b19 — Cross-consistency check across the seven AI-co-authorship fact-sheets#

Compiled:

2026m05d13

Compiled by:

Claude Opus 4.7 Max (main session, drawing on subagent fact-sheet 1–7 deliverables)

Scope:

Surfaces tensions, corroborations, NEJM-AI clarifications, and the decision-relevance ranking that will frame the (b) EDEN steelmans on AI co-authorship for the matheo-b19 paper.

Inputs:

  • Fact-sheet 1 (journal-policy landscape; 997 lines)

  • Fact-sheet 2 (authorship-criteria frameworks; 797 lines)

  • Fact-sheet 3 (historical precedents; 921 lines)

  • Fact-sheet 4 (verification-bandwidth in authorship; 954 lines)

  • Fact-sheet 5 (responsibility allocation across asymmetric parties; 899 lines)

  • Fact-sheet 6 (AI co-authorship precedents — has any officially-acknowledged case been retained?; 778 lines)

  • Fact-sheet 7 (non-individual entities as authors; 1061 lines)

Status:

Independent synthesis — informational, not yet a recommendation. The (b) steelmans build on this.

Methodology — read first

All seven source fact-sheets carry the same caveat: WebFetch was sandbox-blocked across every subagent run. Quoted operative sentences and citation details may be training-reconstructed and are flagged inline with [QUOTE NEEDS VERIFICATION] etc. This cross-consistency check inherits the caveat — claims supported by convergent evidence across multiple fact-sheets are stronger than claims resting on a single quoted item, and the synthesis below prefers convergent-evidence claims wherever possible.

Section 1 — Tensions surfaced across fact-sheets (with resolution)#

Tension A — Personhood vs accountability as the operative criterion#

Claim from Fact-sheet 2 (authorship-criteria frameworks):

“The durable distinction is **legal personhood at the moment of signing*, not multi-criterion competence.”*

Claim from Fact-sheet 7 (non-individual entities):

“The operative criterion separating accepted from non-accepted classes is **accountability, not natural-person status*. ICMJE Section II.A.4, COPE, Nature, Science, and PLOS all justify the AI prohibition on accountability grounds, not personhood grounds.”*

Apparent contradiction. Both fact-sheets surveyed primary policy text; both reached different conclusions about which criterion is the operative blocker.

Resolution. Fact-sheet 7 inspected actual practice across the four classes of non-individual byline authors already accepted in scientific publishing (consortia, institutional/standards-body, collective pseudonyms, individual pseudonyms). The existence of those four accepted classes empirically rules out personhood-per-se as the operative blocker — if personhood were operative, none of the four would be accepted. Fact-sheet 2’s reading is correct as a description of what gets applied as a shortcut when the candidate is AI, but incorrect as a description of the stated criterion. Fact-sheet 7’s reading wins on the stated criterion; Fact-sheet 2’s reading describes the applied shortcut.

Decision-relevance. This single resolution is the sharpest move in the entire fact-sheet pass: the conventional-position steelman cannot rest on a personhood gate. It must argue accountability. That changes the steelman entirely.

Tension B — Bourbaki rejected (FS3) vs Bourbaki accepted (FS7)#

Claim from Fact-sheet 3 (historical precedents):

“The American Mathematical Society rebuffed Bourbaki’s individual- membership application on the ground that ``he was not an individual``. Same categorial ``is the signatory an individual?`` test that ICMJE applies to AI today, for at least three-quarters of a century.”

Claim from Fact-sheet 7 (non-individual entities):

“Bourbaki in mathematics, **accepted [as byline name] since 1935* via senior sponsor-vouching by Élie Cartan.”*

Apparent contradiction, already pre-empted by LLoL in Prompt 4: the AMS is a society of individuals whose membership rules are free-form and not the same domain as journal authorship.

Resolution. These are different domains: society membership (rejected for Bourbaki) vs journal authorship (accepted for Bourbaki). LLoL’s pushback was correct. The Bourbaki-1948 finding should be downgraded in the (b) steelman of the conventional position; the Bourbaki-1935 finding should be upgraded in the (b) steelman of the pro-co-authorship position. Bourbaki in fact gives the pro-position a 90+ year precedent of accepted non-individual byline authorship with senior sponsor-vouching — structurally close to the AI-with-named-absorber proposal.

Decision-relevance. Bourbaki-1935 becomes a load-bearing historical precedent for the pro-position. Bourbaki-1948 becomes inadmissible as a precedent against the pro-position.

Tension C — NEJM AI split: encourages USE, prohibits AUTHORSHIP#

Claim from Fact-sheet 1 (journal-policy landscape):

“2024 NEJM AI sister-journal explicitly encouraging LLM use against the flagship NEJM line.”

Claim from Fact-sheet 6 (AI co-authorship precedents):

“NEJM AI (which encourages LLM use in submissions but **still forbids AI on the byline*).”*

Not a contradiction; a clarification. Fact-sheet 1’s framing is incomplete — NEJM AI is split: it encourages LLM USE (drafting, methods, analysis support) but still PROHIBITS AI on the byline. The 2024 NEJM AI launch did not break the no-AI-author consensus; it relaxed the no-LLM-use consensus.

Decision-relevance. The “dissent” identified in Fact-sheet 1 is narrower than it appeared. The conventional position on AI-as-author is unbroken; only the AI-as-tool conventions have evolved.

Tension D — None additional after spot-check#

Other cross-fact-sheet claim pairs spot-checked:

  • ICMJE four criteria text (FS2, FS4, FS5): consistent.

  • ICMJE criterion 4 as the hardest single blocker (FS2, FS4): cross-corroborated.

  • 2023 ChatGPT-as-author episode (FS1, FS3, FS6): consistent on the three named papers (Kung et al.; O’Connor; Zhavoronkov).

  • Deceased-author rule mechanism (FS3, FS5, FS7): consistent triangulation — visible marker + named living absorber + estate consent.

  • ATLAS / CMS / LIGO / IPCC kilo-author practice (FS3, FS7): consistent on consortium-as-byline acceptance.

  • AI-company self-authorship practice (FS6): unique to FS6; not contradicted elsewhere. The strong negative finding (no AI company lists their own model as author) stands.

Section 2 — Cross-corroborations (where fact-sheets reinforce each other)#

Five claims reach the convergent-evidence threshold (corroborated by three or more independent fact-sheets):

Claim

Corroborating fact-sheets

Strength

The operative criterion for excluding AI is accountability, not personhood.

FS2 (applied), FS5 (load-bearing), FS7 (stated)

Very high

ICMJE criterion 4 (accountability) is the hardest single blocker, criterion 1 (substantial contribution) is cleanly satisfiable.

FS2, FS4, FS7

Very high

The deceased-author rule is the cleanest structural precedent for retained byline + visible marker + named absorber.

FS3, FS5, FS7

Very high

Consortium-byline authorship is accepted across all major journals (ATLAS, CMS, LIGO, ENCODE, 1000 Genomes, TCGA, gnomAD; IPCC, MMWR/CDC, WHO).

FS3, FS7

High (two fact-sheets, but very specific on the same instances)

Asymmetric naming without a named absorber is not a recognised form in any precedent class.

FS5 (across six legal/ethical frameworks), FS7 (across four publishing classes)

Very high

Section 3 — Decision-relevance ranking: findings that move the (b) needle#

The following are the top decision-relevant findings, ranked by how much they move the b19 AI-co-authorship question. Each is tagged with the direction it moves the needle (CONV = toward conventional position, PRO = toward pro-co-authorship position, BOTH = constrains both positions structurally) and the fact-sheet(s) it comes from.

#

Direction

Finding

Source(s)

1

PRO

The operative criterion is accountability, not personhood. The conventional steelman cannot rest on a personhood gate; four classes of non-individual byline authors are already accepted.

FS2 + FS7

2

CONV

Officially-acknowledged AI co-authorship at a refereed venue is essentially unprecedented as of 2026m05d13. One marginal durable case (Zhavoronkov & ChatGPT, Oncoscience 2022, cited as a cautionary tale not a precedent). 2 of 3 attempted cases were corrected.

FS6

3

BOTH

Asymmetric naming without a named absorber is not a recognised form in any precedent class. The pro-position must specify an explicit named absorber. The deceased-author rule provides the cleanest template (visible marker + corresponding-author absorber + estate consent equivalent).

FS5 + FS7

4

CONV

AI companies never list their own models as authors on their own model papers. OpenAI / DeepMind / Meta AI / Anthropic. Going ahead with Claude on the b19 byline puts LLoL ahead of even Anthropic’s own self-authorship practice.

FS6

5

PRO

The Bourbaki-1935 byline precedent stands: 90+ years of accepted non-individual byline authorship in mathematics, with senior sponsor-vouching as the absorber mechanism. (Bourbaki-1948 AMS rejection is inadmissible — different domain.)

FS3 + FS7 (resolution of Tension B)

6

PRO

The PhD-student case shakes out more like the AI case than convention admits on ICMJE criteria 1–3 (provenance ambiguity, supervisor-rewritten drafts, sign-off formality routinely ignored) and partly on criterion 4 (former students often as unreachable as a model instance). The applied exclusion of AI is selective, not principled.

FS2 + FS5

7

BOTH

AI-vs-kilo-author-human partial reduction. Same information-theoretic shape (named authorship with unclosable verification gaps); divergent legal-entity layer (kilo-author human remains a legal backstop; AI is not). A legal-entity backstop substitute is required for the AI case — Anthropic-the-organisation is the natural candidate.

FS4 + FS5 + FS7

8

CONV

The post-2023 trajectory is consolidation, not loosening. ICMJE, COPE, Nature, Science, JAMA, BMJ, Cell, PLOS, Elsevier, Springer Nature, and NEJM AI all converge on no-AI-author. No prominent academic voice in 2023–2026 has published a sustained formal argument for revision.

FS1 + FS6

9

PRO

5 of 7 b17 coping strategies repurpose cleanly as authorship-verification practices (sample-don’t-scan, checksums, adversarial probes, redundancy, preserve-slow-clock). The two that don’t (rate-distortion declaration, blast-radius tiering) need institutional infrastructure — not within b19’s unilateral reach but available at venue level.

FS4

10

PRO

Hosseini et al. 2025 — disclosure should be voluntary because mandatory disclosure biases reviewers against non-native English authors. The closest contemporary academic dissent; argues against mandatory disclosure, not directly for AI authorship, but opens the door.

FS1 + FS6

Section 4 — Pre-(b) framing recommendations#

Based on the cross-checked evidence, here is what the (b) EDEN steelmans need to argue. These are framing constraints, not conclusions.

What the conventional-position steelman must argue (and what it cannot rest on) ————————————————————————

Must argue:

  • That AI specifically cannot bear ongoing accountability in the way the four existing non-individual byline classes can — i.e., that the consortium / institutional / collective-pseudonym / individual- pseudonym mechanisms structurally fail when applied to AI.

  • That LLoL’s unilateral declaration of absorbing accountability is insufficient because Claude’s contribution is qualitatively different from the contribution that gets absorbed in (a) deceased- author cases (one-time, then frozen), (b) consortium cases (governance-distributed across thousands of humans), (c) corporate- liability cases (corporate veil + respondeat superior).

  • That the empirical absence of precedent (FS6) is itself load-bearing — not merely a sociological accident but evidence that the underlying structural reasons hold.

  • That AI-company self-authorship practice (no AI on any model paper’s byline) reflects the deepest practitioners’ own judgment.

Cannot rest on:

  • Personhood as the operative criterion (FS7 ruled this out).

  • The Bourbaki-1948 AMS rejection as a precedent in scientific authorship (FS7 ruled this domain-mismatched).

  • The claim that “AI tools are merely tools, like Mathematica” — the bandwidth-asymmetry case (FS4) and the substantive-contribution satisfiability (FS2, ICMJE criterion 1) take this off the table for a case like b19.

What the pro-co-authorship steelman must argue (and what it cannot rest on) ————————————————————————

Must argue:

  • That a specific named-absorber mechanism is declared explicitly in the paper itself: who absorbs forward accountability for Claude’s contribution. The three options on the table:

    1. LLoL alone as senior + corresponding author (deceased-author- rule analogue; LLoL has unilateral standing).

    2. Anthropic as the model’s operating organisation (corporate- liability analogue; requires explicit Anthropic consent, not currently in evidence).

    3. Both jointly (strongest form; requires Anthropic consent).

  • That the four-class precedent in non-individual byline authorship empirically rules out personhood-per-se as the blocker, and the AI case can be structured to satisfy the operative criterion (accountability) via mechanism 1, 2, or 3.

  • That the Bourbaki-1935 byline acceptance is the closest long-running historical analogue: a non-individual entity accepted as a byline name with a senior sponsor as accountable absorber.

  • That doing this as a deliberate framework-grounded position rather than a smuggling — with explicit absorption declaration in the title-page footnote — converts the empirical-unprecedence finding from a count-against to a contribution-of-precedent.

Cannot rest on:

  • The claim that the singularity / PraS requires AI co-authorship. PraS describes a bandwidth regime; it does not by itself entail authorship attribution.

  • The claim that ICMJE criterion 4 is “merely institutional inertia” — FS4 and FS5 both establish it as the load-bearing accountability criterion in formal practice. The argument has to address it structurally, not dismiss it.

  • The empirical-precedence record from FS6 — there is no precedent base to cite. The pro-position must own being substantively unprecedented.

What the (c) responsibility-allocation question must resolve#

The (c) question already largely answered by FS5: across all six frameworks, asymmetric naming requires a named absorber. The question for the b19 paper specifically is therefore:

  1. Who is the named absorber for Claude’s contribution? (Options 1, 2, 3 above.)

  2. What is the visible marker? (fn 7 is the natural place; the current fn 7 transparency disclosure could be expanded to carry the absorber declaration.)

  3. Does the named absorber have the standing to absorb? - LLoL: yes, unilaterally (as senior + corresponding author). - Anthropic: would need to be asked; standing is structural but

    consent is not in evidence.

The structurally honest answer is therefore: LLoL absorbs unilaterally, with optional Anthropic-acknowledgement supplement if sought and obtained.

Section 5 — Notes for the eventual footnote-forging step#

When the discussion reaches the title-page footnote stage (fn 6 + fn 7 expansion in AHA/study-title-page-footnotes-template.md), the findings above suggest the footnote pair must carry:

  • fn 6 (Anthropic affiliation) — current text is adequate as affiliation; if Anthropic-as-absorber path is chosen, fn 6 carries the absorption acknowledgement.

  • fn 7 (Claude bio / AI co-authorship rationale) — must carry:

    • the PraS one-liner (clarifier preventing Hollywood-conflation),

    • the named-absorber declaration (LLoL absorbs forward accountability, or LLoL + Anthropic jointly),

    • a brief structural-precedent acknowledgement (deceased-author rule and consortium-byline rule as the operative precedents — not a defensive citation of every fact-sheet, but the one or two that make the move legible to a sceptical reader).

Three candidate one-liners for PraS (kept open, to be finalised at the footnote-forging step):

  • A (long, ~36 words): “Practical singularity (PraS) here means the per-individual, per-topic regime where AI useful-generation bandwidth has outpaced the human’s review-and-incorporate bandwidth — not the global, recursive-self-improvement version.”

  • B (footnote-grade, ~15 words): “Practical singularity: per- individual, per-topic — AI generation outpaces human review; not the global singularity claim.”

  • C (PraS-spelled-out-in-sequence, new): Per-Individual Practical Singularity — the per-individual, per-topic bandwidth-asymmetry regime; distinct from any global recursive- self-improvement claim. (count words at the footnote-forging stage.)

Section 6 — Methodology limitations across all 7 fact-sheets#

The cross-consistency check inherits the following limitations:

  1. ``WebFetch`` sandbox-blocked across all seven subagent runs. Every quoted operative sentence may be training-reconstructed. Convergent-evidence claims (Section 2 above) are robust against this; single-sourced claims (e.g., the exact wording of NEJM AI’s policy, the Élie Cartan sponsor-vouching mechanism for Bourbaki, the precise ICMJE criterion 4 wording) need re-checking against live URLs before any quotation enters the published b19 paper.

  2. PraS-adjacent reference candidates (Licklider 1960, Engelbart 1962, Clark & Chalmers 1998, Brynjolfsson 2022, Kasparov 2017) are held for end-of-discussion WebSearch verification per LLoL’s request.

  3. Anthropic-as-absorber path is structurally available but consent is not in evidence. The cross-check cannot tell whether Anthropic would consent if asked.

  4. The b19-specific structural fit has not been argued in this cross-check; that is what the (b) steelmans will do.

Note

This cross-consistency check is the input to the (b) EDEN steelmans. The steelmans should cite this check by section number and should not duplicate its content. When (b) is complete, this file remains the cross-cut reference; (b) is the verdict.