{"path":"research/the-literature-as-a-distributed-data-structure.md","content":"# The literature as a distributed data structure: Dan Roy's resignation read against the corpus\n\n**Date**: 2026-11-04 · **Type**: a public post read against the corpus, with outside research, tests on real cases and recommendations (the all-angles pass) · **Status**: Claude's reading; nothing here is ruled; the decisions only he can make are ledger Arc BG.\n\n**His request, verbatim**: *\"Ground extremely deeply on this. Then relate it to Deliberus and its potentialities. Ground extremely deeply on related parts of the corpus. Consult the corpus graph wisely. Ultrathink\"*, with a screenshot of the post.\n\n**The post, verbatim** (Dan Roy, @roydanroy, 2 November 2026, 1:24 PM in the screenshot's own timezone, 45.3K views at capture):\n\n> Just fully* resigned from JMLR. I still believe in science, but journals needs a complete rethink. AI reviews by top models are now much better than time-pressed human ones. OTOH, one-time gate keeping (accept / reject) doesn't make sense. The literature is going to be moving faster than ever. We need to think about it as a distributed data structure, and start over.\n\n**His asterisk, three minutes later**: *\"\\*Fully because I actually resigned ~2 years ago, and have been dealing with resubmissions.\"*\n\n**Two replies of his the same day and the next, which the post leaves out and which decide how to read it** (his X profile as mirrored by a news aggregator, read 2026-11-04):\n\n- To a reader asking how AI reviews can be good, 2 November 15:21: *\"Context matters. My chatbots have seen dozens of my reviews, etc., and I've given them a lot of guidelines in the past, and so when I ask them to assess work, it's always pretty good.\"*\n- 3 November 08:00: *\"At 1 paper a month, I don't think quality would suffer. One can easily work that much faster with top models. 1 paper a day, for sure, you would be getting a lot of junk. But more, who is consuming this? That's the real question. I suggest we all up our ambition and aim for very impactful problems. And then solve them as quickly as possible.\"*\n\n---\n\n## 0. In short\n\n*The answer in five lines is § 0b, at the end of the note, after the outside research and the tests it rests on.*\n\n## How this was done\n\nThe six parts of the all-angles pass, each done or skipped with its reason:\n\n1. **The question pinned** (§ 1): the post quoted whole; which of its four claims the corpus already rules on and which is fresh. Darwin was off the network this session, so `mannaminne` and the corpus graph's meaning search were unavailable; the inventory was made by file (`rg`), by `corpus_graph.py around` on each home note, and by reading.\n2. **The corpus inventoried before any research** (§ 2): 22 notes read whole this session, listed there by name, plus the ledger rows named in § 1 read whole.\n3. **Research only on what the corpus lacks** (§ 4): two helpers on Sonnet, one on the post's context and the measured state of AI review, one on what *a distributed data structure* has meant in practice and on real cases with published ground truth; their reports verbatim in `the-literature-as-a-distributed-data-structure/`. Five of their saved sources were read whole by the session as well (the TMLR post of 15 October, the arXiv-cap abstract and news lead, ClaimRxiv, JMLR's own history page, the 2001 resignation letter).\n4. **Theory first, then tests** (§ 3, § 5): what the corpus settles on each claim is written before the outside evidence; the tests are real cases with known outcomes, from the corpus and from outside, with the arithmetic recomputed in `cases.py` beside this note.\n5. **Every conclusion set against what is settled** (§ 6): the ruled rows each recommendation touches, and the outside check run on the draft.\n6. **Persisted, then reported** (§ 7, § 8): pointers from the notes this answers, the README index, the ledger rows.\n\n**Files read whole this session**: `peer-review-and-the-reasoning-layer.md`, `the-missing-layer.md`, `ai-slop-and-the-reasoning-layer.md`, `carriers.md`, `science-reasoning-infrastructure.md`, `the-scrutiny-gap.md`, `machine-work-as-a-standing-property.md`, `alien-proofs-and-the-library.md`, `proving-a-pass-reliable.md`, `the-evals-loop-and-our-machine-passes.md`, `fragile-checkers-and-the-verification-bottleneck.md`, `earning-the-right-to-write.md`, `what-confirmed-means.md`, `agents-as-a-consumer-class.md`, `understanding-after-the-proof-machines.md`, `science-lineage-philosophy.md`, `the-ratification-record-and-the-triage-feed.md`, `derivation-sciencepedia-and-the-logic-core.md`, `the-harness-conviction.md`, `lean-social-system-research.md`, `the-reuse-flywheel-prior.md`, `drowning-in-claims.md`, `saturation-and-the-long-tail.md`; in part: `the-residual-error-taxonomy.md` (class 6), `signposts-for-the-whole-project.md` (§ 4), `what-deliberus-is-for.md` (§ 3, § 8.6), `incentives-analysis.md` (§ 2), `the-agent-population-and-what-checking-is-up-against.md` (§ 2d–2e), `fractal-priors-for-success.md` (§ 5), `societal-integration-outlook.md` (§ 1), `what-he-held-in-the-group-2011-2013.md` (§ 1.14), `lean-deliberus-analogies.md` (§ 7), `competitive-landscape.md` (Nanoarguments), `TODO.md` (the order of work). Ledger rows read whole: Q32, R7, R8, R12, T1, T3, Y9, Z1, Z11, AF5, AS1, AS2, AV2, BF1, BF3.\n\n---\n\n## 1. The question, pinned\n\n### 1a. Who, and what the words refer to\n\n**Dan Roy** is at Google DeepMind, on leave as a Canada CIFAR AI Chair, former Research Director of the Vector Institute, and Professor at the University of Toronto in Statistics and Computer Science (his profile text, read 2026-11-04). His homepage lists him as an **Action Editor** of JMLR (read 2026-11-04). The post is a statement by a senior insider of the venue it resigns from, with positional knowledge of what reviewing there costs.\n\n**JMLR was itself born of a resignation.** Leslie Kaelbling founded it in 2000 as a free-to-read journal; in 2001 forty editorial board members of Kluwer's *Machine Learning* resigned publicly over restricted access (an institution paid $1,050 a year) and moved to JMLR. The letter's argument was about **distribution**; it kept the gate as it was: *\"the editorial board, which also serves its traditional function of ensuring rigorous peer review of journal articles\"* (the resignation letter, SIGIR Forum 35(2), read whole; `jmlr.org/history`, read whole). So *start over* has a precedent in the same journal twenty-five years ago, and that start-over changed who owns the pipe and left the gate untouched. Roy's is the first in that lineage aimed at the gate itself.\n\n**TMLR, the continuous-submission journal in the same family, published what it is doing on 15 October 2026** (the Editors-in-Chief's post, read whole). In one year: submissions up 300%; solo-authored papers up 1,300%; 159 authors told their papers will not be sent out for review for a year; desk rejections from 6% in 2023 to about 53%; author quotas by coauthorship since August, already circumvented by second accounts and by dropping the over-quota author. And **since October 2026 every submission gets an AI review that** *\"evaluates whether the claims made in the paper are met with convincing evidence. It does not evaluate interestingness, which is left to human reviewers.\"* The action editor decides; the Editors-in-Chief approve. That split, soundness to the machine and interest to people, is the fact behind Roy's second sentence, and it is already in production in his own journal family.\n\n**arXiv capped submissions at two papers a month per submitting researcher from 1 October 2026** (Nature news, 15 October, the lead paragraph; the rest paywalled). Submissions reached a record 40,363 in September 2026; solo-authored papers rose 70% in a year in mathematics and condensed matter; the break in the thirty-year 9%-a-year trend sits between December 2025 and February 2026, which is when models capable of extended autonomous work shipped (Contaldi, arXiv:2610.34449, abstract, 25 October 2026; the analysis was itself drafted by an agent under the author's direction). That is *\"the literature is going to be moving faster than ever\"* with a number on it.\n\n**ClaimRxiv** (Kejriwal, SEMANTiCS 2026, a blue-sky paper published 13 October 2026, read in part) is the nearest outside statement of the fourth sentence: *\"Papers are the wrong unit of computation\"*; scientific assertions as first-class objects with scope conditions, evidence links, provenance, uncertainty, support and challenge structure, and temporal evolution; and *\"a reasoning engine for science … combining symbolic semantics with defeasible, provenance-aware, and evidence-weighted reasoning\"*. Nothing is built; the paper is the Semantic Web community proposing the move to itself, *\"along the lines of services like arXiv\"*. The corpus did not hold it (searched 2026-11-04: no mention).\n\n### 1b. The four claims, and what the corpus already rules on\n\n| His sentence | Ruled here | Open here | Fresh |\n|---|---|---|---|\n| *journals needs a complete rethink* | the reviews outrank building (2026-08-29); *contributions are added, never overwritten* (D6); the Legibility Rule stays a candidate (never ratified) | the eight places the peer-review note names, none ruled | whether what is happening at JMLR, TMLR and arXiv this autumn changes the project's own order |\n| *AI reviews by top models are now much better than time-pressed human ones* | propose-never-assert; R7 (a second model is not a second observation); T1 (review need not keep pace with a **proven** pass); Y9 (a person reads before anyone proves); *an error rate belongs to the task as written*; AM12 (a pass is taught the method, never the contents); T8 (proven carries its bound and scope) | T3 (how a pass is proven: the procedure); AF5 (the mark opens the pass's faults) | whether *better* has a measurement behind it, and what kind of review it is true of |\n| *one-time gate keeping (accept / reject) doesn't make sense* | R8 (standing work, with a brake); R12 (a later reading retires); AS1 (never bottlenecked); AS2; the no-ratify-act end-state (2026-08-31); Q32 (confirmed retired); confident housekeeping (2026-09-13) | the triage feed's work half (AW2 to AW9) | whether the corpus's *gate on weight, not on existence* answers him, and where the queue goes when the gate goes |\n| *think about it as a distributed data structure, and start over* | the missing layer (the unit is reasoning); always-mint; the record of machine judgments (Z1, Z11); read-only agents; BF1 (the second rung) | R7's structural proposal (independence as structure); Z8 (a computed door) | what *data structure* has to mean for it to work, which the post does not say |\n\n**The open half in one sentence**: Roy names three moves, replace the reviewer, remove the gate, change the unit; the corpus has ruled on the second and third in substance and refuses the first as stated, so the question this pass serves is **in what form an AI review can enter a continuous, claim-level structure without being the judge the record refuses.**\n\n**Unruled decisions that bear, found before the pass rested on them**: the Legibility Rule (candidate since July, never put to him); the component claim about the Anthropic coordination work (unruled positioning, `the-scrutiny-gap.md` § 1a); the eight places of the peer-review note; the pre-submission self-review item in `TODO.md` (open since August); the triage feed's work half (Arc AW). None is assumed settled below.\n\n---\n\n## 2. What the corpus already holds\n\nAngle by angle; *read whole* means this session, by name; a note only searched is written *not read*.\n\n| Angle | Notes read whole | What each holds |\n|---|---|---|\n| Peer review's history and what reform left untouched | `peer-review-and-the-reasoning-layer.md` | Peer review is younger than television (*Nature* 1967); Peters and Ceci 1982 (8 of 9 resubmitted papers rejected); the unbundling table (arXiv 1991 → PLOS ONE → PubPeer → Registered Reports → OpenReview → eLife 2023 → PREreview): every reform made criticism more visible, none made it structured; the reviewer's own argument is the one thing never unbundled from prose; eight places Deliberus could act, pre-submission self-review first since it needs no institution; **a legible review record is a weaponizable one**; run 5: a referee report hand-extracted into 22 claims, six kinds of move, zero attack links (verdict and critique orthogonal) |\n| Science's own attempts at a claim layer | `science-reasoning-infrastructure.md` | nanopubs (85,720 in July 2026), micropublications, scite (1.4 billion citation statements), ORKG, Discourse Graphs (*\"GitHub for scientific communication\"*), Polymath, Registered Reports (96% vs 44% positive results), open review (eLife: 2,853 submissions in 2025, down from 6,396), argument mining; the cross-case pattern: input friction, wrong moment, wrong consumer, wrong granularity, no computation over disagreement, no bridge from prose, no self-audit |\n| Why Deliberus claims an empty slot | `the-missing-layer.md`, `science-lineage-philosophy.md` | conclusions, quality control, agreement-shape, procedure exist; reasoning does not; *science made empirical claims publicly correctable, Deliberus makes reasoning publicly correctable*; Strevens's Iron Rule as the strongest objection and the Legibility Rule as the candidate answer; Longino's uptake condition as the thing peer review makes least visible |\n| The flood, measured | `ai-slop-and-the-reasoning-layer.md`, `carriers.md`, `fractal-priors-for-success.md` § 5, `societal-integration-outlook.md` § 1 | NeurIPS doubled; arXiv slop rejections 4% → 10–12%; OSF closed; Pangram: half of ICLR reviews machine-assisted, a fifth machine-entire; white-font injections; § 4's six consequences of changing the unit (phantom citations impossible, the flood dies because novelty becomes a diff, **review becomes capital**, the loop breaks on lifecycle, the mill's payoff removed, replication routable), all gated on sameness; Tier 1 is *the drowning carrier*; checking infrastructure is born of floods (Csiszar); every precedent took decades |\n| Machines judging machines | `the-scrutiny-gap.md`, `fragile-checkers-and-the-verification-bottleneck.md`, `the-harness-conviction.md` | persuasion beats experts by throughput; the checking apparatus is being handed to the producers; a judge that learned the marker is monotonic; our own quality score removed because it had no rubric; Jagged Judges: 25–71% of verdicts flip under pressure, 56–70% of flips move away from truth; verdict-as-record vs verdict-in-channel; jury majority strength across families predicts fragility (|ρ| 0.59); the harness conviction (quality is a function of harness and grounding, architecture-independent) and its ceiling (coherent absence, frame lock-in) |\n| What *proven* means for a pass | `proving-a-pass-reliable.md`, `the-evals-loop-and-our-machine-passes.md`, `earning-the-right-to-write.md` | a bound, never a zero (59 clean blind items for 5%, 149 for 2%); review need not keep pace, depth compounds (T1); three kinds of truth (checkable, consensus, contested); twelve devils (correlated errors: two models wrong together ~60% vs 33% chance); abstention buys the guarantee (56–68% answered at 80–90% agreement); a person reads before anyone proves (Y9); one fault, yes or no, two rates; the trust ladder (proposing → writing-marked → trusted) with the addressability gate; the splitter's invention rate 36% |\n| One door, once | `machine-work-as-a-standing-property.md`, `what-confirmed-means.md`, `the-ratification-record-and-the-triage-feed.md`, `the-residual-error-taxonomy.md` class 6 | a pass fires once at one door over one batch with one net; coverage unknowable (concept tagging 24.3%, resemblance 1.4%, 0 of 53 minted reports embedded); R8 built for the keeper (715 pairs owed, 553 outdated by one prompt change); the record as memory, version by content; `confirmed` meant *a person or extraction wrote this*, retired; no ratify act exists; contributions and telemetry, nothing else; ratification decay cannot bite until a second person arrives |\n| A library with a gate | `alien-proofs-and-the-library.md`, `lean-social-system-research.md`, `understanding-after-the-proof-machines.md`, `lean-deliberus-analogies.md` § 7 | Mathlib: about 30 maintainers, the queue from 1,500 to over 2,600 in a year, machine-written code misses the standard *\"by a large margin\"*; the checker stops at the proof and the library runs on judgement; a result can be accepted before understood only where something other than a reader checks it; a library can overwrite because every use is re-checked, a graph cannot; the Lake pattern: re-check only what depends on a changed claim; *every AI-generated proof is born an unsolved exposition problem* |\n| What the structure must carry | `the-reuse-flywheel-prior.md`, `drowning-in-claims.md`, `saturation-and-the-long-tail.md`, `agents-as-a-consumer-class.md` | identity must be cheap to decide, and here it is judged (the legal bucket); `owl:sameAs` unified 177,000 names falsely; volume costs only when unconnected or unmerged; the head saturates (72.5%), the tail never closes and is where the value is, and summaries kill it; agents are the ideal consumer of a nonrival good, read-only, instruments as the API; *what does an AI reader owe the graph* is unanswered |\n| Recovering the reasoning a leap hides | `derivation-sciencepedia-and-the-logic-core.md` | SciencePedia checks the end of a chain by two models agreeing; Derivation would check steps by logic; Deliberus checks the middle between two given ends; three layers of logic, one built |\n| His own earlier words on science | `what-he-held-in-the-group-2011-2013.md` § 1.14 (2012) | *\"Science is a subset of consensus tech, and in its current form it's a very rough diamond. It's about time we sharpened it a bit.\"*; every inference checkable; version history keeps every corrected error |\n\n**Not read, named for the record**: `the-tidying-line.md`, `graph-daemons-design-space.md`, `cognitive-commons-and-deliberus.md`, `steadying-an-outside-judge.md`, `incentives-analysis.md` beyond § 2, `structure-versus-scale.md`, `world-models-and-the-ontology-revival.md`, the dogfood-run notes. Each is cited below only through a note that was read.\n\n---\n\n## 3. What the corpus settles, claim by claim, before the outside evidence\n\n### 3a. *Journals needs a complete rethink*\n\nSettled, and the corpus says which rethink survives. Every predecessor in science's own line died at the same wall, asking people to type the structure (nine attempts, `science-reasoning-infrastructure.md`); so *complete* must not mean a new format authors write in. The moment to act is after any text exists, never at publication (the *wrong moment* finding); the gate that stayed through thirty-five years of reform is the one on the reviewer's own argument, which no reform unbundled from prose. And the corpus's own precedent reading (Csiszar) says checking infrastructure is adopted during floods by the institutions the flood threatens, which is where Roy, TMLR's editors and arXiv now stand. The strongest objection on record is Strevens: science worked because it narrowed the official record to empirical testing, and a rethink that widens it needs a restriction as strict in its own terms. The corpus's candidate, *official moves must be graph-legible, role-explicit and challengeable*, is unratified since July; this pass does not re-open it, and § 7 names it where it bears.\n\n### 3b. *AI reviews by top models are now much better than time-pressed human ones*\n\nSettled in the corpus's terms, and the terms are not Roy's. *Better* is not a property of a model. A pass earns trust on one task, with one version, on one kind of material, by a measured bound (T1, T8), and a bound is proven only for the kinds of fault a person's reading found (Y9). A review is not one task. The one referee report the corpus has taken apart holds six kinds of move in 450 words (a verdict, a definitional attack, a causal claim, a duty claim, a sociological observation, a speaker's hedge), and the three kinds of truth say which can be proven at all: *is this claim met by the evidence the paper cites* and *is this quoted passage in its source* are checkable against the text; *does each part stand alone* is consensus; *is this interesting* and *which value premise is right* are contested, where accuracy is undefined past the ceiling of two careful readers' agreement. So a machine review can be *better* on the checkable microtasks, and nothing can make it better on the contested ones. TMLR's October split, soundness to the machine and interest to people, is exactly that table in production.\n\nRoy's reply names the mechanism the corpus holds as a conviction: *\"My chatbots have seen dozens of my reviews, etc., and I've given them a lot of guidelines\"*. That is the harness conviction from the other side: the quality of machine reasoning is a function of harness and grounding, and here the harness is one person's record of reviews. It also names the limit. A harness built from one reviewer's reviews reproduces that reviewer's frame, and the two error classes the harness can never reach are what the corpus is a sample of and the frame it crystallised (the ceiling of `the-harness-conviction.md` § 10). AM12 adds the sharper warning: a pass shown twenty of Roy's reviews copies how they look; a pass told what a sound claim is can judge a paper Roy never reviewed.\n\nWhat the corpus refuses outright is the form in which an AI review would usually enter a journal: as a verdict. Every judge in this pipeline is a single-shot proposer, and a challenge to its answer becomes a claim in the record, never a turn in its conversation (`fragile-checkers-and-the-verification-bottleneck.md` § 2a), because a verdict held in a channel flips under pressure 25–71% of the time and the flips are mostly wrong. The scrutiny gap's leg is the same point from the production side: machine checking handed to the class of system that produces the papers, with half of ICLR's reviews already machine-assisted and white-font instructions planted for the machine reader. And a second model's agreement is a relay, not a witness (R7). So the corpus's answer to *AI reviews are better* is: **on the checkable half, probably, once measured; as a verdict, never; and the comparison Roy draws is between two unmeasured things.**\n\n### 3c. *One-time gate keeping (accept / reject) doesn't make sense*\n\nSettled, by ruling, and measured on the project's own pipeline. *Contributions are added, never overwritten; nothing needs permission to exist* (D6). What is gated is weight, and which form of a thing is shown first (`alien-proofs-and-the-library.md` § 4.9). There is no ratify act; the graph holds contributions and telemetry (the end-state blessed 2026-08-31). A machine pass is owed by a condition on the graph, never fired once at a door (R8, ruled with a brake), because the one-time door was found to be the architecture's own defect: a pass fires once, at one door, over one batch, with one net, and whatever it missed stays missed, so that concept tagging had touched 24% of claims and the resemblance linker 1.4%, with nobody able to say what had been attempted. A later reading retires an older one and keeps it visible (R12). A pass's version is the fingerprint of its instructions, so a changed prompt makes every earlier reading outdated at once and the queue is reported, never drained by itself. All of that is the opposite of accept-once.\n\nThe nuance his line lacks: **the queue does not vanish when the gate goes; it moves.** Mathlib has a perfect checker of truth and still runs its library on judgement, and its queue grew from about 1,500 to over 2,600 waiting submissions in a year. Here the gate on existence is gone and the queue stands at the review of what machines wrote (the sittings, built and unsat; the triage feed, unruled). TMLR shows the same on its side: with no accept-reject at submission time, its gate moved to desk rejection, from 6% to 53%.\n\n### 3d. *We need to think about it as a distributed data structure, and start over*\n\nSettled in substance and underspecified in the post. The corpus's position is that the unit is the claim, not the paper, and what Roy's phrase leaves open is everything that makes the structure work:\n\n1. **Identity.** A data structure whose elements are judged the same or different by reading is in the legal bucket, not the chemistry one; identity work is the permanent core activity, and strict identity used sloppily at scale unified 177,000 names falsely (`owl:sameAs`). A check made once answers every paraphrase only if the paraphrase is recognised as the same claim (`the-agent-population-and-what-checking-is-up-against.md` § 2d): *the race is decided where sameness is decided*.\n2. **What a node carries.** Provenance on every count (asserted, minted, reported); a stance for reported content on a link that does not propagate (always-mint); a date and a validity that can lapse; a version for every machine reading, named by its content, write-once (Z1, Z11); a confession value on every closed list.\n3. **What never happens.** Nothing is overwritten; a retired link stays visible and restorable; a person's words are never changed in place (AL10). A library can overwrite because every use of a changed statement is re-checked and a change that matters breaks something loudly; a graph of reasons cannot, because nothing breaks (`alien-proofs-and-the-library.md` § 4.5).\n4. **What the reader meets.** A graph is tail-preserving and every summary door in front of it is tail-killing; the rare claim, the dissent and the unwritten crux are where the value sits and are the first things a summariser drops (`saturation-and-the-long-tail.md` § 3). *Who is consuming this?* is Roy's own third-day question, and the corpus's answer is that the main consumer of a nonrival structure is a machine reader, read-only, with the honesty instruments as its API.\n5. **What no structure reaches.** What the corpus is a sample of, and the frame it was built inside. A data structure cannot know what was never submitted to it.\n\nAnd one more thing the corpus holds that the phrase does not: the structure's own checking layer is part of the structure. An AI review that is a reading attached to a claim, versioned, retired by a later reading, kept visible, moving no strength until examined, is not a gate and not a judge. It is what the data structure does to itself. **That is the form in which Roy's second and third sentences stop contradicting each other.**\n\n## 4. What the outside research adds\n\nTwo helpers' reports, verbatim in `the-literature-as-a-distributed-data-structure/` (`report-tweet-and-ai-review-2026-11-04.md`, `report-data-structure-and-ground-truth-2026-11-04.md`). Every number below is recomputed in `cases.py` from the counts the helpers tagged as read; the read-status tags are theirs.\n\n### 4a. The post, in its context\n\nRoy was a JMLR **Action Editor**, not an editor-in-chief; he is on the board's snapshots of January and December 2024 and off the live page, which fits *resigned ~2 years ago*. The one earlier statement of his on the subject is the *marginalia* page of his own site: errata, clarifications, missed citations and links to follow-on work are *\"important, but underappreciated, contributions to the scientific literature. Ultimately, the nature of a scientific document needs to be rethought\"*, collected meanwhile in a wiki (now a 404). That is the precursor of the fourth sentence, and it is the corpus's own position stated from the editor's chair: the unit is smaller than the paper and the record should hold the marginalia. On Bluesky, 23 November 2024, two years before the post: *\"Reviewing is dead\"*, and *\"No one gets tenure by reviewing well\"*, which is an incentive diagnosis, not a capability one. Nobody in the thread elaborated the data structure; the one reply the helper could retrieve asks *\"But what is the solution?\"*, and Roy did not answer it.\n\nTwo facts the corpus had wrong or lacked. **JMLR was founded in 2000 by Leslie Kaelbling, a year before the forty editors left Kluwer's *Machine Learning***; the 2001 letter endorsed an existing journal, and its subject was access (an institutional subscription cost $1,050), not the review model. Roy's is the first resignation in that lineage about the gate itself. And **the carriers are moving at once**: TMLR tripled in a year, solo-authored submissions up thirteenfold, desk rejection from 6% to about 53%, 159 authors suspended, and an AI soundness review on every submission since October, the human action editor still deciding; arXiv capped authors at two papers a month from 1 October after a record 40,363 submissions in September, 51% more than the September before; ICLR 2027 caps authors at twenty papers; IEEE retracted 1,417 conference papers in one day on 23 October. The carriers note's Tier 1, *the drowning carrier*, is now a description of the people in the post.\n\n### 4b. The measured state of AI review: two registers, opposite results\n\nThe literature splits by what it measured, and the split is the one § 3b predicted from the three kinds of truth.\n\n**As feedback to authors, on checkable faults, with ground truth: better, by two to three times.** The cleanest test is the newest. Brian et al. (medRxiv, 25 October 2026) re-ran the BMJ trial's planted-error papers, where 491 human reviewers had found 2.40 of 9 major errors (27%), through fifteen models: 4.73 to 8.65 of 9 (53% to 96%), and on 58 trials published in 2026 a sensitivity from 46% (Haiku 4.5) to 93% (Opus 5.5). The human baseline is stable across twenty-eight years of planted-error studies: 34% of major errors in Baxt 1998 (those recommending acceptance found 17%), 25% in Godlee 1998, 29% in Schroter 2008. King and Worthington (28 October) measured helpfulness rather than errors and found the same direction: Claude Opus 4.8's reviews preferred to paid expert reviews in 134 of 134 pairs by a GPT-5.5 judge and 26 of 30 judgments by three former editors, with incorrect claims about the manuscript at 5.5% against 15.9%. AAAI-26 gave every one of 22,977 papers a labelled machine review and 5,834 survey respondents rated it above the human on six of nine criteria, worst on *overemphasized minor issues*.\n\n**As a verdict: worse, inflated, correlated and purchasable.** Baumann et al. pooled 8,015 ICLR 2026 papers: averaged human scores predict acceptance with AUC 0.822, averaged AI scores 0.710; AI means 7.3 and 6.1 against the humans' 4.3, correlation 0.15; a $0.25 rewrite of hedges and emphasis words raises the AI score by 0.45 of a point; fully machine-written reviews are more alike across papers than reviews with human input. BadScientist's fabricated papers were accepted by a model panel at up to 82%, often with the integrity problem flagged in the same review. On real published errors the same models find a fifth: SPOT's best recall is 21.1% at 6.1% precision, and TMLR's own evaluation of the reviewer it chose finds it misses 70% to 82% of inserted claim-breaking errors. The ICML 2026 randomised trial of reviewer policy found near-zero effect on decisions (27.0% against 26.5% accepted) and 22.5% of the no-model arm used one anyway; Pangram counted 21% of ICLR 2026 reviews fully machine-written; seventeen arXiv papers from fourteen institutions carried white-font instructions to the machine reviewer.\n\nSame model class, opposite results. What differs is the task: *find the planted error* has a ground truth and the model finds it; *is this paper good* has none and the model's answer is a rating of its prose. The human gate is not the standard either: two NeurIPS committees judging the same papers disagreed on 26% in 2014 and 23% in 2021, and in both years about half of what one committee accepted the other rejected (random would be 35% to 38%); Peters and Ceci's twelve resubmitted published papers were rejected eight times of nine. So Roy's second sentence is true on the one task where it can be measured and false on the task a journal would hand the machine, and the corpus's split of the review into kinds of truth is what the literature's split reproduces.\n\n### 4c. The gate, and what runs without one\n\nThe gate's errors run both ways and the slow one costs most. Akerlof's *Lemons* was rejected three times as trivial and has 23,235 citations; Higgs's particle prediction exists because Physics Letters rejected the paper without it; Campanario counts 24 future Nobel laureates who met referee resistance on the prize-winning work. The other way: arsenic life passed *Science*'s gate in December 2010, was refuted in print in nineteen months and retracted after fourteen years and seven months; LK-99 never met a gate and was refuted in public in 25 days. The Retraction Watch record of 68,154 dated retractions has a median lag of 1.36 years, a 90th percentile of 6 years, and a median of 4.26 years where the cause is fabrication.\n\nWhat runs without a gate is small and shows review changing the paper in a minority of cases. eLife's publish-review-curate model sent 27% of 12,228 submissions to review and improved the evidence-strength wording of 44% of them between reviewed preprint and version of record; its intake halved in 2025 after Clarivate withdrew its impact factor. Peer Community In recommended 248 preprints in 2025 at about €718 each. The Unjournal found 5 of 22 assessed papers updated substantively and one with a public link from evaluation to correction. COS's Lifecycle Journal published 34 papers in 2025 and the pilot has closed. Fewer than half of 119 living evidence syntheses updated once in two years. Of 2020 preprints, 7% received a single comment in seven months. PubPeer, the one continuous checking layer at scale, gets an editorial notice for 21.5% of the papers it flags; public discussion yields 6.5 times the retractions of a private report, and ten commenters wrote 27% of the first comments on the 800 highly cited retractions Bishop counted. The Linux kernel, the one continuous review process that runs at the volume arXiv now sees, reviews half of what arrives on its busiest list and ranks the rest by the submitter's standing.\n\nThe defenders of the gate say one thing the corpus must answer. Shalizi: without referees every reader redoes the same filtering, and the division of labour reinvents review. Boyer: publication in a named venue is a signal, and withholding it reads as low quality, so a gateless record faces a coordination problem, not an epistemic one. Desmond: a gate removed is replaced by citation counts and search ranking, which are gates with worse owners. Tao, August 2026: *\"I do not believe that human referees can be removed from the publication process.\"*\n\n### 4d. What *a distributed data structure* has meant to those who proposed or built one\n\nNobody has built Roy's structure, and the proposals disagree on the unit. **The paper**: Lorig's forkable archive (July 2026, no implementation), where a deposit is source, commentary is attributed and voted, forks link their parent and there is no accept bit, and whose author lists seven objections against himself, sparse participation first; Zharnikov's research-as-repository. **The claim**: Nielsen's 2007 micropublication; Kejriwal's ClaimRxiv (October 2026), a scoped assertion node with evidence, provenance and epistemic status, five relations (supports, challenges, refines, replicates, supersedes), and the sentence that names what it has not solved, that claim identity across papers cannot be string similarity; Tardigraph, a git repository eight days old with 1,690 claims carrying a locator and 161 contradictions kept rather than resolved. **The proof**: Tao's pipeline of generation, verification, exposition, publication and canonicalisation with authorship decoupled per stage; Palomar, a registry of Lean snapshots checked mechanically and then by a model, with no human review; Hexagon, which accepts fully machine-generated mathematics and rejects only for obvious defects. **The run**: XScientist's exploration DAG with content hashes and failed branches, for research machines did.\n\nThe software analogues give the mechanics the proposals lack. Git is content-addressed, so a version is named by what it contains (Z11's rule, from 2005). PyPI's yank, npm's deprecate and Cargo's yank are non-destructive withdrawal with a stated reason, visible to users, reversible, and ignored by anyone who pinned the old version: R12 in a package registry. CRDTs converge every replica on the same state with no coordinator and leave exactly one case to the application: two concurrent writes to the same property of the same object, which is sameness. And every one of them, as the helper notes, converges on *state* and defines nothing by evidence; the kernel and the registries name review capacity and consumer pinning as their limits and measure no error rate.\n\nThe critics of claim-level stores do not reject them; they name what flattening loses. De Waard: flat triples drop epistemic status and the link to evidence. Al Khatib with de Waard: tentative statements get stored as fact, and a claim and its premise sit in different sections. Hyland: a hedge is the writer's negotiation with the reader. Kejriwal against his own proposal: early flattening loses study design, effect sizes, time and negative evidence. The corpus has a ruled answer to each (a hedge is the author's stance stored beside the claim, AM6; every count labels its provenance; a reported position is minted and attributed with a non-propagating link), which is what the proposals lack, and the proposals have what the corpus lacks: a carrier (TMLR, eLife, arXiv), a scale (Semantic Scholar's 225 million papers, Tardigraph's 274 in eight days against the graph's 31), and a user.\n\n---\n\n## 5. Tests on real cases\n\nThe tests are the twenty-two cases of `cases.py`, each a published count run through the readings of the claim it bears on. The readings are: **Roy's**, the sentence as written; **the corpus's**, § 3's statement of it; and **the defender's**, the strongest opposite reading found in § 4. A reading *holds* where the case comes out as it predicts, *fails* where it does not.\n\n| Claim | Case (number in `cases.py`) | Outcome | Roy's reading | The corpus's | The defender's |\n|---|---|---|---|---|---|\n| 1. Rethink | TMLR's year (4): desk rejection 6% to 53%, reviewed share of submissions 94% to 47%, reviewed volume still 2.0x | the journal survived by moving its gate forward, not by removing it | holds | holds: the moment to act is after any text exists | holds: the gate was conserved |\n| 1. Rethink | arXiv (22): 51% in a year, cap of two a month | the archive without a review gate added a volume gate | holds | holds | holds |\n| 2. AI better | planted errors (16): humans 27%, models 53% to 96%; 46% to 93% sensitivity on 2026 trials | 2 to 3x on the checkable task, against ground truth | holds | holds | fails |\n| 2. AI better | helpfulness (12, 1, 3): 134/134, 26/30; overlap 1.08 to 1.11x a second human; 6 of 9 criteria | preferred as feedback | holds | holds | fails |\n| 2. AI better | the verdict (17): AUC 0.710 vs 0.822, bought for $0.25, 21% of ICLR reviews fully machine | worse as a judge, gameable, homogenising | fails | holds: as a verdict, never | holds |\n| 2. AI better | real published errors (16, 18): SPOT 21% recall at 6% precision; the checker 83% precision, 60% recall on injected; TMLR's reviewer misses 70 to 82% | the planted result does not transfer to real faults whole | fails as stated | holds only with the bound and scope T8 requires | holds in part |\n| 3. Gate | repeatability (15, 14): 23 to 26% of verdicts flip; half of accepted; 8 of 9 published papers rejected | the accept bit is a draw of referees | holds | holds: what is gated is weight, not existence | fails on repeatability |\n| 3. Gate | wrong both ways (21): Akerlof, Higgs, 24 laureates; arsenic life 14.6 years | the gate's slow error costs most and a held objection would have closed both | holds | holds | fails |\n| 3. Gate | what runs without one (20, 13, 19): 7% commented, 21.5% noticed, 5 of 22 updated, Lifecycle closed | removing the gate yields silence, not review | fails: *doesn't make sense* says nothing of what replaces it | holds: a reading a machine writes does not wait on a commenter; the queue moves | holds: the signal and the coordination are lost |\n| 3. Gate | the queue moves (3c, 22): Mathlib 1,500 to 2,600 waiting; the kernel reviews half, ranks by standing | no system removed the gate; each moved it to the queue | fails | holds | holds |\n| 4. Structure | the objection stream (18, 19): 4.66 flags a paper, 83% real; PubPeer 21.5% to notice | the demand is measured and nothing holds it | holds | holds: the structure's checking layer is part of the structure | not addressed |\n| 4. Structure | the built ones (4d): Tardigraph 1,690 claims in 8 days; ClaimRxiv names identity unsolved; CRDTs leave the same-property case to the user | the unit is the claim and the unsolved part is sameness | underspecified | holds: *the race is decided where sameness is decided* | de Waard's loss list holds, and the corpus answers each |\n\n**Where each reading fails, and what predicts it.** Roy's reading fails exactly where the sentence stops: *better* does not say at what, and *doesn't make sense* does not say what replaces it, and the cases that sit in those gaps (the verdict; the gateless silence) go against him. The defender's reading fails wherever a ground truth exists (planted errors, repeatability) and holds wherever the question is social (the signal, the coordination, who shows up). The corpus's reading holds on every case **with one condition it already states**: the machine's advantage is bounded by the kind of fault a reading found (T8), and the planted-error result does not transfer whole to real errata, where the same models find a fifth. What predicts the transfer failure is whether the error was constructed to be findable or arose in the wild; a pass proven on the first kind carries no bound on the second, which is Y9's reason for reading before proving.\n\n**The failure cases for the corpus's own reading.** Two on record. First, the precision side: the models flag 14% to 29% of error kinds on clean control reports, and SPOT's 6% precision means sixteen flags per real fault on real errata. A machine objection attached to a claim, moving no strength until examined (AG12), is the right shape for that, and it still fills the queue: at 4.66 flags a paper and 40,000 papers a month, the objection stream is 190,000 claims a month that nobody asserts, which R8's brake of four calls a sweep was built never to produce and which the triage feed (AW1, two halves ruled) has no ranking for. Second, participation: every running gateless system shows that a place for the objection is not an objector. The corpus's answer, that the standing passes write the reading nobody comments, is an answer only once a pass is proven, and none is (§ 6).\n\n**One exact case.** The post's own reply, *\"My chatbots have seen dozens of my reviews\"*, run through AM12 (a pass is taught the method, never the contents): a reviewer harnessed on twenty of Roy's reviews reproduces Roy's frame and reads a paper outside it as Roy would misread it, which is the frame lock-in class the harness cannot reach; a reviewer told what a sound claim is judges the paper Roy never saw. No count exists for this; it is the corpus's prediction, and the hivemind result (fully machine reviews more alike across papers, +18% to +40% in controlled runs) is its nearest measurement.\n\n---\n\n## 6. What does not fit, the rows each recommendation touches, and the assumptions ranked\n\n### 6a. What does not fit the corpus's reading\n\n- **Anonymity.** Every proposal and every running system meets it: Lorig's real names chill criticism; PubPeer's commenters were sued and a court kept them anonymous; the peer-review note's own limit says a legible review record is weaponizable and anonymity protects the reports most worth having. The corpus attributes every contribution and has no anonymous contributor design; a blob's membership is the nearest thing. This is a cost the structure carries and no ruling addresses.\n- **Certification and credit.** Boyer's coordination problem and Lorig's lost signal: a journal's name is a credential on a paper, and a badge is a weight on a claim. Nothing in the design certifies anything, and *who is consuming this* (Roy's own second question) has, for a career, the answer *a hiring committee*, which reads credentials and not graphs. Lorig § 6.3 names it; the corpus's carriers note assumes the carrier supplies it.\n- **Strevens.** A rethink that widens the official record past empirical testing needs a restriction as strict in its own terms; the Legibility Rule is the corpus's candidate and is unratified since July. Roy's post has no restriction at all.\n- **The machine judging the comparison.** The strongest number for Roy's second sentence (134 of 134) is a model preferring a model's review, and the human check is ten pairs; R7 says a second model's agreement is a relay. The editors' 26 of 30 is the real evidence and it is small.\n- **Roy's third-day answer and BF3.** *Who is consuming this* gets, in the corpus, the answer *a machine reader, read-only*; and BF3 (open) says the record is never served as reward or published evaluation while learning from it as data is the use. A literature that is read mostly by machines is a literature trained on, and the corpus's reader and its training target are the same object seen from two sides.\n- **Readiness.** Every potentiality below is lineage and not roadmap (the scale ladder's gate), and it says so: in production every machine pass is off (2026-10-24); one of 24 passes is standing work (the keeper); no pass has a proven bound; nobody has read a pass's output in its kept form (`quality-control-for-deeper-steps.md`). A cost of now, with what ends it named in TODO's order of work; not a cost of the design.\n\n### 6b. The ruled rows and canon lines each recommendation touches\n\nRead whole this session; for each, whether the recommendation extends, narrows or reverses it.\n\n| Recommendation | Touches | Relation |\n|---|---|---|\n| 1. Nothing changes on the post; it is filed as evidence | D6 (added, never overwritten); R8 (standing work); R12 (later reading retires); the 2026-08-31 end-state (no ratify act); T1, T8, Y9 | extends each with an outside measurement; reverses nothing |\n| 2. The two registers as the first outside calibration of the three kinds of truth | the harness conviction and its BF2 scope; AM12; the three kinds of truth (`the-evals-loop-and-our-machine-passes.md`, unruled) | extends; the planted-vs-real transfer failure narrows T8's *proven on English extracted claims* to *proven on the kind of fault the reading found*, which T8 already says |\n| 3. The OpenReview thread as the next real-material register | the Pipeline Testing Principle (substance over meta); AV2 (the outreach gate); the demo of 21 November; T1's depth cap | extends the testing principle to a register where the review is attached to its object; consistent with AV2, since nobody is asked to try anything; waits on the reviews and the demo |\n| 4. *Review as capital* named as the potentiality; the entry point stays pre-submission self-review | D6; AG12 (a machine's sentence is a claim when written); the peer-review note's eight mechanisms; TODO line 4451 | extends the eight with a ninth that the literature's own numbers now motivate; reverses nothing |\n| 5. A soundness reading per claim as the project's form of AI review | R8; R12; Z11; AM12; the no-verdict principle; TMLR's split | extends standing work with a pass the carriers now run as a finished output; narrows nothing; depends on the reviews finishing first |\n\n**The outside check on this draft** (`scripts/outside_check.py --file`, run 2026-11-04): it found no conflict; it reads the note as extending AM12, BF1 and the harness conviction, repeating D6, R8, Y9 and R7, and depending on T8, and it could not tell two things the author also could not: whether the triage feed's work half (AW1) could absorb an objection stream of 190,000 claims a month, and whether the certification gap of § 6a is anything AL7's stance on agreement bears on. Both stand in § 6a and § 6c as open. Its line is carried into each row below.\n\n### 6c. The pass's assumptions, weakest to strongest\n\n1. **That a claim-level record would get the objections PubPeer gets.** Weakest: 7% of preprints commented, 1 of 22 Unjournal papers publicly corrected, ten commenters writing 27% of PubPeer's first comments. Nothing in the corpus has been measured against a motivated contributor, and this would need one.\n2. **That the planted-error advantage transfers to real faults.** SPOT says it does not transfer whole; the checker over published papers (83% precision, 60% recall on injected) says it transfers in part. Unresolved, and T8 is built for exactly this gap.\n3. **That a machine reading attached to a claim beats a comment thread.** Unmeasured: no pass here has a proven bound, and PubPeer's thread, with no weight on anything, got 6.5 times the retractions of a private report.\n4. **That sameness can be decided by reading at the volume of a literature.** The keeper reads 19 of 20 real pairs right; ClaimRxiv names the problem unsolved; the legal profession devotes itself to it. Holds at twenty pairs and is untested at twenty thousand.\n5. **That the task, not the model, sets the error rate.** Strong: one model class, two tasks, opposite results (Brian et al. against Baumann), which is the harness conviction measured from outside.\n6. **That the one-time verdict is a draw of referees.** Strongest: measured twice at NeurIPS, once at Peters and Ceci, and in every planted-error study since 1998.\n\n**Where the thinking may be wrong.** The corpus's reading answers Roy's third sentence by saying the gate moves to the queue, and then treats the queue as this project's existing work (sittings, triage). That is true of Deliberus and does not answer Roy: a journal's queue is read by editors paid in prestige, Mathlib's by maintainers paid in nothing, and the corpus's by a founder. The honest form of the claim is that *no structure removes the attention gate; a structure decides what the attention is spent on*, and that is a weaker claim than § 3c's.\n\n---\n\n## 7. Recommendations, in order of dependency\n\nEach names the alternatives it beat. Nothing here is acted on; items 3 to 5 are the decisions, as ledger Arc BG.\n\n1. **Nothing in the pipeline changes on the post, and the post is filed as evidence under the rulings it confirms** (D6, R8, R12, T1, T8, Y9). The corpus held every move the post names before the post: the unit as the claim (2008), standing work over the one-time door (R8), the retired-and-visible reading (R12), the no-verdict judge. *Beat*: reading the post as a new direction (there is none in it the corpus lacks); ignoring it (loses the first outside measurements of three rulings and the signal that the drowning carrier is now resigning in public). *Changes*: six pointers, one line each, from the notes this answers.\n\n2. **The two registers of the review literature are recorded as the first outside calibration of the three kinds of truth.** On checkable faults with ground truth, a top model finds two to three times what a time-pressed human finds; as a verdict it predicts the human verdict worse than humans do and is bought for a quarter. *Beat*: a new canon line (not wrong by default: no session has asserted that AI reviews are better flat, and the harness conviction with BF2 already carries the task-not-model reading); a note of its own (the measurement belongs beside the table it calibrates). *Changes*: a dated line in `the-evals-loop-and-our-machine-passes.md` § 3 and in `the-scrutiny-gap.md`.\n\n3. **The next real-material register, after the reviews and the demo, is one OpenReview thread: a submission, its public reviews, the rebuttal and the decision, mapped together** (BG1). It is the one register where the review is attached to the thing it reviews and the decision is on record, so it tests what no earlier run could: does a reviewer's objection land on the claim it is about, and does TMLR's soundness-versus-interest split fall out as the checkable-versus-contested kinds. Quota only, on the Claude engine, by hand where the engine's ceiling bites. *Beat*: a journal paper with no public review (no ground truth on the objection); an Agents4Science paper (machine-written, reviewer agreement 19 of 250, unclear what it would test); a TMLR rejected submission (the same test with a weaker decision record, and a second candidate if the first is thin).\n\n4. **The potentiality named on the peer-review note is *review as capital*, and the entry point stays pre-submission self-review** (BG2). Everyone now builds the feedback pass as a finished output: TMLR's soundness review, Google's Paper Assistant Tool, King and Worthington's. What none of them keeps is the objection: an AI review is read once by one author and discarded. In the graph an objection is a claim attached to the claim it attacks (D6, AG12), so the next paper that asserts the same claim inherits the objection unanswered, and the review a field has already done compounds instead of repeating. That is the one thing a data structure does that a reviewer, human or machine, cannot, and it is what Roy's fourth sentence asks for without saying. *Beat*: naming the AI reviewer itself as the potentiality (the carriers have it and it is the verdict the corpus refuses); naming interoperability (ClaimRxiv, Nanoarguments, Tardigraph; real, cheap, and it waits on AV2 since it is outreach).\n\n5. **The project's own form of AI review, when a pass is designed for it, is a soundness reading per claim, never a verdict on the paper** (BG3): for each claim extracted from a paper, whether the evidence the paper cites meets it, stored as a reading with its version, retired by a later reading, moving no strength until examined. It is the half TMLR gave the machine, done the way R8 and R12 require, and it is the pass whose bound Brian et al. and the checker over published papers have measured from outside. *Beat*: an interest reading (contested; no ground truth; the half TMLR kept for people); a paper-level reading (the badge boundary failure in a new place). Designed after the reviews; not before a pass now built has its bound.\n\n6. **Parked, not decisions.** The carriers note's Tier 1 gets Roy, TMLR's editors and arXiv's Dietterich by name as the people now in the water (a dated line; outreach waits on AV2). The competitive landscape gets ClaimRxiv, Hexagon, Palomar, Tardigraph and Lorig's archive in one paragraph. The objection-volume arithmetic (190,000 machine objections a month at arXiv's scale against a brake of four calls a sweep) is a design question for the triage feed's ranking, filed in § 5 and in TODO beside AW1's open rows.\n\n---\n\n## 8. The decisions, as ledger rows\n\nArc BG in `session26-open-rulings.md`: BG1 (the OpenReview register), BG2 (review as capital), BG3 (the soundness reading). Each carries its **For**, its **Because** in both views, its **Meets** line and the outside reader's line.\n\n**BG2 (review as capital) RULED 2026-11-04**, the evening it was asked, his word the offered answer: *\"Yes, rule BG2\"*. Canon: through AG12 and D6, with this note (§ 7, recommendation 4); graded predicted. The ripples read over the 428 open rows changed none. BG1 and BG3 wait, one at a time.\n\n---\n\n## 0b. In short, the answer in five lines\n\n1. Roy names three moves: replace the reviewer, remove the gate, change the unit. The corpus holds the last two as rulings and refuses the first as a verdict.\n2. On the one task with ground truth (planted errors) a top model finds two to three times what a time-pressed human finds; as a judge it is worse, inflated and bought for $0.25. The task sets the rate, not the model.\n3. No running system removed the gate; each moved it to a queue, and the gateless ones get a comment on 7% of papers.\n4. A machine review entering as a reading attached to a claim, versioned, retired by a later reading, visible, weighing nothing until examined, is the data structure's own checking layer; that is where his second and third sentences stop contradicting each other.\n5. The potentiality is review as capital: an objection kept on the claim it attacks, inherited by the next paper that asserts it. Deliberus has the shape and none of the proof; the proof is the reviews' work.\n\n"}