connections all session.
  • LIT-104 (Carson) was not retried: its earlier attempt found it closed-access.

The trap for next time: PhilArchive is not a paywall. A session whose network reaches the Wayback Machine, or an owner who fetches the PDFs by hand, can close three of the six. For LIT-181 the Wayback copy is known to exist (2026-04-08).

A slip in the batch brief

The brief for this pass was made by editing the consciousness pass’s brief with sed. One phrase was broken across a line and survived: “consciousness-tag verdict”. One reader reported it and read it correctly as the epistemology tag. The brief is now fixed. When cloning a brief, grep it for the old tag name before handing it out.

The ethics tag, read through

2026-09-26 21:59:51 · curation · record

A pass over the ethics tag, the same shape as the epistemology pass: audit the tag, then read every unread work on it in full.

The audit

Thirty-four works carried the tag. Twenty had been read. Five had been tried in earlier passes and could not be reached, so they were not retried: LIT-117, LIT-129, LIT-137, LIT-186 and LIT-197. That left eight to read.

One untagged work now carries ethics. LIT-206 (Goldstein & Lederman), in its §7, argues that attributing desires to a chatbot does not settle welfare. It also argues that on a social-contract view such desires could still warrant moral consideration. A reader browsing AI moral status would expect it.

One work lost the tag. LIT-101 (Fitouchi, André & Baumard, moral disciplining) is descriptive moral psychology and cultural evolution. It is explicitly agnostic about whether puritanical norms work, and it draws no normative conclusion. Its tags are now social-science and cognition. The reader suggested religion too, but the paper mentions religion only in passing, so it was not added.

The readings

Four works were read in full, and all are Active:

  • LIT-098, Veit, on health concepts and social justice. Read from the PhilSci-Archive preprint; the journal version sat behind a bot challenge.
  • LIT-101, the moral-disciplining paper. Read from the accepted manuscript on OSF.
  • LIT-179, Schmauder et al., Algorithmic Nudging.
  • LIT-198, Hooker’s SEP entry Rule Consequentialism.

The corrections, as in the epistemology pass, mostly took claims away from the seeds:

  • Veit. The skim had him “conceding” that Kukla can hold the institution of medicine fixed. That is the second horn of his own dilemma: holding medicine fixed makes the account normativist “in name only”.
  • Schmauder et al. The seed said AI systems “can learn nudges that exploit biases nobody has documented”. Both case studies exploit documented biases. The undocumented case is a conjecture.
  • Hooker. The seed had the view survive by “acceptance by less than everyone and a prevent-disaster rule”. That misassigns the replies. Ranking rules by the expected good of teaching and accepting them answers the collapse objection. The prevent-disaster rule answers only rule-worship. The 90% threshold is there for punishment rules.

Four works that could not be reached

These stay Deferred, and each status note says what was tried:

  • LIT-128 and LIT-214: Taylor & Francis refused the requests.
  • LIT-174: a book with no open copy.
  • LIT-183: Springer closed access. Its publisher abstract does state the conclusion, and the status note now carries it.

So of the thirty-four works on the tag, nine remain unread, all for access reasons. None of the nine is openly available anywhere a legitimate route found.

The cognition tag, read through

2026-09-26 22:13:54 · curation · record

A pass over the cognition tag, the same shape as the epistemology and ethics passes: audit the tag, then read every unread work on it in full.

The audit

Twenty-seven works carried the tag. Nineteen had been read, and LIT-128 had been found unreachable in the ethics pass. That left seven to read.

Two works gained the tag:

  • LIT-147 (Pernu, mental causation via neuroprosthetics) is philosophy of mind, which the tag’s blurb names.
  • LIT-140 (Mallory, LLMs as stochastic measuring devices) takes the deflationary side of the machine-minds debate. It was unread, so it joined the list, making eight.

The readings

Six works were read in full, and all are Active: LIT-099, LIT-140, LIT-149, LIT-153, LIT-155 and LIT-178. LIT-140 was read from the author’s own September 2024 draft. The OUP chapter and the revised preprint could not be reached, and the NOTE says so.

Corrections worth knowing:

  • Ivy & Mroczko-Wąsowicz (LIT-099). The paper stays neutral on whether any single percept is penetrated. The skim had it asserting that none is.
  • Martínez (LIT-149). Footnote 7 makes the account teleosemantic, which the skim missed. Without teleofunctions, any system can be made to look rate-distortion efficient.
  • Angius et al. (LIT-153). The paper favours mechanistic explanation. Its primer describes transformer training as autoencoding, which is wrong.
  • Prinzing & Vazquez (LIT-155). “Outperform all other majors” ranks point estimates; it is not a test. All the test scores are self-reported.
  • Bottou & Schölkopf (LIT-178). The essay leaves open whether alignment can turn the “fiction machine” into a “verification machine”. It never uses the word “sycophancy”, which was the seed’s gloss.

Two citations were corrected from “et al.” to the two authors they are.

Retags from the readings

  • LIT-155 now leads with social-science and adds epistemology.
  • LIT-178 and LIT-153 drop linguistics. Its blurb excludes language “as models encode them”, which is what both are about. LIT-178 now leads with cognition.
  • LIT-153 and LIT-149 gain neuroscience: most of their evidence is neural.

Two works that could not be reached

LIT-134 (Prinz) and LIT-148 (López-Rubio) are paywalled with no open copy. They stay Deferred, and each status note says what was tried.

The philosophy-of-science tag, read through

2026-09-26 22:33:24 · curation · record

A pass over philosophy-of-science, the record’s largest philosophy tag with 81 works. Most had been read in earlier passes. Nine works on the tag had never had a close reading attempted.

The audit

Four already-read works gained the tag:

workwhy
LIT-021, the emergence editorialemergence in science; LIT-141 and LIT-150 already carry the tag
LIT-077, Grangier et al.an informal argument about what makes quantum randomness quantum
LIT-079, causal claims in economicsthe tag’s blurb names “causation and evidence”
LIT-086, Barandesa reformulation offered as an interpretation of quantum theory

The technical quantum-foundations papers (structure theorems, graphical calculi) were left alone. The quantum-foundations blurb already covers what the theory says about the world, and adding a second word for the same thing would say nothing new.

The readings

Six works were read in full:

Findings worth keeping:

  • Frigg & Werndl (LIT-114). The SEP entry has three frameworks, not two. The Boltzmann Equation (H-theorem, Lanford’s theorem) was missing from the skim. The entry gains information-theory and metaphysics.
  • Takola & Schielzeth (LIT-127). They present the time-slice, prospective and trajectory views of the niche as complementary, not as a forced choice. They concede the concept is “mostly of a metaphorical value”. The paper is behavioural ecology explicating its own term, so it now leads with natural-sciences. It does not cite Trappes (LIT-169). Its “Trappes et al.” is a different preprint.
  • Luque (LIT-205). The skim’s section map was largely wrong. Hamilton’s rule is only a footnote. The disputes the paper engages are van Veelen’s and Nowak’s: that the equation is a tautology, and that it lacks predictive power.
  • Grindrod et al. (LIT-208). Proposed: the claim that stability is “roughly proportional” to frequency is never measured.

A missing word

LIT-208’s reader found its philosophy-of-science tag unjustified: the paper is philosophy of language. linguistics excludes meaning “as models encode [it]”, so neither of the paper’s tags fit. Four more held works had the same gap. ADR-009 adds philosophy-of-language to the vocabulary and the philosophy group, and retags all five.

Three works that could not be reached

These stay Deferred, each with a status note:

  • LIT-108 (Weyl): in copyright, lending-only scans.
  • LIT-201 (Damuth & Heisler): paywalled at Springer.
  • LIT-180 (Wigner): open in the PNAS archive, but not fetchable from here.

For Wigner, the reader’s retries were refused by the session’s permission check. Nobody tried to route around the refusal. The next session should simply try again.

The last three philosophy tags, and what the philosophy passes could not reach

2026-09-26 22:47:04 · curation · record

This pass closes out the series of tag passes over the philosophy group. The three tags left were small:

  • logic: 3 works;
  • philosophy-of-mathematics: 7 works;
  • philosophy-of-language: 5 works, the topic ADR-009 added in the previous pass.

Only two of their works had never had a close reading attempted: LIT-106 and LIT-145. So this pass is one reading, one access failure, and an audit of all three tags.

The audit

worktag addedwhy
LIT-032, a type theory for synthetic ∞-categorieslogica type theory is a formal theory of proof
LIT-095, two-level type theorylogicthe same
LIT-054, a complete graphical calculus for Spekkens’ toy theorylogicits result is a soundness and completeness theorem for an equational proof system
LIT-151, Wallace on math-first structural realismphilosophy-of-mathematicsit asks what reading a theory as mathematics commits one to
LIT-207, Roberts on the fiction of artificial mindsphilosophy-of-languageit argues that chatbots’ self-ascriptions are systematically untrue, and treats engaging with them as make-believe

The sheaf-theoretic contextuality paper (LIT-016) was considered for logic and left alone. Its logic is the vehicle of a result about quantum theory, not its subject.

The reading

LIT-145, Turner’s Ultrafilters as Propositional Theories, is Active. It was read from the author’s 2021 preprint. The 2025 Wiley version was not reachable. Its abstract ends at Łoś’s theorem, so whether the preprint’s compactness section survived into print is unverified. The reader found four slips in the preprint’s proofs. The NOTE lists them.

LIT-106 (Marabini) could not be fetched. It is marked open access, but every route to it failed. Its status note says which.

What the philosophy passes could not reach

Every work in the group has now been read in full or had a reading attempted. Twenty-one remain Deferred for access, and each status note records what was tried.

Open, but not fetchable from this environment (five). These are the cheapest to close: a session whose network reaches the Wayback Machine and PhilArchive, or a PDF supplied by hand.

Closed access (sixteen). These need a library copy:

In two passes, philosophy of science and this one, a reader’s fetches were refused by the session’s permission check. Each reader reported the refusal and did not work around it. That is the behaviour wanted. It also means that some of the five open works were never actually tried by every route.

The information-theory tag, audited

2026-09-26 22:53:26 · curation · record

Every one of the sixteen works on information-theory had already been read in full. One of them, LIT-064, has since moved to the anthology. So this pass is an audit only: which works about information did not carry the tag?

The tag’s blurb is: entropy, channels, coding and capacity read for themselves, including information-theoretic accounts of work, memory and inference outside machine learning. Four works meet it and now carry the tag:

workwhy
LIT-005, Bourrat 2024its subject is Krakauer et al.’s three mutual-information measures of individuality, and which of them measure it
LIT-025, Mediano et al.ΦID, an integrated-information decomposition, defines causal emergence
LIT-027, Varley & Hoela partial information decomposition of past–future mutual information
LIT-010, Vinjanampathy & Andersquantum thermodynamics, whose account of work runs through Landauer and entropy, which the blurb names

Considered and left alone:

  • LIT-049 (Bourrat 2023) and LIT-150 (Rizi). Their coarse-graining accounts of individuality and emergence mention mutual information in passing, but their machinery is prediction, not information.
  • LIT-136 (Floridi on personal identity). It is philosophy of information in name: its tool is levels of abstraction, not entropy or channels.
  • LIT-044 (Stigler on maximum likelihood). Fisher information appears there as history of statistics, which probabilistic-modeling already covers.

Representation learning, description length and Riesz: an owner’s reading list filed

2026-09-26 23:15:41 · curation · record

The owner asked for a list of 24 works to be filed, grouped under their own headings:

  • Knowledge Graphs;
  • Relating K-complexity to minimum sufficient statistics;
  • Riesz Representation Theorem;
  • Representation learning as a spectral approximation;
  • Platonic Representation Hypothesis;
  • Compling ∩ Renormalization Group Theory.

Two were already held, in the anthology

  • Huh et al., The Platonic Representation Hypothesis, is ANTH-LIT-458.
  • Gallego & Orús, Language Design as Information Renormalization, is ANTH-LIT-695.

Neither is duplicated here.

Twenty-two were filed

Four dossier agents seeded them. Twenty-one were skimmed from the full text. Vera, Piantanida & Rey Vega (LIT-240) is paywalled, so it is seeded from its abstract. All are Deferred: filed, not yet read.

They needed words the vocabulary did not have. ADR-010 adds two: representation-learning and learning-theory.

The ten works chiefly about ML practice carry anthology-candidate. These are:

  • the six self-supervised-learning papers, five of them spectral accounts;
  • knowledge sheaves;
  • the information and conditional-entropy bottlenecks;
  • MDL generalization for representation learning.

The Kolmogorov-complexity and MDL theory, the compression bounds, the Riesz material and the SEP entry are theory read for itself, and carry no flag.

Corrections the seeds made to the list

The list’s venues and identifiers were checked against the works:

  • Knowledge sheaves appeared at AISTATS 2023, not 2021. 2021 is the year of arXiv v1.
  • Assran et al. is an ICLR 2023 poster, and Tan et al. appeared at ICLR 2024. The list gave “arXiv” for both.
  • Sun et al. was in the ICML 2023 proceedings before its arXiv posting.
  • Sefidgaran, Zaidi & Krasnowski first appeared at NeurIPS 2023.
  • Simon et al. is arXiv 2303.15438.
  • Sefidgaran et al. 2022: its second author is “Amin Gohari” on arXiv.
  • Grünwald & Vitányi 2004: the draft carries the authors’ own warning that it contains errors, and has no journal version.

One relation

Sefidgaran, Zaidi & Krasnowski argue that Vera et al.’s bound on I(X;U) is vacuous in realistic settings. It is recorded as corrects. The claim comes from a skim, and a reading should confirm it.

”Do we even need Riesz?”

The owner’s note under the Riesz heading asked this. The dossier agent’s answer, from reading the four pages: yes, as a lemma, not as a reading.

  • Riesz is what makes kernels exist. Evaluation at a point is a continuous functional, and Riesz turns it into the kernel vector k_x.
  • It is why a linear probe or “concept direction” can be treated as a vector.
  • It is why adjoints, and so the spectral arguments in the self-supervised learning papers, make sense.
  • Adler’s REU paper ends on Radon–Nikodym, the density result that KL divergence and mutual information, and so the bottlenecks, rest on.
  • Bra–ket is notation.
  • GNS is a suggestive analogy: a state determines its representation up to unitary equivalence. That echoes the Platonic hypothesis’s kernel comparisons, but no representation-learning paper found uses it.

None of the four pages makes these connections itself. They are the dossier’s own and are marked as such.

Riesz, Radon–Nikodym and GNS in representation learning: the connections pursued

2026-09-26 23:42:28 · curation · record

The owner filed the Riesz representation theorem, bra–ket notation and the GNS construction next to a reading list on representation learning, and asked “do we even need Riesz?”. A dossier agent answered with four connections:

  • Riesz makes kernels exist;
  • a linear probe is a functional, and so a vector;
  • Radon–Nikodym underwrites KL, mutual information and the bottlenecks;
  • GNS says a state determines its representation up to unitary equivalence, echoing the Platonic hypothesis.

The source pages made none of these connections themselves. The owner asked for them to be expanded and pursued. Three research agents took one strand each and read the relevant theorems in the papers’ own texts:

  • Riesz and reproducing kernels;
  • Radon–Nikodym and density ratios;
  • GNS and unitary equivalence.

The one object the strands meet in

The three strands converge on a single object: the positive-pair density ratio K⁺(a,b) = p⁺(a,b) / (p(a)p(b)). It is four things at once.

  1. A Radon–Nikodym derivative. It is dP⁺/d(P⊗P), and InfoNCE, the logistic loss and the spectral contrastive loss all have it as their population optimum. InfoNCE and skip-gram reach it in the log domain, as PMI (THEORY-001).
  2. A Mercer kernel. Its kernel principal components are the eigenfunctions of the positive-pair Markov chain. Their span is the minimax-optimal linear basis for view-invariant targets (THEORY-007).
  3. The kernel of the conditional-expectation operator on L²(p). The record’s own derivation: (Tf)(a) = E[f(A′) | A = a] = ∫ K⁺(a,b) f(b) p(b) db. T is Hilbert–Schmidt exactly when 1 + χ²(P⁺‖P⊗P) < ∞. L²(p) is the space the commutative GNS construction builds from the state f ↦ E_p f (THEORY-005, Proposed: no source states this whole).
  4. The exponential of the PMI kernel. Huh et al.’s Platonic hypothesis proves convergence to the PMI kernel under its assumptions (ANTH-LIT-458, §4.2).

The owner’s reading list therefore has one topic under four headings. The spectral self-supervised learning papers, the density-ratio view of contrastive learning, and the Platonic hypothesis all describe the same operator.

Answers, connection by connection

Riesz makes kernels exist: true, and stated. Muandet et al. (LIT-260) state Riesz (Thm 2.4) and derive the reproducing property from it (Prop. 2.1).

  • The kernel mean embedding is a Riesz representer. It is the representer of taking an expectation (Lemma 3.1). This is the one place in the family where Riesz is invoked by name.
  • Self-supervised learning can be read as separating mean embeddings. Li et al.’s SSL-HSIC (LIT-248) turns this into an objective: it maximizes the average squared MMD between each image’s view distribution (THEORY-003).

Two parts of the original connection did not survive:

  • The representer theorem does not use Riesz. Schölkopf, Herbrich & Smola’s proof (LIT-261) uses the reproducing property and orthogonal projection instead.
  • No spectral SSL paper invokes Riesz. Their self-adjointness comes from the symmetry of the positive-pair distribution (THEORY-009).

So, do we need Riesz? Yes, for defining a reproducing-kernel space abstractly, for kernel mean embeddings, and for the adjoint of a general bounded operator. No, for the finite-space spectral theorems the self-supervised learning papers actually prove.

Asymmetric methods are where Riesz would earn its keep. Predictor and stop-gradient methods (BYOL, SimSiam) and cross-modal pairs (CLIP) are not covered by the symmetric theory. There the natural object is the singular value decomposition of a non-self-adjoint operator, which needs the adjoint in earnest. No such account was found. That is the open question worth pursuing next.

A probe is a functional: true, finite-dimensional, and already held. Park, Choe & Veitch (ANTH-LIT-606) use a “Riesz isomorphism” to map a concept’s probe direction to its steering vector. Their substance is that training does not fix the inner product.

Harvey, Lipshutz & Williams (LIT-257) prove the representation-side fact. What a regularized linear readout decodes depends only on the representation’s normalized kernel. CKA, CCA and GULP are averages of readout agreement (THEORY-008).

No paper treats probes on an infinite-dimensional feature space as bounded functionals.

Radon–Nikodym underwrites mutual information: true. Belghazi et al. (LIT-256) write mutual information as the expected log Radon–Nikodym derivative. The Donsker–Varadhan bound is tight exactly there.

Poole et al. (LIT-246) prove that InfoNCE lower-bounds mutual information for every critic and never exceeds log K (THEORY-006). Tschannen et al. (LIT-247) show that tightening the bound does not explain why the representations are good.

Two caveats from this strand:

  • The conditional entropy bottleneck (LIT-226) works with entropies, not the Radon–Nikodym form.
  • Vera et al. (LIT-240) could not be read, so whether it states the absolute-continuity condition is unverified.

GNS: true as a parallel, stated by one source, and used by no representation-learning paper. Jorgensen & Tian (LIT-259) name the kernel-to-Hilbert-space construction “the GNS construction” (Remark 1.34). They prove a kernel’s minimal realizations are unique up to a unitary (Cor. 1.35).

On the representation side:

  • Harvey, Larsen & Williams (LIT-250) prove that Procrustes distance equals the Bures distance between kernels (THEORY-004). So the kernel is the invariant, and a representation is one realization of it.
  • Nielsen et al. (LIT-253) show that a softmax model’s distribution fixes its embeddings only up to an invertible linear map. Closeness in KL does not bound representational dissimilarity.
  • The Platonic hypothesis’s convergence, read carefully, is convergence of kernels, up to the symmetry group of whatever is observed (THEORY-002, Proposed).
  • Huh et al. measure with mutual nearest neighbours, which is much weaker than kernel equality.

The strand’s own caveat stands: GNS of E_p gives L²(p), which contains no learned representation. What connects them is the training objective, through the density ratio.

Status of what was filed

  • Works: sixteen, all Deferred with Skimmed notes. Two works were found by two strands (HaoChen et al., Johnson et al.). Each is filed once, with both skims in its note.
  • THEORY documents: nine, all Proposed. Six are theorems proved in their sources that this record has only skimmed; each promotes on a close reading. Three are inferences across papers, and each names the kind of result that would settle it.
  • Relations. The operator claim and the self-adjointness claim extend the eigenbasis claim, which extends the shared-optimum claim. The readout and Platonic claims extend the kernel-invariance claim.

Open questions the strands left

  • Are Simon et al.’s NTK-based kernel (LIT-242) and K⁺ the same operator in any limit? Simon et al. say their representations differ from the optimum-based accounts.
  • Rank-d truncation of PMI versus of exp(PMI): do InfoNCE-style and spectral-style representations differ in the top-d subspace? No paper compares them.
  • Is χ²(P⁺‖P⊗P) < ∞ stated anywhere as the condition for the spectral account on continuous inputs?
  • Do models with high mutual-nearest-neighbour alignment also have high Bures alignment of their centred kernels?
  • Worth filing next: NeuralEF (arXiv 2205.00165, read but not filed), its follow-up (arXiv 2210.12637), Aronszajn (1950), and Wang & Isola’s alignment/uniformity paper. Whether its uniformity term is a squared mean-embedding norm is unchecked.

Contextuality: the threads brought together

2026-09-27 02:19:45 · curation · record

The owner asked whether the record had an account of contextuality. It had the works: five were read in full. What it lacked was the account: no tag to find them by, and nothing connecting the two formal threads they form.

The owner asked for the tag, and for the threads to be brought together with theory entries. ADR-012 adds the tag. A research agent read the five works’ notes and, where needed, the texts. It drafted the claims below and seeded the four bridging works the record lacked:

  • Spekkens (2005), the bridge between the two notions (LIT-266);
  • Abramsky, Barbosa & Mansfield’s contextual fraction (LIT-265);
  • Budroni et al.’s 2022 review (LIT-263);
  • Dzhafarov, Zhang & Kujala on behavioural and social data (LIT-264).

The account in seven claims

claimstatus
THEORY-012: Kochen–Specker contextuality is the failure of context distributions to glue into one global distribution. Signed ones always exist, so negativity is no marker.Active (LIT-016, read)
THEORY-016: generalized noncontextuality is equivalent to a positive quasiprobability representation, which is a rigid, exact frame.Active (LIT-003, read)
THEORY-015: Kochen–Specker noncontextuality is measurement noncontextuality plus outcome determinism, and preparation noncontextuality forces the determinism. Every Kochen–Specker proof therefore refutes the universal notion.Proposed (Spekkens 2005, skimmed)
THEORY-011: the generalized notion is strictly broader: a qubit is already preparation-contextual.Proposed (skimmed)
THEORY-014: in both formalisms, classicality is positivity of a representation that always exists over the reals.Proposed (the record’s inference joining two read theorems)
THEORY-013: behavioural and word-meaning data published as contextual show none, once context-dependent marginals are separated out.Proposed (skimmed; a working hypothesis of its authors)
THEORY-010: Grangier & Auffèves’s “contextual objectivity” is an ontological postulate, not contextuality in either formal sense.Proposed (the record’s placement)

The shape. The first two claims are the two threads, each proved in a work the record has read. The third and fourth are how the threads relate:

  • the Kochen–Specker notion sits inside the generalized one;
  • the containment is strict.

Both rest on Spekkens (2005), the one source that states the link, which is seeded and not yet read.

Inferences. The fifth claim notices that the threads share a mathematical shape; it is marked as the record’s inference. The sixth takes the topic outside physics. The seventh places a sixth work, Grangier & Auffèves, that uses the word differently.

Relations. The third claim extends the first two, the fourth extends the first, and the fifth extends the first two.

What was not claimed

  • Restricted fragments. Spekkens’s proof that preparation noncontextuality forces outcome determinism uses quantum theory’s full preparation set. Whether it holds inside a stabilizer or interferometric fragment is open. That bears on Catani et al.’s unproved claim that its toy theory is noncontextual.
  • Sheaf ⇄ GPT. Whether a sheaf-noncontextual empirical model yields a noncontextual model of the corresponding GPT fragment is not addressed by either work.
  • Representation learning and retrieval. No held or seeded source connects contextuality to the kernel THEORY documents or to van Rijsbergen (LIT-262). One observation is recorded as an open question: measurements that are all functions of one random variable always have a global section. So linear probes on one fixed representation cannot be Kochen–Specker-contextual. Any contextuality in a learned system would have to come from the representation changing with the context. That is the record’s reasoning from LIT-016 Prop 3.1, not a published result.
  • The contextual fraction. It would extend the first claim as a graded measure. It is left as a candidate until its supplementary proofs are read.

Tag audit

Tagged:

Not tagged:

To promote the Proposed claims

  • A close reading of Spekkens (2005) would move the third and fourth claims.
  • Kunjwal & Spekkens (2015), not held, is the noise-robust successor worth filing next.
  • Reading Grangier’s own derivations would settle the seventh.

The Geometry of Information Retrieval, read

2026-09-27 02:28:52 · curation · record

The owner supplied the text of van Rijsbergen’s The Geometry of Information Retrieval (LIT-262): front matter, Preface, Prologue, chapters 1–6 and the index. The appendices and the annotated bibliography were not included.

The book had been seeded that morning from the publisher’s excerpts and one review. It is now read, and it is Active with a Read note (NOTE-239).

Two corrections to what the record said

  1. Riesz is the book’s own point. The seed attributed the functional-as-vector identification to Kantor’s review. Chapter 6 (p. 74) states it, citing Riesz & Nagy: query matching is a linear functional, and so an inner product with a vector.

  2. The book models non-commuting measurement in retrieval. The contextuality curation entry of this date reported Kantor’s review as saying the book has no retrieval analogue of non-commuting measurements. That is not true of the book:

    • relevance and aboutness are explicitly allowed not to commute (AR ≠ RA), and the sequence A → R → A can change the second answer (pp. 21–22);
    • chapters 2 and 5 develop compatibility and the failure of distribution.

    What the book lacks is data showing these effects in retrieval, not the notion. That entry is left as written, since the journal is history, and this one corrects it.

What the reading adds to the record’s threads

  • The book never discusses contextuality. It is therefore not tagged contextuality. But its order effects are exactly the question-order setting that Contextuality-by-Default analyses (LIT-264, THEORY-013). That analysis found a Hilbert-space model of question order to predict no contextuality. Whether retrieval judgements show the book’s order effects at all is untested, by the book or anyone the record holds.
  • GNS. The book does not mention it and works in the opposite direction, from state to probability via Gleason. THEORY-004’s GNS link stays an inference.
  • The strongest parts are two chapters. Chapter 5 proves that the subspace conditional is a Stalnaker conditional, equal to the Sasaki hook. Chapter 6 recasts standard retrieval operations as trace calculations. Everything else is a language offered, not a result shown.

The Lattice Representation Hypothesis, re-read against the geometry of IR

2026-09-27 02:48:40 · curation · record

The owner allowed a work to be held in both records (ADR-013). They asked for Xiong’s The Lattice Representation Hypothesis of Large Language Models to be re-read here in light of van Rijsbergen’s The Geometry of Information Retrieval (LIT-262). The anthology holds it as ANTH-LIT-460.

Reason for a second entry (ADR-013 requires one). The anthology reads the paper for concept geometry in LLMs. This record reads it as a piece of the older mathematics of retrieval: formal concept analysis, Galois connections, and the logic of lattices of classes. That is van Rijsbergen’s second chapter, and the anthology holds none of it.

The paper was read in full (arXiv v3, ICLR 2026, with appendices). It is Proposed here with a Read note (LIT-267, NOTE-240).

What the re-reading found

  • The construction is van Rijsbergen’s, with one change. The paper’s concept lattice is the (tr, in) Galois connection his chapter 2 builds for inverted files; its closed pairs are his monothetic kinds. The change: the object–attribute incidence is a thresholded linear functional on an embedding, not a given index-term incidence.
  • Theorem 1 is formal concept analysis’s standard fact. Any incidence has a complete concept lattice. The paper itself says the geometry plays no further role once the incidence is fixed.
  • Distributivity is never discussed, and it fails. The reader’s own checks, marked as such in the note:
    • the object-level lattice can fail distributivity (a three-object diagonal context gives M₃), exactly as van Rijsbergen shows for kinds;
    • the region lattice is Boolean for linearly independent directions;
    • it is non-distributive for dependent ones.
  • No projectors, no trace-rule probability, no non-commuting operations. So van Rijsbergen’s quantum-probabilistic layer has no counterpart here, and contextuality has no purchase on the text. Two further observations, both the reader’s own:
    • the paper’s soft incidence is a sign-sensitive per-attribute marginal, not a probability measure on the lattice;
    • its inclusion score is not reflexive.
  • The evidence. On five WordNet domains, linear probes recover an incidence generated by GPT-4o at F1 69.7–83.2. No lattice law is tested.

The anthology’s entry needs correcting

Per ADR-013 this record reports these and does not edit the anthology. I checked the first against the paper’s text; the rest are the reader’s findings.

  1. The join. ANTH-LIT-460 says the join is “the conic hull of the two concepts’ attribute directions”. Definition 7 defines it as the set union R(Y_A) ∪ R(Y_B), which “can be approximated by the conic hull”. A union of cones is generally not a region of the family, so it is not a lattice join. The experiments use neither: they use coordinatewise min and max of attribute profiles.
  2. The evidence. The attribute schema and the “ground-truth” incidence were generated by GPT-4o (§4), not taken from WordNet, which supplies only the hierarchy.
  3. The complete lattice proved is the finite object-level concept lattice, not a lattice of regions.
  4. The soft version is thresholded to a crisp incidence before any lattice appears. No soft lattice is proved.
  5. The canonical-form condition is that Dc = τ is solvable. Linear independence of the directions is sufficient, not the condition itself.
  6. The causal inner product appears only in the preliminaries. The experiments use Fisher-LDA directions on raw hidden states.

Items 1, 3 and 4 bear on ANTH-THEORY-034’s account, which the anthology should re-check.

Attention as sparse distributed memory, and ZipNN, read for this record

2026-09-27 03:55:48 · curation · record

The owner asked for nucleation readings of two works. Both were read in full and both are Active.

Bricken & Pehlevan, Attention Approximates Sparse Distributed Memory

LIT-269, NOTE-242. Also held in the anthology as ANTH-LIT-641.

Reason for a second entry (ADR-013 requires one). The anthology reads the paper for attention practice, mainly its QK-norm retrodiction. This record reads it as associative memory:

  • Kanerva’s content-addressable retrieval;
  • the cerebellar mapping;
  • the SDM read as a kernel smoother, set against van Rijsbergen (LIT-262) and the kernel THEORY documents.

What the reading found

  • The paper never says “kernel”, but its construction is one. The weight SDM gives each stored memory is the overlap of two Hamming balls, which is a positive-definite kernel of distance. Locally it is approximated by exp(β·cos), the von Mises–Fisher kernel. Softmax attention with unit vectors and a fitted β is then a normalised kernel smoother over the memories.

  • Neuron-level SDM is linear attention with a fixed binary feature map. This is the reader’s observation, not the paper’s.

  • The approximation holds over a narrower range than the paper claims.

    • It holds for distances below the radius d, not up to 2d. At n = 1000 the fit over the full range has R² ≈ 0.68.
    • The overlap is a staircase, equal at distances 2k−1 and 2k.
    • The paper’s analytic β is off by about 2–2.5×, because it approximates a far-tail binomial with a Gaussian. The β the experiments use comes from regression instead.

    These are the reader’s recomputations.

  • The biology is weak. The cerebellar mapping restates Kanerva’s rather than re-arguing it. For attention it would need about 2⁶⁴ neurons, which the paper itself calls implausible. How a circuit would write every token in one forward pass is not addressed.

  • Against van Rijsbergen. His tr(ρP) is a broad degree-2 cosine kernel that ranks documents and is normalised over the query. SDM and attention use a sharp, tunable kernel, return a reconstructed vector rather than a ranking, and normalise over the stored memories. Same family, different job.

The anthology’s entry needs correcting

This record reports these and does not edit the anthology (ADR-013).

  1. The interpolation claim. ANTH-LIT-641 and ANTH-THEORY-097 say the learned QK-norm band β ∈ [10, 25] interpolates between SDM’s three optima. The paper claims it only for two of them, the critical-distance and signal-to-noise optima. I checked this against the text. Its Fig. 7 caption says the memory-capacity optimum, β = 35.5, is “much larger than the largest” learned βs.
  2. The residual stream. ANTH-LIT-641 says the paper interprets “the residual stream” through the SDM frame. The paper never mentions it (also checked). Its §4 covers the feed-forward block, LayerNorm, value norms, multi-head attention and stacking.
  3. Where the optima come from. ANTH-THEORY-097 credits all three to Kanerva’s book and his 1992 review. The paper cites Kanerva (1993) for two of them and derives the memory-capacity optimum itself.
  4. The GPT-2 check. The abstract’s “we confirm … in pre-trained GPT2” overstates Appendix A.2. GPT-2’s effective β mostly lies in [3, 12], below the QK-norm band, and GPT-2 does not normalise its queries and keys.

ZipNN: Lossless Compression for AI Models

LIT-268, NOTE-241. Not held in the anthology. It is tagged anthology-candidate because it is chiefly about ML practice.

What the reading found

  • The exponent carries almost all the redundancy.
    • About 40 of its 256 values occur, and 12 cover about 99.9% of parameters. The exponent compresses to about 33%, while sign and mantissa stay near 100%.
    • Whole models compress to about 66% (BF16), 83% (FP32) and 34–48% for rounded “clean” models.
    • On Llama-3.1-8B, ZipNN is 17% smaller and 62% faster than Zstd.
  • The information theory is thin.
    • The paper never computes an entropy or compares its codes to one.
    • Its explanation of the exponent skew is asserted, not argued. The paper itself says “No matter the reason”.
    • Part of that explanation is wrong: Adam’s ε does not add noise to weights, and 2⁻²³ is exponent 104, not 99.
    • Its “ExaByte per year” savings claim read “per month” in v1, which the paper’s own 6 PB/day figure rules out.
  • No bearing on the record’s description-length works. Lossless coding of the float format compresses how the numbers are written, not the function the network computes. It does not touch Kolmogorov complexity or MDL (LIT-224, LIT-225, LIT-238), or the lossy compressibility of the Sefidgaran bounds (LIT-233, LIT-236).

Housekeeping

Two notes filed earlier today (NOTE-239, NOTE-240) cited Proposed documents without acknowledgement directives. The directives are added.

Quantum structure claimed in LLM language, read and rejected

2026-09-27 04:16:24 · curation · record

The owner asked for a nucleation reading of Entropy 28(6):622. It is Aerts et al., Identifying Quantum Structure in AI Language: Evidence for Evolutionary Convergence of Human and Artificial Cognition: DOI 10.3390/e28060622, arXiv 2511.21731, LIT-270.

Access. MDPI refused automated requests. The full published text was read from Europe PMC’s open-access XML. The figures (captions only), the supplementary materials and the rows the published Table 4 omits were not read.

Verdict: Rejected. Neither of the paper’s two pieces of evidence supports its conclusion that LLM “code is structurally quantum”. The first of the two failures is already on the record.

  • The Bell test fails on context-dependent marginals. The paper runs its 2011 “The Animal Acts” test on ChatGPT and Gemini and reports CHSH values above 2. The published v1 abstract called this “quantum entanglement”; the journal abstract says “non-classical probability model”.

    But its own tables show outcome marginals that change with context. The paper’s CHSH is computed as if the marginals were consistent. Under that confound, a value above 2 does not show that no classical model exists.

    This is exactly what the record’s THEORY-013 records, from LIT-264’s Contextuality-by-Default analysis of the same test design on human subjects. That analysis found the published human violation disappears once marginals are handled. The paper does not cite it.

  • The Bose–Einstein fit is a relabelled Zipf curve. It is fitted to the word counts of one story that Gemini wrote after being told about the research. From the paper’s own constants, the fitted curve is essentially a Zipf–Mandelbrot law. Its Maxwell–Boltzmann comparator is an exponential, which cannot fit word frequencies. There is no goodness-of-fit statistic.

  • An internal inconsistency, checked. Table 3 prints E(A′,B) = +1.000, where the text gives 0.556. The reported CHSH of 3.09 follows from 0.556; the table’s value gives 3.53.

Worth noting for later. A published comment on v1 raises the same points: Sienicki, arXiv 2601.06104. It is not held.

NTK of matrix product states, read and rejected

2026-09-27 04:32:47 · curation · record

The owner asked for a nucleation reading of arXiv 2111.14046, Guo & Draper, Neural Tangent Kernel of Matrix Product States: Convergence and Applications (LIT-271). Neither record held it. The full 19-page v1 (the only version; no published version was found) was read, and the reading is NOTE-244.

Verdict: Rejected. The paper’s one sound calculation is that, in the sequential infinite-bond-dimension limit, the MPS NTK is Σ_k φ(x_k)·φ(x′k) Π{l≠k} σ_l² φ(x_l)·φ(x′_l). That is the MPS’s own GP product kernel times the constant Σ_k σ_k⁻², which follows in a line from the model being multilinear in its tensors. I checked that reduction when filing, and the paper never states it. Tensor Programs II (ANTH-LIT-557) covers this kind of architecture rigorously. The distinctive claims do not hold up:

  • The lazy-training lemma is circular. It uses the gradient at initialisation for all times and bounds entries, not the kernel.
  • Positive-definiteness is not proved. The abstract says it holds “without any extra assumptions of the data set”. The appendix proof shows at most semi-definiteness, and strict definiteness fails whenever m > Π|s_i|, the finite feature dimension the GP theorem itself assumes.
  • “Over-fitting is naturally avoided” contradicts the paper’s own memorisation result.
  • Several constants are off by factors of n or 2
  • There are no experiments.

The Born-machine closed form P_x(t)=1/m−(1/m−P_x(0))e^{−4mKt/Z} is correct; the reader checked it numerically.

Filing choices. It is held here, not in the anthology, because it is a theory paper with nothing for practice to act on. Its NTK ancestry is already in the anthology (ANTH-LIT-360, ANTH-LIT-557, ANTH-LIT-558). The tags are learning-theory, probabilistic-modeling (the GP limit and Born machines) and mathematics. The reader proposed quantum-foundations; I dropped it, because MPS here is a function class borrowed from physics, and nothing in the paper is about the foundations of quantum theory. No existing nucleation work is genuinely connected. LIT-242 (stepwise SSL as kernel PCA) was considered and set aside.

Trap. apply_read.py’s new-NOTE path truncates the title at its first colon, which leaves the note’s heading disagreeing with its paper’s title. It bit LIT-270’s note too, and was fixed by hand both times.

DisCoCat, founding paper and Frobenius sequel, read

2026-09-27 04:57:53 · curation · record

The owner asked for a reading of arXiv 1401.5980: Kartsaklis, Sadrzadeh, Pulman & Coecke, Reasoning about Meaning in Natural Language with Compact Closed Categories and Frobenius Algebras (LIT-272, NOTE-246). While it was being read, the owner added the paper it builds on, arXiv 1003.4394: Coecke, Sadrzadeh & Clark, Mathematical Foundations for a Compositional Distributional Model of Meaning (LIT-273, NOTE-245). Neither record held any DisCoCat or pregroup work before. The 2014 paper is related to the 2010 one as extends.

Both are Active, as specifications rather than results.

  • The 2010 paper states no theorem and runs no experiment. Its sentence spaces are 1-d or 2-d truth values in a toy model, and it gives no method for building S or the verb tensors from data. It is still where the ε/η composition and the single sentence space come from. Its claim that the category distinguishes grammatical ambiguities fails, because the pregroup is posetal. Its printed similarities are unnormalised inner products, under its own cosine definition.
  • The 2014 paper is the citable categorical statement of the Frobenius lift of verbs. Its experiments are the authors’ COLING 2012 numbers, re-reported. I checked Table 2 against COLING 2012’s Table 1 and they are identical. The chapter drops the COLING caveat that copy-object’s lead is “statistically insignificant”, and its abstract’s “verify the theoretical predictions” has no prediction behind it. The best result is ρ = 0.172, against a human bound of 0.620.

Where basis dependence enters. The 2010 composition is invariant under an orthogonal change of basis. The Frobenius copying map of 2014 is not: one random rotation moved a sentence–subject cosine from 0.008 to −0.53. The basis dependence belongs to the Frobenius extension, not to DisCoCat as such. That is the point of contact with van Rijsbergen’s basis-free geometry (LIT-262).

A status overridden. The 2014 reader proposed Superseded, by its COLING 2012 antecedent and by the later “Frobenius anatomy” papers and Kartsaklis’s thesis. None of those is in either record, and Superseded names its successor. The antecedent also lacks the functorial recasting, which is the chapter’s stated contribution. So the status is Active, with the overlap stated in the status note. If COLING 2012 or the thesis is filed later, this should be revisited.

Tags. The 2014 paper keeps representation-learning, because its vectors and verb matrices are built from a corpus. The 2010 paper drops it, because it learns nothing. Neither takes quantum-foundations: both borrow the categorical quantum formalism and say nothing about quantum theory.

A defect fixed in passing. LIT-271 reached main with two number: keys, a lint violation. I had hand-written number: tmp… into the seeded frontmatter, and luria concretize then prepended the real number. luria new lit writes no number: for a temporary code, and neither should a hand-seeded file. The stray line is removed here, and no other document carries a duplicate.

Where a Hilbert space gets its basis: Carroll connected, THEORY drafted

2026-09-27 06:11:49 · curation · record

The owner asked for a reading of arXiv 2103.09780. It is Carroll, Reality as a Vector in Hilbert Space, which the record already held as LIT-123, read in full the day before (NOTE-134, Active). A second reading would have added nothing. What had changed was the record around it: van Rijsbergen, the two DisCoCat papers and the Riesz/GNS kernel thread were all filed after NOTE-134 was written. On the owner’s go-ahead, this contribution does two things.

NOTE-134 gains a connections paragraph (version 3). The reading itself is unchanged. Its “Bearing on the record” line saying no THEORY depended on the paper is updated.

THEORY-017 is new, Proposed. It says that in a Hilbert-space model only unitarily invariant structure is intrinsic, and that the record’s Hilbert-space works differ in where they get a basis or factorisation. There are four answers:

  • Carroll takes it from the dynamics.
  • Van Rijsbergen takes it from the observable.
  • DisCoCat 2010 needs none; its composition is orthogonally covariant.
  • Kartsaklis et al. 2014 stipulates a corpus basis.

THEORY-004 is the same point for learned representations, so the new document extends it. Two inferences are marked as the document’s own, not the sources’:

  • Contextuality needs more than one basis, so a single-stipulated-basis model such as LIT-272’s cannot be contextual.
  • Distributive, Boolean concept structure is evidence that a basis was chosen.

Why Proposed. The piece that would turn “basis”, “Frobenius algebra” and “classical data” from analogies into one notion is Coecke, Pavlovic & Vicary’s theorem. That theorem is cited by LIT-272 and not held here. So the promote_when asks for that paper to be filed and read, and for a check that each placement in the table rests on something measured or proved, not inferred. It also names what would refute the account.

Checked while writing. A first draft said LIT-272’s basis-dependent model “scores as well as its invariant rivals”. It doesn’t have invariant rivals in that sense. Kron and Multp are elementwise too, so basis-dependent, and the only rotation-invariant baseline, additive, scores lowest (ρ = 0.050 against 0.163–0.172). The caveat now says so. The pattern is worth noting: on that one task, the basis-dependent models lead.

Candidate follow-up. Filing Coecke, Pavlovic & Vicary (arXiv 0810.0812) would let this THEORY be promoted. It would also give LIT-272’s Frobenius-algebra claims (NOTE-246, C3) a source in the record rather than a citation.

Coecke, Pavlovic & Vicary read; THEORY-017 restated, still Proposed

2026-09-27 06:26:32 · curation · record

The owner asked for arXiv 0810.0812 next: Coecke, Pavlovic & Vicary, A New Description of Orthogonal Bases (MSCS 23(3), 2013). THEORY-017’s promote_when had named it. It is filed as LIT-274, read in full as NOTE-247, and Active.

What it proves.

  • Thm 5.1. In FdHilb, commutative †-Frobenius algebras are in bijection with orthogonal bases, the basis being the vectors the comultiplication copies. The bijection is of sets, for each fixed space.
  • §6. The special ones are exactly the orthonormal bases, with the vectors fixed exactly, phases included.
  • Corollaries 7.1–7.2. These give categorical equivalences with finite sets labelled by positive reals, and with FinSet.
  • The proof. Its key step rests on the cited spectral theorem for finite-dimensional commutative C*-algebras.
  • Language and scope. The paper says “classical data”, never “classical structure”. It says nothing about infinite dimensions, distributions, contextuality or projection lattices.

What changed in THEORY-017. It now states the correspondence as proved: orthogonal bases for commutative †-Frobenius algebras, orthonormal bases for special ones. It also gains a caveat that the correspondence holds on complex spaces only.

  • The real-space counterexample. The reader found it, and I checked it numerically: on ℝ², the complex numbers with multiplication scaled by 1/√2 form a commutative special †-Frobenius algebra with no nonzero copyable vector. The associativity, Frobenius, special and commutativity laws all hold, and the copy residual shrinks only as the vector goes to zero.
  • What it means for the table. LIT-273, LIT-272 and THEORY-004 work over the reals. A basis there still induces a Frobenius algebra, which is all LIT-272 uses, so the table’s placements stand. But “basis” and “Frobenius algebra” are one notion only in the complex case.

Why it stays Proposed. The first promote_when condition is met. The second is not. Two placements rest on the readers’ own numerical checks, not on the papers’ claims: LIT-273’s orthogonal covariance and LIT-272’s rotation sensitivity. The promote_when now says so. Promoting on the strength of what the readers computed would be exactly the move the field exists to prevent.

NOTE-246 corrected (version bump). Two sentences overstated this paper. The spider normal form is not in it, though LIT-272 cites it for that. And the basis correspondence needs † and a complex Hilbert space. Its claims table’s C3 was accurate and is unchanged.

Candidate follow-up, not filed. For infinite dimensions, the reader found by search, but did not read, Abramsky & Heunen (arXiv 1011.6123). That is where THEORY-017’s finite-dimensional boundary would be tested.

Abramsky & Heunen read; THEORY-017’s finite-dimensional caveat narrowed

2026-09-27 06:43:22 · curation · record

The owner asked for arXiv 1011.6123: Abramsky & Heunen, H*-algebras and nonunital Frobenius algebras (AMS PSAPM 71, 2012). The previous entry had named it as the test of THEORY-017’s finite-dimensional boundary. It is filed as LIT-275, read in full as NOTE-248 (arXiv v3, the final version), Active, and it extends LIT-274.

Checked against the text when filing: Lemma 3, Theorem 22, Proposition 23 and Theorem 35 read as the reader reported.

What it proves.

  • Lemma 3. A unital Frobenius algebra exists in Hilb only in finite dimension.
  • Theorem 22. Without the unit, a commutative special †-Frobenius algebra on a complex Hilbert space of any dimension is induced by an orthonormal basis exactly when it satisfies Ambrose’s H*-axiom, equivalently when it is semisimple, or a directed colimit of finite unital ones. It adds an approximate-identity form for separable spaces.
  • Proposition 23. It leaves open whether the axiom is automatic, i.e. whether a radical part can exist.
  • Tools. It uses Gelfand theory and Ambrose, not GNS or C*-algebras.
  • Relations. In categories of relations the axiom is proved, for quantale-valued relations only under Theorem 35’s hypothesis q² = 1 ⇒ q = 1. The abstract’s “always coincide” overstates this.

THEORY-017 (version 3). The “stays finite-dimensional” caveat is narrowed. A discrete basis is extra structure supplied as an algebra on the space in any dimension. Only continuous observables, which the paper defers, remain outside, and that is where Carroll’s and the GNS strand’s “an algebra of observables must be supplied” applies. The real-scalar caveat now holds in every dimension: the reader found that ℂ regarded as a real algebra satisfies the H*-axiom and has no copyable vectors. I checked the other axioms numerically in the previous contribution; the H*-axiom check is the reader’s. The document stays Proposed, on the same remaining promote_when condition.

Small corrections recorded in the NOTE. The paper’s “Theorem 1 [CPV09]” is Coecke, Pavlovic & Vicary’s §6 (special/orthonormal) form, not their Theorem 5.1 (orthogonal). The orthogonal case appears here only as a sketch in a footnote.

Candidate follow-up, not filed. Poinsot, Hilbertian Frobenius algebras (arXiv 2003.04149, 2020). By its abstract, which is all the reader saw, every commutative special Hilbertian Frobenius algebra is semisimple. If the body bears that out, Proposition 23’s question is closed and THEORY-017 can drop “was left open there”.

Measurement as sheafification, read and rejected

2026-09-27 07:26:15 · curation · record

The owner asked for arXiv 2512.12249 v2: Ghose, Measurement as Sheafification: Context, Logic, and Truth after Quantum Mechanics. It is filed as LIT-276, read in full as NOTE-249, Rejected. The reader read v2 and compared it word by word with v1. No journal version was found.

Verdict. It is an interpretive essay with no theorem, model or computation of its own. Its correct formal core is Abramsky–Brandenburger’s global-section criterion, which the record holds first-hand as LIT-016 and THEORY-012. Each of its three additions is asserted without a construction, or is wrong as stated:

  • “Measurement is sheafification.” It gives no site, no topology and no sheafified presheaf. Sheaf toposes are not Boolean in general (Sh(ℝ) is not). And classical probability distributions do not form a sheaf, since the marginal problem breaks gluing.
  • The seven-valued Ghose–Patra logic as “a finite Heyting algebra”. It gives no order and no operations. The structure its Appendix B implies is not a lattice, and completing it gives the Boolean algebra 2³.
  • The σ–λ dynamics. λ(σ) is never specified, nothing is computed, and there is no prediction.

v2’s added Čech obstruction is always zero. I checked this against the text when filing. The appendix defines the class as [δs] for a 0-cochain s. That is the class of a coboundary, so it vanishes for every s, and “[δs] ≠ 0” can never hold. The paragraph that follows describes Abramsky, Mansfield & Barbosa’s actual construction correctly, but only by citation.

Beyond its comment, v2 also:

  • deletes v1’s introduction, leaving refs [1]–[8] uncited;
  • deletes v1’s only empirical claim, the micellar-surfactant one;
  • adds a quantum-gravity paragraph.

Bearing. It supports no THEORY. Its slogan “Boolean needs a sheaf” resembles THEORY-017’s inference “Boolean structure needs a basis”, and the NOTE warns against reading it as support. The slogan is false as stated, and THEORY-017’s inference rests on different grounds.

Candidate follow-up, not filed. Abramsky, Mansfield & Barbosa’s cohomology of contextuality is the correct version of what v2’s appendix attempts, and neither record holds it. Filing it would give THEORY-012 and LIT-265 their cohomological companion.

Cohomology of contextuality read; Ghose note corrected, THEORY-012 bounded

2026-09-27 07:38:16 · curation · record

The owner asked for Abramsky, Mansfield & Barbosa’s cohomology of contextuality, the follow-up the Ghose entry named. It is arXiv 1111.3620, The Cohomology of Non-Locality and Contextuality (QPL 2011, EPTCS 95, 2012). It is filed as LIT-277 and read in full as NOTE-250 (v2, compared with v1 word by word). It is Active, and it extends LIT-016.

What it shows.

  • The obstruction. For a section s of context C₁, the obstruction γ(s) is the class of δ⁰(s, s₂, …, sₙ) in the relative group Ȟ¹(𝒰, F_{C̄₁}), where F is the presheaf of ℤ-combinations of support sections.
  • What vanishing means. γ(s) = 0 iff s extends to a compatible family of ℤ-combinations (Prop 4.2), which need not be a global section. So a non-vanishing class witnesses possibilistic or strong contextuality, and only in one direction: it is sufficient, not necessary.
  • Where it works. γ is non-zero for the PR box, GHZ, the 18-vector Kochen–Specker set and Peres–Mermin.
  • Where it fails. γ vanishes on all of Hardy’s model, and on 9 of 15 sections of a strongly contextual Kochen–Specker cover.

The reader recomputed every example over ℤ and ℤ/2.

A v1 error, verified. v1’s “strongly contextual” false-positive cover, {A,B,C}, {A,D,E}, {B,D,E}, {A,D,F}, {A,E,G}, is not strongly contextual. I enumerated it when filing: it has exactly two global Kochen–Specker assignments. v2 silently replaced it.

NOTE-249 corrected (Ghose). Its description of this construction was written from general knowledge. “Not the class of a 0-coboundary” was wrong as worded: γ(s) is a coboundary’s class, taken relative to C₁ and built from data that are not cochains of the relative presheaf. The note now says what Ghose’s [δs] drops (the relativisation, the anchoring and the cover), and why that makes it vanish. The verdict on Ghose is unchanged.

THEORY-012 bounded (version bump). Its “does not say” gains a line: the cohomological witness is a computable relaxation of the exact global-section criterion, not an equivalent. The claim itself is unchanged.

Candidate follow-ups, existence checked only.

  • Abramsky, Barbosa, Kishida, Lal & Mansfield, Contextuality, Cohomology and Paradox (CSL 2015, arXiv 1502.03097).
  • Carù (arXiv 1701.00656 and 1807.04203).
  • Okay, Roberts, Bartlett & Raussendorf, Topological proofs of contextuality (arXiv 1701.01888).

Contextuality, Cohomology and Paradox read; THEORY-012’s cohomology line refined

2026-09-27 08:14:56 · curation · record

The owner asked for Contextuality, Cohomology and Paradox: Abramsky, Barbosa, Kishida, Lal & Mansfield, arXiv 1502.03097, CSL 2015. It is filed as LIT-278 and read in full as NOTE-251. That covers v2, which matches the published LIPIcs text, and v1, which holds the only proof of Prop 16. It is Active, and it extends LIT-277.

What it proves.

  • Theorem 21. For a connected cover, AvN_R(S) ⇒ SC(Aff S) ⇒ CSC_R(S) ⇒ CSC_ℤ(S) ⇒ SC(S). So a model with an All-vs-Nothing argument, over any commutative ring, is strongly contextual, and its cohomological obstruction is non-zero on every section.
  • Theorem 4. Stabiliser states of n-qubit “AvN triples” of Paulis supply such models (qubits only).
  • GHZ(3). §4.1 is a complete hand proof, the case NOTE-016 had recorded as only sketched.

The paradox half proves little.

  • One exact case. The length-4 Liar cycle is the PR box.
  • The rest is analogy. The cohomology of paradox is deferred.
  • Scope of the encoding. It fails for the classic Liar (N = 1) and for N = 2. For N ≥ 3 the Liar cycles are All-vs-Nothing over ℤ₂; I checked that when filing. The strong contextuality then follows from Theorem 21, though the paper doesn’t say so.

NOTE-250’s gaps are not closed. The reader found that Hardy’s affine closure is the whole event sheaf, so no All-vs-Nothing argument reaches Hardy. By Theorem 21, the §8 strongly contextual cover of LIT-277 has no All-vs-Nothing argument over any ring. The paper mentions neither, and v2 deletes v1’s sentence about Hardy.

THEORY-012 (version bump). Its cohomology line now says where the witness is complete: on All-vs-Nothing models. It also says that both of LIT-277’s documented failures lie outside that class. The derivation for covers failing the GCD condition is marked as the reader’s own. The claim is unchanged.

Silent corrections between versions.

  • v2 corrects v1’s GHZ parity equations without comment.
  • v2 adds “all known quantum examples of strong contextuality are of AvN type” with no citation.

Follow-up. By its abstract (unread), Carù’s arXiv 1701.00656 disproves LIT-277’s Conjecture 8.1. It is now the most direct next reading on this line.

Carù read: the cohomology conjecture is false; THEORY-012 updated

2026-09-27 08:29:06 · curation · record

The owner asked for arXiv 1701.00656: Carù, On the Cohomology of Contextuality (QPL 2016, EPTCS 236). It is filed as LIT-279 and read in full as NOTE-252 (v1, the only version). It is Active, and it corrects LIT-277.

The counterexample holds.

  • The model. It sits on the two-party, two-setting, four-outcome Bell cover.
  • Checks. The reader reproduced NOTE-250’s counts first, as a calibration: Hardy 13/13, PR box 0/8, AMB §8 cover 9/15. Then, with exact ℤ and ℤ/2 linear algebra, the reader confirmed that the model is strongly contextual and that γ = 0 on all 22 sections. I checked the strong contextuality by hand when filing; the γ computation is the reader’s.
  • Beyond the paper.
    • The reader found, by linear programming, a no-signalling probability model with this support.
    • The model has no All-vs-Nothing argument. That follows from LIT-278’s Theorem 21 by contraposition, and the reader also checked it directly over ℤ₄.

Scope of the refutation. It refutes Conjecture 8.1 for symmetry of the cover, and for “no measurement confined to one context”. It does not touch covers whose contexts pairwise intersect (AMB’s own §8 cover is one), or symmetric models: this model has 2 automorphisms, against the PR box’s 16.

The rest of the paper is weaker.

  • The hierarchy collapses. The higher-obstruction hierarchy (Prop 5.5) is empty above level 0 for every no-signalling model, by the paper’s own Prop 5.6. So the abstract’s “hierarchy of levels of contextuality” is the one level AMB already had.
  • Prop 6.1 is false in its “only if” direction. The PR box is a counterexample.
  • No positive completeness result.

It closes neither the Hardy gap nor the §8 cover.

THEORY-012 (version bump). Its cohomology line records the refutation and the open cases. The claim itself is unchanged.

Follow-up, existence only. Carù, arXiv 1807.04203, Towards a complete cohomology invariant, is the natural next reading on this line.

Carù 2018 read: complete on cycles, conjecture refuted

2026-09-27 16:16:44 · curation · record

The owner asked for arXiv 1807.04203: Carù, Towards a complete cohomology invariant for non-locality and contextuality. It is filed as LIT-280 and read in full as NOTE-253 (v1, the only version, 46 pp.). It is Active, and it extends LIT-279. No published version was found. The author’s 2019 DPhil thesis was not obtained, so whether it repairs §8 is unverified.

What holds: Theorem 7.7. On cyclic covers, ℤ/2 cohomology of the (N−1)-th iterated joint model is exact for logical and strong contextuality. The reader implemented the construction and checked it on 659 random contextual cycle models. It detects:

  • Carù 2017’s (2,2,4) model at level 3, on 22 of 22 sections;
  • Hardy’s one contextual section at level 3.

The paper claims both, and the computation confirms them. So both of the record’s gaps on cycles are closed.

What fails.

  • Conjecture 9.1 is refuted. Under the paper’s Definition 7.1, LIT-277’s §8 cover is undetected at every level. The reader built the lifted witness family explicitly up to level 4, which has 1350 contexts. I re-ran the reader’s level-1 witness script when filing, and its independent compatibility check passes.
  • Prop 8.2 is false on the paper’s own Table 5 model.
  • The paper’s recount of the §8 cover is wrong (4/15 against the correct 9/15).
  • The worked examples use a weaker, context-wise definition than the theorems. That weaker version detects the §8 cover at level 2; that finding is the reader’s.
  • The abstract’s “vast majority of empirical models” is established only for cycles.
  • Cost is never analysed. The iterated objects grow exponentially: 5, 10, 30, 150, 1350 contexts on the §8 cover.

THEORY-012 (version bump). Its cohomology line records the partial repair on cycles and the refutation off them, and “still open” is restated. The claim itself is unchanged.

The line so far.

  • LIT-277: the witness.
  • LIT-278: complete on All-vs-Nothing models.
  • LIT-279: incomplete on symmetric covers.
  • LIT-280: complete on cycles, with the general conjecture false.

The next step, if the owner wants it, is the 2019 thesis, to see whether §8 was repaired.

Carù’s 2019 thesis read: nothing repaired

2026-09-27 21:57:33 · curation · record

The owner asked for Carù’s DPhil thesis, Logical and Topological Contextuality in Quantum Mechanics and Beyond (Oxford, 2019, supervised by Abramsky). The reader obtained it from the Oxford Research Archive (open, CC BY; DOI 10.5287/ora-kznbp4k4y, which I checked resolves to the ORA record). It is filed as LIT-281, read as NOTE-254, Active, and it extends LIT-280.

  • Read in full: Chapters I–V and VIII. Chapter IV was diffed against arXiv 1807.04203.
  • Skim-read, statements only: Chapters VI (AvN triples) and VII (minimum resources for strong non-locality). The brief allowed this.
  • Date: ORA gives the year only, so published: is its deposit date.

The question was whether the thesis repairs the 2018 preprint beyond cycles. It does not.

  • Republished unchanged. Chapters III–IV republish LIT-279 and LIT-280, with “joint” renamed “line”.
    • Def 7.1 → Def IV.26, with the same universal definition.
    • Thm 7.7 → Thm IV.32, which holds.
    • Prop 8.2 → Prop IV.34, reworded but still false on the same table.
    • Conjecture 9.1 → Conjecture IV.41, restated as open.
  • Other errors carried over. The wrong level-0 equation of §6.1 and LIT-279’s false Prop 6.1 (now Prop III.8) are both kept.
  • Two new claims are false.
    • “No model without the CCP is known” (p. 88). The §8 cover is one.
    • The §8 cover has a unique false ℤ/2 family (p. 91). A second exists.
  • The reader’s computation on the §8 cover. Under the thesis’s own definition, the reader ran the 2018 scripts unchanged, since the definitions are unchanged. The §8 cover of LIT-277 is detected on 6 of 15 sections at every level from 0 to 3, and its non-detected section lifts compatibly through level 4. The context-wise reading, which the thesis never states, detects all 15 sections at level 2.

What is new in the thesis.

  • Probabilistic line models (Thm IV.40).
  • Chapter V (with Abramsky) recasts contextuality in valuation algebras. It gives inference algorithms with bounds such as O(k^n·l^{k(n−1)+1}), never implemented.
  • Chapter VI characterises All-vs-Nothing stabiliser subgroups by AvN triples. Its proof was not checked.
  • Chapter VII: strong non-locality needs three qubits in the GHZ class.

THEORY-012 (version bump). The cohomology line cites the thesis for the cyclic result, and says its extension and conjecture are restated without repair. It adds that neither paper nor thesis analyses cost. The claim itself is unchanged. The reader’s optional stabiliser-characterisation line was not added, because its proof was not checked.

Where the line stands. A complete, computable cohomological test is established only on cyclic covers. LIT-277’s §8 cover remains the standing counterexample to the general conjecture, and the context-wise variant that does detect it has not been published by anyone. The bibliography’s “Car17 = arXiv:1701.00242” is the proceedings volume; LIT-279 is 1701.00656.

Sen’s capability approach, assembled from nine texts

2026-09-28 02:56:43 · curation · record

The owner asked for Amartya Sen’s ideas to be represented, and suggested assembling the capability approach from shorter texts and reviews rather than reading The Idea of Justice (2009) in full. Nothing by or about Sen was held before this entry.

A scoping pass checked what was legitimately and fully reachable. Nine texts were then read in full, by four readers working in parallel, and each is filed Active with a Read NOTE.

RoleWorkLITNOTE
Sen, core statementEquality of What? (Tanner Lecture, 1979)LIT-289NOTE-258
Sen, social-choice groundingThe Possibility of Social Choice (Nobel lecture, 1998)LIT-282NOTE-263
Sen, formal resultThe Impossibility of a Paretian Liberal (JPE 1970)LIT-285NOTE-255
OverviewRobeyns & Byskov, SEP The Capability Approach (2011; 2025 revision read)LIT-283NOTE-260
OverviewWells, IEP Sen’s Capability Approach (2012)LIT-284NOTE-257
SynthesisRobeyns, Wellbeing, Freedom and Social Justice (2017)LIT-287NOTE-259
Critique, the listNussbaum, Capabilities as Fundamental Entitlements (2003)LIT-288NOTE-256
Critique, justice methodValentini, A Paradigm Shift in Theorizing about Justice? (2010/2011)LIT-290NOTE-262
The 2009 book, indexedZwolinski, review of The Idea of Justice (2010)LIT-286NOTE-261

The relations added are: the Nobel lecture extends the 1970 paper, and both Nussbaum and Robeyns extend the 1979 lecture.

The account, piece by piece, and where each piece comes from

Each claim below is marked first-hand (read in a Sen text here) or second-hand (reported by a secondary source, which is named).

  1. The negative argument (first-hand, 1979). Three measures of advantage are each insufficient, even for the needs-based part of equality:

    • utilitarian equality (equal marginal utility), because sum-maximisation gives the less efficient converter of income into utility less income, pp. 203–204;
    • total-utility equality, completed as leximin, because a contented disabled person with low aspirations shows no utility deficit, pp. 217–218;
    • Rawls’s primary goods, which are “fetishist” because they take goods “as the embodiment of advantage, rather than taking advantage to be a relationship between persons and goods”, p. 216.

    The disabled person’s claim is tracked by none of them.

  2. The positive proposal (first-hand, 1979). “Basic capability equality”, meaning “a person being able to do certain basic things” (p. 218). Sen offers it only as “a partial guide” to the equality part of morality, and says the negative thesis stands without it (p. 220). The index of capability bundles yields only a partial ordering, completed by culture-dependent conventions (p. 219).

    The 1979 text never uses “functioning”, “capability set”, “agency” or “freedom”. Accounts that credit it with the full later vocabulary, or with replacing utility and primary goods, read later Sen back into it. Robeyns does this (p. 76 n. 39; NOTE-259), and so does the IEP entry in places.

  3. The mature vocabulary (second-hand).

    • Functionings are achieved beings and doings; capabilities are the real opportunities to achieve them; the capability set is the menu (per the SEP and the IEP).
    • Conversion factors are why equal resources yield unequal capabilities (per all three secondary sources).
    • Adaptive preferences are why utility misleads (per the IEP, quoting Commodities and Capabilities).
    • Well-being freedom is distinguished from agency freedom, and freedom from achievement. This is attributed to the Dewey Lectures (1985) by the SEP (§2.5) alone; the IEP never mentions it. No Sen text for it was reachable.

  4. The social-choice grounding (first-hand, 1998).

    • What escapes Arrow is interpersonal comparability. Even ordinal comparisons break the exact impossibility, and partial comparability often suffices. Capabilities enter only as one candidate answer to “comparison of what?”, which Sen declines to make exclusive.
    • The liberal paradox is different. Sen proved it in 1970, and the reader checked the proof. It is explicitly not resolved by comparability (p. 199).
    • A correction to the scoping report. It ran these steps together.

  5. The “no fixed list” stance and public reasoning (second-hand). Sen refuses “one pre-determined canonical list” (SEP §3.3, quoting “Human Rights and Capabilities” 2005, p. 158). The two overviews disagree about what replaces the list:

    • the IEP credits Sen with a social-choice procedure of “intersection” and “ranges of weights” (§3b);
    • the SEP says he has “never explained in detail” how selection should work (§3.3).

  6. Justice (second-hand, via two readers of the 2009 book).

    • Comparative, not transcendental. Sen argues for a comparative, “realization-focused” approach against “transcendental institutionalism” (Valentini; Zwolinski).
    • Positional objectivity and the impartial spectator (Zwolinski).
    • Freedom is “an inescapably plural idea” (Zwolinski, IoJ p. 305).
    • Human rights whose realisation need not be the state’s task (Zwolinski, IoJ p. 364).

    The two readers agree on what the book says and disagree on whether it matters. The SEP adds that the capability approach is not itself a theory of justice (§4). The IEP calls Sen a “comprehensive consequentialist”, and the SEP calls that contested.

The critiques as the record now holds them

  • Nussbaum: the list and the threshold. The approach yields no theory of justice without a definite, revisable list of central capabilities, owed up to a threshold. She gives ten. “Freedom” is not a general good, since some freedoms are trivial or unjust. She endorses Sen’s arguments against GNP, utility and resources. Her threshold is named but not specified in this paper, and Sen’s side appears only as her reports of his oral remarks.
  • Valentini: method. Each of Sen’s complaints against transcendental theory either misreads Rawls or is true but inconsequential. “If there are transcendental institutionalists, Rawls is not one of them.” She says nothing about capabilities.
  • Robeyns: the modular reply. The core is eight “A-module” commitments; particular theories fill seven compulsory-but-open modules. She uses this to argue that the standard critiques (no list, individualism, power, liberalism) hit particular capability theories, not the approach. The approach is committed to ethical, not ontological, individualism. She is an advocate and Sen’s former student; read her as a reconstruction that generalises Sen, including dropping his “reason to value” clause.
  • Pogge’s resourcist critique is not represented. None of the nine texts states his argument. The scoping report expected Robeyns 2017 and the SEP to carry it, and both readers found that neither does. Pogge 2002 itself is paywalled.

Errors found in the secondary sources (checked against Sen 1979)

  • The SEP inverts Sen’s pleasure-wizard. In Sen the pleasure-wizard is person B, not the disabled person A (p. 203). I checked this against the text when filing. The SEP also merges his argument against utilitarianism with his separate argument against total-utility equality (§§1.1, 2.6).
  • Robeyns over-attributes to the 1979 lecture.
    • She has it argue that almost any theory of justice is egalitarian in some space (p. 94 n. 2). The lecture makes no such argument; the reader thinks the thesis is from Inequality Reexamined (1992), unverified.
    • She counts it as endorsing her whole A-module (p. 76 n. 39).
  • The IEP is internally inconsistent on when the approach began (“1980s” against 1979), and drops Sen’s condition from the utilitarian argument.

What is missing, and why

Sen’s own statements. None of these is reachable legitimately, and all are represented only second-hand through the SEP and the IEP:

  • the Dewey Lectures (1985), the source for agency freedom;
  • Capability and Well-Being (1993);
  • Capabilities, Lists, and Public Reason (2004);
  • Elements of a Theory of Human Rights (2004);
  • Human Rights and Capabilities (2005);
  • Development as Capability Expansion (1989).

Critics. Pogge (2002) and Anderson (1999) are also unreachable.

Blocked hosts. Sen’s Harvard page, the Tanner Lectures site, PhilArchive and Open Book Publishers’ own site refused automated requests. The owner may be able to fetch copies in a browser. The Dewey Lectures and the 2004 Lists paper would most improve the account.

Dating conventions.

  • Valentini’s published: is the November 2010 working paper, which is her first appearance (ADR-002).
  • Zwolinski’s issue date is inferred and unverified.
  • Nussbaum was read from her own preprint, so her page numbers are not the journal’s.

Existing record links.

  • LIT-202 (preference formation) is cited for adaptive preferences, but only thematically: it cites Elster and Khader, not Sen.
  • LIT-146 treats Nussbaum’s capabilities as a value theory for animals.

No THEORY document was filed. The candidate claim — that Arrow is escaped by comparability, while the Paretian-liberal paradox is not — rests for its Arrow half on Sen’s 1970 book, which is not held.

Frankfurt’s account of higher-order volitions, assembled from seven texts

2026-09-29 20:09:53 · curation · record

The owner asked for Harry Frankfurt’s account of higher-order volitions to be represented. Nothing on it was held before this entry. A scoping pass found that the core texts are all out of legitimate reach:

  • Frankfurt’s 1969–1994 papers, including “Freedom of the Will and the Concept of a Person” (1971);
  • Watson’s “Free Agency” (1975);
  • Bratman’s papers.

The copies of the 1971 paper found on course sites were not used. Seven texts were read in full and each is filed Active with a Read NOTE:

RoleWorkLITNOTE
Frankfurt, first-hand (late view)Tanner Lectures, Taking Ourselves Seriously / Getting It Right (2004)LIT-293NOTE-264
The 1971 apparatus and the repliesMcKenna & Coates, SEP Compatibilism (2004; 2024 revision read)LIT-291NOTE-270
Overview of autonomyBuss & Westlund, SEP Personal Autonomy (2002; 2018 revision read)LIT-296NOTE-267
Overview of autonomyDryden, IEP Autonomy (2010)LIT-297NOTE-266
Critique of the hierarchyVelleman, What Happens When Someone Acts? (1992)LIT-295NOTE-265
Critique of the late viewVelleman, Identification and Identity (2002)LIT-292NOTE-268
The 1969 casesRobb, SEP Moral Responsibility and the Principle of Alternative Possibilities (2020)LIT-294NOTE-269

The account, and where each piece comes from

Each claim below is marked first-hand (read in Frankfurt here) or second-hand, with the source named.

  1. The 1971 hierarchy (second-hand, SEP Compatibilism §4.2.1).
    • First-order desires are for actions; second-order desires are desires about one’s desires.
    • The will is the “effective first-order desire”, the one that moves one “all the way to action”.
    • A mere second-order desire (the therapist who wants to know what craving is like) differs from a second-order volition, a desire that a given desire be one’s will.
    • The unwilling addict has a second-order volition against the craving that moves him. The willing addict endorses it.
    • The formula for free will: “One acts of her own free will if and only if her action issues from the will she wants”. I checked the wording against the entry when filing.
    • Page references are to the Watson (ed.) 1982 reprint.
  2. Wantons and persons (second-hand, and only in older revisions).
    • Only persons have second-order volitions. Wantons are “passive bystanders to their wills”. “The wanton is not a person, and so is not a candidate for freely willed action.”
    • This is in the SEP Compatibilism revisions of 2004–2015 only (e.g. the Fall 2019 archive edition). The 26 Nov 2019 revision dropped it.
    • None of the other six texts states it, and the Tanner Lectures mention only “wantonly unreflective” animals and small children (p. 172).
  3. Freedom of action vs freedom of the will (first-hand, Tanner p. 177). Frankfurt restates the distinction and his compatibilism (p. 178). No secondary source in the batch does.
  4. The mature view (first-hand, Tanner Lectures). I checked the quotations against the lecture PDF.
    • Identification is largely a default “willing acceptance”, “whether with a welcoming approval, or in weary resignation”.
    • Free will is “a certain volitional unanimity”: the governing desire agrees “with a higher-order volition” and conflicts with no desire “at any higher order”. That is wholeheartedness.
    • Caring is wanting to go on wanting.
    • Love is volitional. Its necessities “may be irresistible, but they are not coercive”. It is “the only legitimate authority” for our normative attitudes.
    • The lectures do not use the terms “second-order desire” or “second-order volition”, do not discuss the regress, and never name Watson. The case for love’s authority rests on the stipulation that lovers are never ambivalent.
  5. The regress and Watson (second-hand, and the sources disagree).
    • The IEP and SEP Compatibilism credit Watson (1975) with the authority/regress objection. SEP Personal Autonomy states the regress as its own argument, credits it to nobody, and files Watson 1975 as a rival (valuational) account.
    • Velleman (1992) reports the regress as Watson’s objection, which Frankfurt conceded.
    • None of the sources reports Watson’s valuing-versus-desiring distinction in any detail.
  6. Frankfurt’s replies (second-hand, dated inconsistently).
    • Decisive commitment: the IEP and Velleman place its germ in 1971 itself; SEP Compatibilism says 1971 “rested solely” on higher-order volition.
    • Wholeheartedness: the IEP dates it to 1987, SEP Compatibilism to 1994.
    • Satisfaction (1992): only SEP Personal Autonomy, n. 1, mentions it.
    • Manipulation: SEP Compatibilism quotes Frankfurt (2002, p. 27) accepting that a character given by a manipulator still leaves a person responsible.
    • Resignation: the resigned willing addict is excluded by the 1994 amendment according to SEP Compatibilism, but Frankfurt’s own 2004 text counts acceptance “in weary resignation” as identification.
  7. Critiques.
    • Velleman 1992: on the “standard story” of action the agent drops out. Any state defined without the agent’s participation (second-order desires, Watson’s values, decisive commitments) can be disowned, and any state defined to include it presupposes what it was meant to explain. His alternative: the agent is whatever plays the agent’s functional role, the desire to act in accordance with reasons.
    • Velleman 2002: this targets the late view. An essence made of volitional necessities would yield authenticity, not autonomy. Wholeheartedness is a defensive ideal (the Rat Man). Frankfurt’s “Reply to J. David Velleman” in the same volume was not obtained.
    • Manipulation and mesh objections: per SEP Compatibilism and Personal Autonomy.
    • Bratman’s self-governing policies (2007): per SEP Compatibilism only. Bratman 1996 and 2004 are reported by no source.
  8. Frankfurt cases (1969; second-hand, Robb). These cases are the strongest attack on the principle that responsibility requires being able to do otherwise, and there is “not much consensus”. The 1969 case appears only in Fischer’s retelling. The bridge to the hierarchy (§5.1) is that identification or endorsement is one of two compatibilist accounts of responsibility after Frankfurt cases.

Gaps, and what would fill them

  • Frankfurt 1971, 1976, 1987, 1992 and 1994, and Watson 1975, are all paywalled. Watson is the largest loss: his objection is reported, but with contradictory attributions.
  • Stump 1988 has a legitimate author-linked copy on PhilArchive, which blocks this environment. The owner could fetch it in a browser.
  • The mid-2019 SEP Compatibilism revision holds the wanton/person text. It was read only for that section.
  • LIT-117 (Enoch, Agency, Shmagency, Deferred) is the constitutivist target that Velleman’s functional proposal invites.
  • LIT-130 (personal identity and ethics) bears on Velleman 2002.
  • LIT-202’s “Frankfurt-style case” is the 1969 device, not the hierarchy.

The agent drops out of the causal story in Velleman 1992, and Sen’s agency freedom (LIT-289, second-hand through LIT-283) is a neighbouring notion. The record does not yet connect them beyond this sentence.

The owner’s prior-art novelty map, registered in full

2026-09-29 21:50:47 · curation · record

The owner asked for every reference in their prior-art novelty map to be registered in this record. The map is a private working document on concepts as projectors, the linear representation hypothesis and their neighbours.

How it was done. Three agents worked in parallel, split by domain:

  • quantum foundations, logic and representation theory;
  • ML, information theory and the thermodynamics of learning;
  • philosophy and linguistics.

For each reference they checked both records, pinned down the exact work and verified its identifier. Identifiers were verified against Crossref, arXiv or the publisher. Where no DOI covers the first edition, a stable Internet Archive URL is used instead: Riesz 1907, Ryle 1949 and Armstrong 1978.

The result. The map names 63 works once multi-part works are split: Aerts–Gabora I and II, Döring–Isham I–IV, Artin–Wedderburn as two papers, and Cramér–Rao as Rao 1945 and Cramér 1946.

  • 6 were already held. Four were held first-hand: LIT-262, LIT-273, LIT-016 and LIT-220. LIT-254 and LIT-267 were also already held.
  • 57 are new seeds, all Deferred, from metadata and published descriptions. None has been read.

Held in both records (ADR-013). Six of the new seeds are also held in the anthology: Park 2023, Huh 2024, McCandlish 2018, Shwartz-Ziv & Tishby 2017, Power 2022 and Nanda 2023. The reason for a second entry is the one ADR-013 asks for: this record holds them for the owner’s prior-art map, the anthology for ML practice. Each seed names its anthology entry in its body.

Near misses the agents caught. These look held but are not:

  • LIT-263 is a review of Kochen–Specker, not the 1967 paper.
  • LIT-270 is Aerts 2025, not Aerts–Gabora 2005.
  • LIT-175 is Worrall 2007, not 1989.
  • LIT-045, LIT-184, LIT-217, LIT-218 and LIT-219 are Ladyman or Ross works, but not Ladyman 1998 or Every Thing Must Go.
  • LIT-042 and LIT-224 are not Shannon 1949.
  • The record’s information-bottleneck items (LIT-240, LIT-226) are follow-ups, not Tishby, Pereira & Bialek 1999.
  • The Riesz cluster (LIT-230, LIT-243) is not the 1907 paper. That paper is registered now.

Loose references, filed under the agents’ proposed reading. These are flagged in each entry and open to revision by the owner:

  • Kondor: Kondor & Trivedi 2018.
  • Aristotle: Categories ch. 10 and De Interpretatione ch. 7, in Ackrill’s translation.
  • Armstrong: Universals and Scientific Realism, both volumes as one entry.
  • Gelfand: Gelfand–Naimark 1943. Its DOI is the 1994 reprint.
  • Stone duality: Stone 1936.
  • Pontryagin: 1934.
  • Peter–Weyl: 1927.
  • Walsh: 1923.
  • Shannon limit: Shannon 1949.
  • Artin–Wedderburn: both papers.
  • Haag: the 1996 second edition.
  • Van Trees and Rao: DOIs of later reissues, dated to first publication.
  • Murota et al. 2010: the map’s author list merges two papers. The registered one is Murota, Kanno, M. Kojima & S. Kojima. The companion, Maehara & Murota 2010, is not registered.

Not registered, deliberately. These are adjacent to the map’s references but not named by it:

  • Isham–Butterfield parts II–IV;
  • Lyons vol. 2;
  • the Ganter–Wille second edition;
  • Wille 2001;
  • Sommers 1964;
  • Gelfand 1941;
  • Stone 1937;
  • Haag’s first edition;
  • Maehara & Murota 2010;
  • Fréchet’s 1907 note.

One inconsistency flagged for later reading. For Engels 2024 and for Tishby, Pereira & Bialek 1999, the agent’s venue details were from memory. The entries keep only what was verified.

The map’s references and their codes

As the map cites itWorkCodeStatusAlso in anthology
Aerts & Gabora, “A theory of concepts and their combinations” (2005)A theory of concepts and their combinations I: The structure of the sets of contexts and propertiesLIT-340new seed
Aerts & Gabora, “A theory of concepts and their combinations” (2005)A theory of concepts and their combinations II: A Hilbert space representationLIT-312new seed
Busemeyer & Bruza, Quantum Models of Cognition and Decision (2012)Quantum Models of Cognition and DecisionLIT-316new seed
van Rijsbergen, The Geometry of Information Retrieval (2004)The Geometry of Information RetrievalLIT-262already held
Coecke, Sadrzadeh & Clark, DisCoCat (2010)Mathematical Foundations for a Compositional Distributional Model of MeaningLIT-273already held
Isham & Butterfield (1998)A topos perspective on the Kochen–Specker theorem: I. Quantum states as generalized valuationsLIT-325new seed
Döring & Isham (2008)A topos foundation for theories of physics: I. Formal languages for physicsLIT-343new seed
Döring & Isham (2008)A topos foundation for theories of physics: II. Daseinisation and the liberation of quantum theoryLIT-335new seed
Döring & Isham (2008)A topos foundation for theories of physics: III. The representation of physical quantities with arrows δ̆°(A): Σ̲ → ℝ̲≽LIT-299new seed
Döring & Isham (2008)A topos foundation for theories of physics: IV. Categories of systemsLIT-310new seed
Abramsky & Brandenburger (2011)The sheaf-theoretic structure of non-locality and contextualityLIT-016already held
Birkhoff–von Neumann 1936The logic of quantum mechanicsLIT-306new seed
Kochen–Specker 1967The problem of hidden variables in quantum mechanicsLIT-298new seed
Stone dualityThe theory of representations for Boolean algebrasLIT-353new seed (reading of a loose reference)
Zaslavsky 1975Facing up to arrangements: face-count formulas for partitions of space by hyperplanesLIT-332new seed
Wick–Wightman–Wigner 1952The intrinsic parity of elementary particlesLIT-339new seed
Haag 1996 (algebraic QFT, Local Quantum Physics)Local Quantum Physics: Fields, Particles, Algebras (2nd ed.)LIT-321new seed
Artin–WedderburnOn hypercomplex numbersLIT-309new seed (reading of a loose reference)
Artin–WedderburnZur Theorie der hyperkomplexen ZahlenLIT-331new seed (reading of a loose reference)
Murota–Kanno–Kojima–Maehara 2010A numerical algorithm for block-diagonal decomposition of matrix *-algebras with application to semidefinite programmingLIT-352new seed (reading of a loose reference)
Peter–WeylDie Vollständigkeit der primitiven Darstellungen einer geschlossenen kontinuierlichen GruppeLIT-329new seed
Pontryagin (duality)The theory of topological commutative groupsLIT-333new seed (reading of a loose reference)
Gelfand (Gelfand–Naimark / Gelfand duality)On the imbedding of normed rings into the ring of operators in Hilbert spaceLIT-313new seed (reading of a loose reference)
Walsh (Boolean Fourier; Walsh 1923)A closed set of normal orthogonal functionsLIT-300new seed
O’Donnell, Analysis of Boolean FunctionsAnalysis of Boolean FunctionsLIT-346new seed
Riesz (representation theorem)Sur une espèce de géométrie analytique des systèmes de fonctions sommablesLIT-326new seed (reading of a loose reference)
Eckart–YoungThe approximation of one matrix by another of lower rankLIT-330new seed
Park, Choe & Veitch 2023The Linear Representation Hypothesis and the Geometry of Large Language ModelsLIT-304new seedANTH-LIT-606
Xiong 2026The Lattice Representation Hypothesis of Large Language ModelsLIT-267already heldANTH-LIT-460
Elhage et al. 2022 (superposition)Toy Models of SuperpositionLIT-323new seed
Huh, Cheung, Wang, Isola 2024 (PRH)The Platonic Representation HypothesisLIT-302new seedANTH-LIT-458
Kornblith et al. 2019 (CKA)Similarity of Neural Network Representations RevisitedLIT-254already held
Cohen–Welling 2016Group Equivariant Convolutional NetworksLIT-314new seed
Bronstein et al.Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and GaugesLIT-319new seed
KondorOn the Generalization of Equivariance and Convolution in Neural Networks to the Action of Compact GroupsLIT-305new seed (reading of a loose reference)
Engels-style non-linear featuresNot All Language Model Features Are One-Dimensionally LinearLIT-322new seed
McCandlish–Kaplan–Amodei 2018An Empirical Model of Large-Batch TrainingLIT-348new seedANTH-LIT-017
Shwartz-Ziv–Tishby 2017Opening the Black Box of Deep Neural Networks via InformationLIT-324new seedANTH-LIT-508
Tishby IBThe information bottleneck methodLIT-338new seed
Xu–Raginsky 2017Information-theoretic analysis of generalization capability of learning algorithmsLIT-347new seed
Landauer 1961Irreversibility and Heat Generation in the Computing ProcessLIT-328new seed
Shannon limit E_b/N0 ≥ ln2Communication in the Presence of NoiseLIT-317new seed (reading of a loose reference)
Still–Sivak–Bell–Crooks 2012Thermodynamics of PredictionLIT-327new seed
Goldt–Seifert 2017Stochastic Thermodynamics of LearningLIT-308new seed
Van Trees Detection, EstimationDetection, Estimation, and Modulation Theory, Part ILIT-351new seed
Cramér–RaoInformation and the Accuracy Attainable in the Estimation of Statistical ParametersLIT-349new seed
Cramér–RaoMathematical Methods of StatisticsLIT-311new seed
Watanabe SLT (2009)Algebraic Geometry and Statistical Learning TheoryLIT-354new seed
Power 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic DatasetsLIT-341new seedANTH-LIT-538
Nanda 2023Progress measures for grokking via mechanistic interpretabilityLIT-345new seedANTH-LIT-085
Gärdenfors, Conceptual Spaces (2000)Conceptual Spaces: The Geometry of ThoughtLIT-334new seed
Wille, “Boolean Concept Logic” (2000)Boolean Concept LogicLIT-337new seed
Ganter–Wille 1999Formal Concept Analysis: Mathematical FoundationsLIT-344new seed
Lyons Semantics (1977)Semantics, Volume 1LIT-307new seed
Aristotle (antonym/contrary/contradictory)Categories and De Interpretatione (Clarendon Aristotle Series, trans. and notes J. L. Ackrill)LIT-350new seed (reading of a loose reference)
Ryle 1949 (category mistake)The Concept of MindLIT-318new seed
Sommers 1963 (sortal trees)Types and OntologyLIT-303new seed
Worrall 1989Structural Realism: The Best of Both Worlds?LIT-315new seed
Ladyman 1998What is Structural Realism?LIT-342new seed
Ladyman–Ross, Every Thing Must Go (2007)Every Thing Must Go: Metaphysics NaturalizedLIT-301new seed
Dennett “Real Patterns” (1991)Real PatternsLIT-220already held
Armstrong (immanent universals)Universals and Scientific Realism (Vol. I: Nominalism and Realism; Vol. II: A Theory of Universals)LIT-336new seed (reading of a loose reference)
Rodriguez-Pereyra (resemblance nominalism)Resemblance Nominalism: A Solution to the Problem of UniversalsLIT-320new seed

Close readings of the prior-art novelty map’s references

2026-09-29 23:06:37 · curation · record

The owner asked for readings of the 57 Deferred seeds that registering the prior-art novelty map added, wherever a work could be found to read.

How it was done. Eight agents split the seeds by subject: quantum cognition and structural realism; thermodynamics of learning; topos quantum theory (Döring–Isham, Isham–Butterfield); classic mathematics and physics; interpretability; equivariance and Boolean analysis; learning dynamics and information theory; and the philosophy and textbook tail. Each worked only from legitimate copies: arXiv, publishers’ free archives, GDZ, Math-Net.Ru, a university preprint series, and Internet Archive scans of journal issues in the US public domain (renewal records checked). Shadow libraries and unauthorised uploads were refused even where they were the only copy found.

The result. 36 works were read and 21 could not be reached.

Zaslavsky’s memoir (LIT-332) matters most of these: the map relies on it for topes, and neither the AMS copy (blocked) nor the MIT thesis was reachable.

Metadata corrections from the readings.

  • published: moved to first appearance (ADR-002):
  • Stone 1936 (LIT-353) takes its printed title, plural “Representations”.
  • Gelfand–Naimark (LIT-313) gains the Math-Net.Ru URL of the 1943 original, since the DOI is the 1994 reprint. Toy Models (LIT-323) gains its transformer-circuits URL.
  • Tags:
    • LIT-343 gains philosophy-of-science (paper I is a programme paper).
    • LIT-310 swaps contextuality for mereology and philosophy-of-science, since paper IV is about composite systems, not contextuality.
    • LIT-348 (McCandlish) loses information-theory: the paper has no entropy, mutual information or channel.

What the readings say about the map. These findings are for the map’s owner to act on. Nothing in the map was edited.

  • Row 8 (the Stone leg).
    • Isham–Butterfield I (LIT-325) already owns the chain from Kochen–Specker to Stone duality (§2.3).
    • Stone 1936 (LIT-353) contains no topology, Stone space, duality or ultrafilters; those are in Stone 1937, which is not registered.
    • In the finite case, ultrafilters are atoms, which may be faces rather than only topes.
  • Orthomodularity and negation.
    • Birkhoff–von Neumann (LIT-306) propose the modular law, not the orthomodular one. Citing them for “orthomodular” is a misattribution.
    • Döring–Isham use a Heyting negation, which clashes with the map’s orthocomplement rows, and give no finite witness.
    • Kochen–Specker (LIT-298) prove dimension 3 only. They never state “dimension ≥ 3”.
  • Row 7 (predicate harmonics and irreps).
    • “Harmonics = characters” is textbook abelian harmonic analysis (O’Donnell, LIT-346), so KNOWN, not SYNTHESIS. Walsh 1923 (LIT-300) has no characters at all.
    • “Multiplets = degeneracies” is the Schur corollary. Kondor–Trivedi (LIT-305) supply the machinery but never state it, and Peter–Weyl (LIT-329) does not state the converse “degeneracy implies symmetry”.
    • Abelian characters give no multiplets. The cube’s degeneracies come from the nonabelian signed-permutation group, which the equivariance reader checked numerically.
    • The equivariance papers (LIT-314, LIT-305, LIT-319) and Engels et al. (LIT-322) all take the group as given. So “they impose; you detect” holds, and Engels’ weekday and month circles are the natural validation target.
    • Murota et al. (LIT-352), the nearest algorithmic precedent, has a Theorem 3.1(C) that fails over ℝ, the reader checked numerically. It also has no noise model.
  • Row 12 (gradient as channel).
    • McCandlish (LIT-348) owns batch = gradient SNR and the critical batch. His progress law saturates, whereas the map’s log(1 + SNR) keeps growing.
    • The channel frame belongs to Xu–Raginsky (LIT-347), not McCandlish.
    • Shwartz-Ziv–Tishby (LIT-324) is uncited prior art for “gradient SNR organises training”.
    • The IB paper (LIT-338) says nothing about gradients.
  • Row 13.
    • Goldt–Seifert (LIT-308) and Still et al. (LIT-327) bound entropy production. They do not support a kT ln 2 heat floor per judgment, since heat is due only at erasure.
    • The three “temperatures” (McCandlish’s ε/B, Gibbs β, IB β) are formal analogies, none of them k_B·T.
  • Rows 2 and 5.
    • Eckart–Young (LIT-330) is unique as a matrix only when λ_r > λ_{r+1}, and the factors are never unique.
    • Riesz 1907 (LIT-326) states the representation without proof or uniqueness.
    • Park et al.’s (LIT-304) causal inner product is fixed only by an arbitrary choice.
    • None of the three supports a “Riesz-forced unique factorisation”.
    • Wick–Wightman–Wigner (LIT-339) make cross-sector superpositions mixtures, not truth-value gaps.
  • Structural realism. Worrall 1989 (LIT-315) is epistemic structural realism, so it is the wrong citation for “to be is to be an invariant of structure”. Ladyman 1998 (LIT-342), the ontic source, was unreachable.
  • Aerts–Gabora (LIT-340, LIT-312). Their models are single-basis, so the map’s “non-distributive” gloss should go.
  • Platonic Representation Hypothesis (LIT-302). Its measured convergence is small (a mutual-kNN score of 0.16), and the proof shows only a shared kernel. The map’s “PRH-measured” needs these qualifiers.
  • Toy Models (LIT-323). The reader’s own inference, checked numerically, is a bridge to the map: in the efficiently packed shapes the features form a tight frame, a rank-one POVM whose weights are the paper’s feature dimensionalities. That would make superposition a compressed Boolean context, a candidate reading of “Boolean shadow”.

Disagreements with the anthology (ADR-013). Several readings found faults in the anthology’s entries for the six dual-held papers. For example:

  • The anthology’s Park et al. entry reports “four quadruples” where the paper has one.
  • Its Platonic Representation Hypothesis entry omits the 0.16 ceiling.
  • Its grokking entries do not record that the headline run used no weight decay, or that dropout also produces grokking.
  • Its McCandlish entry claims an “order of magnitude” fit that fails for the SVHN models.

Each is in the NOTE concerned, reported and not fixed. They are for an anthology pass to take up.

Traps. The filing script built regex replacements from reading text, and a backslash in a summary broke it partway through. Use a function as the replacement when the text is not your own.

Supplemental readings for the prior-art novelty map’s misattributions

2026-09-29 23:37:04 · curation · record

The close readings of the prior-art novelty map’s references (record/curation.d/2026/09/29/230637.md) found several places where the map cites the wrong work or a work that could not be reached. They were recorded as findings but not followed up. The owner asked for the follow-up. This batch registers and reads the works the map should cite instead.

How it was done. Three agents shared one brief, split as follows:

  • Stone 1937, Husimi, Zaslavsky’s announcement and Stanley;
  • the two symmetry-detection papers;
  • Russo & Zou, Bennett and Earman.

Each gave a verdict on whether its work supplies what the map needs. Only legitimate sources were used: arXiv, J-STAGE, an author-linked PDF, and the Internet Archive’s public-domain scan of Trans. AMS 41.

Result. Nine works were registered. Seven were read and two were unreachable.

Row 8: Stone duality and topes.

  • Stone 1937 (LIT-358) is the source of the topological duality, so row 8 should cite it in place of Stone 1936 (LIT-353). Its Theorems 1, 2 and 4 give the prime-ideal space as a totally disconnected, locally compact Hausdorff space, compact exactly when the ring has a unit. Stone calls this a “mathematical equivalence”, not a duality, and never states it for morphisms. It says nothing about ultrafilters, hyperplanes or regions. The paper also contains Stone–Čech, Stone–Weierstrass and Banach–Stone. Stone’s link to character-group duality (p. 393) is a conjecture in the text, not a result. LIT-353 now points at it.
  • Stanley’s arrangements notes (LIT-355) are the readable proof of Zaslavsky’s theorem that the record lacked. They show that region and face counts depend only on the intersection poset. The notes never use “topes” and do not cite Zaslavsky, so they are cited alongside LIT-332, not instead of it.
  • Zaslavsky’s announcement (LIT-359) is open access at the AMS, but every route to it was blocked from this session. It stays Deferred, and anyone with ordinary web access can fetch it.
  • “Ultrafilters ↔ topes” still has no source. It needs an oriented-matroid text plus the owner’s own argument, as NOTE-272 and NOTE-290 said.

Orthomodularity. Husimi 1937 (LIT-357) derives the Birkhoff–von Neumann lattice from statistical premises. It uses the orthomodular law as an unnamed proof step, but never states or names it. The usual credit to Husimi is therefore a reading of his axioms. NOTE-275 now carries that correction. A source that states the law as an axiom (Kalmbach, or the Piron line) is still missing. Husimi also raises the “meaningless proposition” (truth-value gap) reading (p. 769) and declines it. That bears on row 5.

Row 7: “they impose; you detect”.

  • Lenc & Vedaldi 2015 (LIT-363). They fit linear maps on a trained network’s feature space that realise named image transformations. So symmetry has been measured in learned representations, and the group acts on representation space. That weakens the “where the group acts” distinction drawn in NOTE-274, NOTE-282 and NOTE-288.
  • Cohen & Welling 2014 (LIT-362). They learn the irreps of a commutative group from pairs related by an unknown element, but the group class is fixed in advance. This is the “learn irreps from data” prior art. The map’s “Cohen–Welling” is the 2016 paper, which imposes the group.
  • What survives as novel. Both papers need paired inputs, and either a hypothesised transformation or an assumed group class. Neither looks at an operator’s spectrum, so degeneracy plays no role in either. The owner’s claim survives, narrowed to where the group comes from: eigenvalue multiplicities, with no pairs and no hypothesised group.
  • Two corrections to earlier notes. NOTE-282 said abelian symmetries give “no degeneracy”. That is false over ℝ for SO(2) and ℤ/n with n ≥ 3, whose rotation irreps force pairs; that is Cohen & Welling 2014’s setting and Engels et al.’s (LIT-322). NOTE-288’s relay of Lenc & Vedaldi overstates that paper. Both notes now carry dated corrections.

Row 12: the learner as a channel. Russo & Zou (LIT-356) is the original of the bound that Xu & Raginsky (LIT-347) extend, so row 12 should cite both.

  • Their Lemma 1 chains information over adaptive steps, and it is the origin of LIT-347’s eq. (36).
  • Their Lemma 2 prices one noisy query at ½·log(1 + SNR). That is the nearest existing result to the map’s C_step = d·log₂(1 + SNR). It covers queries answered with added independent noise, not minibatch gradients, so it cannot be cited for SGD without a further argument.
  • NOTE-306 stated their quantity in Xu & Raginsky’s notation, and now says so.

Row 13: the heat floor. Bennett (LIT-360) settles it against the map.

  • Only logically irreversible steps carry a cost, and the cost is entropy in the environment, which “need not” be heat.
  • Bennett calls “an intrinsic cost of order kT for every elementary act of information processing” a misconception.
  • The defensible form is “k_B T ln 2 of entropy per predicate bit erased or overwritten”. This agrees with NOTE-295 and NOTE-296.
  • Bennett gets the title of Landauer 1961 wrong, so cite him for the principle’s scope and not as a substitute for LIT-328.

Row 5: superselection and truth-value gaps.

  • Earman 2008 (LIT-361) is paywalled and is not on PhilSci-Archive. The only copy found was a draft reposted in a third party’s reading-group folder, which was not used.
  • The truth-value-gap reading still has no source that asserts it. Husimi raises and declines it, and Wick–Wightman–Wigner make cross-sector superpositions mixtures (NOTE-304).

Structural realism, which needed no new reading. The earlier curation entry said only the unreachable Ladyman 1998 (LIT-342) could stand in for Worrall 1989 (LIT-315) on ontic structural realism. That was wrong: the record already holds ontic sources, read earlier. They are:

  • LIT-217 (Ladyman, “Science, Metaphysics and Structural Realism”);
  • LIT-219 (Ladyman & Ross, “Patterns All the Way Up”);
  • LIT-218 (OSR applied to economics);
  • LIT-222 and LIT-223 (the category-theoretic debate over radical OSR).

The map’s “to be is to be an invariant of structure” should be sourced there.

Traps. The Wayback copy of Zaslavsky’s announcement was refused by this session’s egress policy rather than by the archive. A session with different network settings may reach it directly.

Signal detection theory, registered and read

2026-09-30 01:55:42 · curation · record

The owner asked whether signal detection theory (SDT) is covered in this record. It was not. The only related entries were Van Trees 1968 (LIT-351, the owner’s prior-art map cites it for row 14’s detection model, P(flip) = Q(γ/σ) with a Cramér–Rao bound), Rao 1945 (LIT-349) and Cramér 1946 (LIT-311). All three are Deferred and unreachable. This batch registers the founding and standard SDT sources and reads the ones that could be reached.

How it was done. Two agents shared one brief.

  • One took the statistics and engineering line: Neyman & Pearson 1933, and Peterson & Birdsall.
  • The other took the psychophysics line: Tanner & Swets 1954, Green & Swets 1966, and an open tutorial on the measures.

Each also judged whether its work could stand in for Van Trees on row 14. The sources used were a public-domain Internet Archive scan of the November 1954 Psychological Review, the publisher’s own free PDF, and DTIC scans mirrored in the Internet Archive’s US-government documents collection.

Result. Five works registered: three read and two unreachable.

  • Peterson & Birdsall 1953 (LIT-364), Skimmed.
    • What was read. The University of Michigan technical report, Parts I and II, registered as the report. The 1954 IRE paper with Fox is its journal version, given in the venue. Fox is thanked in the report, and becomes an author only in 1954.
    • Findings.
      • One receiver that outputs the likelihood ratio is optimal under the payoff criterion, the Neyman–Pearson criterion and the posterior approach alike.
      • It introduces the ROC and proves that its slope equals the threshold.
      • In white Gaussian noise the log-likelihood ratio is equal-variance normal with detection index d = 2E/N₀, and the optimal receiver is a matched filter.
      • Its index d is the square of the later d′. It never uses d′ or “ideal observer”.
  • Tanner & Swets 1954 (LIT-368), Read. The founding psychophysics citation.
    • It names d′ (as √d, after Peterson & Birdsall), separates sensitivity from criterion, and draws the ROC.
    • It gives the optimal criterion β = prior odds × payoff ratio.
    • With three observers it predicts four-alternative accuracy from yes-no d′, and rejects the “false alarms are guesses” threshold model. The paper says “four” observers once, but only three appear in the data.
  • Stanislaw & Todorov 1999 (LIT-367), Read. The compact source for the formulas: d′, β, c, A′, B″, A_z, and 2AFC accuracy Φ(d′/√2), each with its assumptions. The reader checked the worked example by hand. It gives no d′ formula for m-alternative choice with m > 2.
  • Neyman & Pearson 1933 (LIT-365), unreachable. OpenAlex marks it open access, but every Royal Society URL returned a challenge page, and the Wayback copy was blocked by this session’s egress policy. The only other copies were reposts, which were not used. A session with ordinary web access should be able to read it.
  • Green & Swets 1966 (LIT-366), unreachable. In copyright, and the only copy is controlled lending. Its identifier is an Open Library URL, since the book has no DOI.

What the record now says about SDT. The core of SDT now rests on sources read first-hand: the likelihood-ratio decision rule, the ROC, the equal-variance Gaussian model, and the split between sensitivity (d′) and criterion (β, c).

  • The engineering statement is Peterson & Birdsall.
  • The psychophysical application is Tanner & Swets.
  • The measures are Stanislaw & Todorov.

The likelihood-ratio lemma itself is still second-hand until Neyman & Pearson is read. The only thematic link to earlier holdings is LIT-099 on framing effects in perception.

Row 14 of the map.

  • The detection half can now cite Peterson & Birdsall instead of Van Trees. The equal-prior error Q(√d/2) = Q(d′/2) follows in one line from its eqs. 4.3–4.8, but it is never printed. So it is cited as derived, not quoted. LIT-351 now says so.
  • Tanner & Swets and Stanislaw & Todorov supply the tail-at-criterion form. Error rates are Gaussian tails at a criterion measured in noise SDs, which is the Q(γ/σ) form, and Stanislaw & Todorov print the 2AFC result.
  • The Cramér–Rao half still has no reachable source. None of the three has estimation theory.

Traps. Royal Society and DTIC pages sit behind challenges from here, even for open-access items. The Internet Archive’s government-documents mirror of DTIC scans is the legitimate route for pre-1960 US government technical reports.

Theory backfill: twenty-five accounts drawn from the record’s readings

2026-09-30 03:09:05 · curation · record

The owner pointed out that the record held hundreds of read works and very few theories. The count was 17, not nine, but the point stood. THEORY-001 to 009 are on contrastive-learning kernels, 010 to 016 on contextuality, and 017 on Hilbert-space intrinsic structure. Twenty-one topics had no theory at all, among them philosophy of science (about 40 read works), ethics (about 20), society and governance (about 22), agency, consciousness, metaphysics, learning theory and probabilistic modelling.

How it was done. Five agents drafted theories under one brief, each in its own domain:

  • information and learning theory, with thermodynamics;
  • representation learning and symmetry;
  • logic, quantum logic and probability;
  • agency, ethics and mind;
  • philosophy of science, metaphysics and society.

The brief had four rules:

  • draw only from readings the record already holds, with no new reading;
  • every source: must be a read work, never a Deferred one;
  • default to Proposed unless the readings settle the claim;
  • a Rejected account is welcome when a reading refutes it.

The drafts were reviewed before filing.

  • Titles. Titles over about 280 characters were shortened to the claim. The qualifications stay in the body.
  • Dropped. One draft was dropped: the Arrow/Paretian-liberal contrast duplicated THEORY-027.
  • Rejected accounts. The two Rejected drafts are restated so that the title is the account being rejected, matching how the record’s other statuses read.

Midway, the owner asked for papers-feed evidence on learning phases in the information-bottleneck (IB) sense. Four papers were registered and read in this PR:

Saxe and Zucchet are also held in the anthology as ANTH-LIT-509 and ANTH-LIT-450 (ADR-013). THEORY-035 now cites Saxe first-hand, and a phases theory (THEORY-039) came out of the four readings.

The twenty-five, by domain.

  • Information, learning and thermodynamics.
    • THEORY-030: the scope of Landauer’s principle.
    • THEORY-026: nonpredictive memory dissipates, and with feedback, prediction is not what efficiency requires.
    • THEORY-028, Active: information-usage bounds on bias and generalisation.
    • THEORY-035, Rejected: IB compression as the reason deep networks generalise.
    • THEORY-033: the critical batch as gradient noise-to-signal.
    • THEORY-039: the later phases of training differ in kind, and none is a measured compression of I(X;T).
  • Representation and symmetry.
    • THEORY-019: symmetry forces degeneracy, but degeneracy does not identify a symmetry.
    • THEORY-021: packed superposition as a compressed POVM.
    • THEORY-022: the grokking circuit and language-model circles, where the group comes from the task.
    • THEORY-018, Active: the softmax identifiability gauge, and the causal inner product as a stipulation.
  • Logic, quantum logic and probability.
    • THEORY-020, Active: Stone representation, with the Kochen–Specker obstruction lying in how contexts overlap.
    • THEORY-041, Active: Birkhoff–von Neumann proposed modularity, not orthomodularity.
    • THEORY-042, Active: superselection makes cross-sector superpositions mixtures.
    • THEORY-031, Active: likelihood-ratio optimality, and when d′ is criterion-free.
    • THEORY-037: the topos logic is Heyting.
    • THEORY-024, Rejected: sheafifying contextual data restores Boolean logic.
  • Agency, ethics and mind.
    • THEORY-029: the regress/authority objection to hierarchical accounts.
    • THEORY-040: the manipulation objection.
    • THEORY-025: conversion and adaptation defeat utility and resource metrics.
    • THEORY-027, Active: the Paretian liberal, unlike Arrow, survives interpersonal comparison.
    • THEORY-023: whether AI is conscious is underdetermined by current evidence.
  • Philosophy of science and metaphysics.
    • THEORY-032, Active: what the Yoneda lemma does and does not say about relational identity.
    • THEORY-038: category theory does not support radical ontic structural realism.
    • THEORY-034: the reach of Chakravartty’s dilemma.
    • THEORY-036: Dennett’s real-pattern criterion versus what rainforest realism adds.

In total: 9 Active, 14 Proposed and 2 Rejected. Several extend THEORY-017.

What the four phase papers showed.

  • Saxe et al. refutes all three IB claims as general statements:

    • the information-plane phases depend on the nonlinearity and on the estimator;
    • compression and generalisation dissociate in all four combinations;
    • full-batch gradient descent compresses as SGD does.

    The drift-to-diffusion transition in gradient signal-to-noise is real, but it is generic.

  • Zucchet et al. show three phases in fact learning: marginals, then a plateau while the recall circuit forms, then storage. That is the reverse of Nanda et al.’s memorisation, then circuit, then cleanup (LIT-345).

  • Prakash & Martin report a late collapse in generalisation, tracked by a spectral weight metric. Their headline claim that α < 2 heralds the collapse is contradicted by their own weight-decay control.

  • Zhou et al. show a chaos-to-confinement transition: a parameter perturbation stops changing the run’s basin, and the tangent kernel settles into a cone.

None of the four measures a network-intrinsic fall in I(X;T). A circuit forming during a flat loss is shown by intervention only in Nanda and Zucchet.

Corrections found on the way.

  • Reported to the anthology per ADR-013, not fixed.
    • ANTH-LIT-509 says Saxe has no open route; the PDFs are linked from co-author Kolchinsky’s page. It also overstates the “three independent estimators”.
    • ANTH-LIT-450 gives an optimal imbalance exponent “between 1 and 2”, which is not in the paper: the sweep covers 0–1, and the optimum is 0.6–0.8.
  • In this record. The agency drafter found that the 2026-09-28 Sen curation entry is wrong to call conversion factors second-hand. Sen 1979 asserts them (p. 219), and the 1998 lecture names four sources of them (pp. 194–195). THEORY-025 records this.

Traps. Draft theories that refer to each other cannot use codes until they are filed. Name them by draft id, then substitute the temporary codes after luria new theory, as was done here.

Backlinks