Candidate edition: Publication candidate prepared 2026-11-11 from the unchanged v0.3 research dossier at repository 7fc15ec249fba391018310ade811c5bd24e1099f. This is a proposed editorial successor of that dossier, not an adopted publication or a new dossier identity. Original dossier SHA-256 72a0f5ab049e606f26de5a305415a3d72b32d17f49e147e3130f130b242fa3d2; 34,198 bytes. Original title/version, source handles and uploaded-input lineage remain below. Candidate scope: Shared-reader grammar and bounded analysis, using the existing 6–7 November retained-source readings only. No fresh retrieval, literature search, custody readback, identity assignment, source admission, route registration or deployment occurred. Source inventories preserve unresolved research pointers; proposed lead and qualified claims identify their narrower basis.

Status: PUBLICATION CANDIDATE / RESEARCH-ONLY UNTIL OWNER DISPOSITION / LOCAL SOURCE BYTES PARTLY INSPECTED / NO GOVERNED CAPTURE
Subject: "Attention Is All You Need" (Vaswani et al., 2017) — seven arXiv versions and a NeurIPS proceedings edition whose reported results differ across representations and internal passages, what the paper claimed, what individual public reviews said, and what later use of the Transformer causes readers to remember it as having established
Scope: source inventory, authority/posture, record-contradiction map, provisional five-claim matrix, capture hazards, reverse-source candidates
Research date: 2026-11-06
Supersedes: attention-is-all-you-need-demo-corpus-dossier-v0-2 (2026-11-06), which superseded v0.1. This corrected copy of the uploaded v0.3 was reconciled on 2026-11-06 with the completed local source pull. All seven PDFs are now hash-distinct and read; the score comparison includes their internal passages; the unsupported historical-HTML inference is withdrawn; the proceedings PDF is read from the official source; and parsing is attributed to the arXiv representations that contain it. The eight-author set, individual-review posture and PC-02 reference-versus-identity distinction remain. Preserved input SHA-256: 1f69b70ef3e03de0c5c465e33a627fadb805f86b6040b32f2e4978da63b9f941.
Lineage: derives from and does not replace the Personal Canon design dossier personal-canon-pc02-one-paper-three-researchers-aiayn-dossier-v0-1 (claim objects AIAYN-C01 to C09; sources PAPER-S01 to S05). This dossier is the public demo-corpus restatement in the shared reader grammar; where the two overlap, PC-02's claim identifiers are referenced, not re-minted.
Drafting basis: the initial inventory used web readings on the research date. This reconciliation also uses retained bodies and SHA-256 readbacks in the 2026-11-06 source pull (SOURCE_MANIFEST.json, downloads/<source_key>.json and reviews/science-byte-inspection.json). The selected arXiv PDFs, official proceedings page/PDF, public reviews and two prior/later-work preprints have been locally read. These are editorial research readings, not native Acquisition receipts, accepted source identities or governed bindings. Remaining VERIFY items name unperformed checks rather than denying the bytes already acquired.
Explicit non-claims: This dossier creates no CaptureReceipt, SourceEdition, SRS receipt, custody, admission, final standing, adopted article prose, or canonical Counterpedia identity. It decides nothing about the paper's merit.

A. SUBJECT / DISAMBIGUATION

Common name

Attention Is All You Need (the Transformer paper)

Exact scope

The 2017 paper by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser and Illia Polosukhin, in its arXiv editions (1706.03762, v1 12 June 2017 through v7 2 August 2023) and its NeurIPS 2017 proceedings edition (Advances in Neural Information Processing Systems 30, paper ID 3058), together with the public NeurIPS reviews; the specific quantitative and architectural claims each edition makes (an architecture without recurrence or convolution, built from attention and feed-forward layers); and the gap between those claims and what the paper is commonly cited as having shown.

This dossier is not a history of the Transformer, a survey of later models, an account of the authors' later careers, or a judgment about which edition is "correct".

Important identifiers / terms

• arXiv:1706.03762; versions v1–v7; DOI 10.48550/arXiv.1706.03762

• NeurIPS 2017 (NIPS 2017), Long Beach, 4–9 December 2017; proceedings paper ID 3058; proceedings hash 3f5ee243547dee91fbd053c1c4a845aa

• Transformer; self-attention; multi-head attention; scaled dot-product attention; positional encoding

• WMT 2014 English-to-German and English-to-French; newstest2014; BLEU

• Transformer (base) and Transformer (big)

• "3.5 days on eight P100 GPUs"

• English constituency parsing (WSJ / Penn Treebank)

• Equal-contribution footnote; "listing order is random"

Naming / semantic traps

1. "The paper" names several editions and representations. All seven arXiv PDFs were acquired, their embedded v1–v7 labels read, and their SHA-256 digests found to differ. The proceedings PDF and its web abstract are separate representations. A version label or rounded listing size is not a digest; the actual byte comparison is recorded under S02. Different binary hashes do not, by themselves, establish different scientific claims: normalized extracted v6/v7 text differed only in the embedded version/date in this inspection.

2. "The Transformer got 41.8 BLEU on English-to-French" needs a representation and passage. In arXiv v7, the abstract and Table 2 report 41.8, while §6.1 prose reports 41.0. The proceedings HTML abstract reports 41.1; its PDF reports 41.0. Earlier PDFs add further differences: v2 has 41.2 in the abstract, 41.0 in Table 2 and 41.17 in §6.1; v3 has 41.0, 41.0 and 41.17. The complete selected-PDF readings are in S02. These are attributed statements, not an adjudication that every number represents an identical experimental configuration.

3. "28.4 BLEU on English-to-German" is also edition-dependent. The proceedings abstract page reports 27.5; arXiv v7 reports 28.4.

4. arXiv is not the venue. The paper was refereed at NeurIPS 2017; its reviews are public. The arXiv record is the version history, not the peer-reviewed edition.

5. "Introduced attention" is wrong. The paper's background cites earlier attention mechanisms, including Bahdanau et al. (S07). Its architectural claim removes recurrence and convolution from the sequence model while retaining attention and feed-forward layers; the title does not erase those other components.

6. "Proved attention is all you need" is a title, not a result, and "attention only" is not the architecture. The Transformer removes recurrence and convolution from the sequence model; it retains feed-forward layers, residual connections, layer normalization, embeddings and positional encodings alongside attention. The inspected arXiv versions contain two translation tasks and an English constituency-parsing task; the inspected proceedings PDF does not contain that parsing experiment. Its own text leaves restricted attention for very long sequences and other modalities to future work (PC-02 AIAYN-C06, C07).

7. "The paper introduced GPT / BERT / LLMs" is anachronism. Those are later works that used the architecture. The paper claims translation quality, parallelizability and training cost.

8. "Eight Google authors" is approximately right and the footnote is more careful. The paper lists equal contribution with random order and per-author contributions; two authors are marked as having done the work while at Google (one University of Toronto affiliation is listed). Read the footnote in the captured edition rather than paraphrasing it.

9. "Most-cited paper of the century" is a dated, source-specific ranking. Citation counts and rank lists change by database and date; any such sentence needs its source and date (see S08, OPTIONAL).

10. Training cost is version- and configuration-specific. The v7 abstract associates 3.5 days on eight GPUs with the big English-to-French model; the PDF specifies P100 hardware. The v2 abstract instead says 4.5 days. Preserve the edition and experiment rather than describing one training-time number as a property of Transformers (S02).

Temporal bounds

The selected arXiv editions range from v1 on 12 June 2017 through v7 on 2 August 2023. The proceedings are NeurIPS/NIPS 2017. The seven PDF versions, v1/v7 listings, proceedings PDF/HTML and public review page were retained and read on 6 November 2026; exact passage comparisons were additionally recorded on 7 November. This 11 November candidate makes no new source-retrieval observation.

Version dates and retrieval dates answer different questions. The proceedings HTML reading is a 2026 observation of a proceedings page, not proof of its publication-day 2017 contents. The original unsupported historical-HTML inference remains withdrawn. A dated v5 PDF cannot supply a missing historical HTML snapshot.

The separate PC-02 design dossier remains research_artifact_only and is not superseded or assigned new claim identity here. This publication candidate proposes a distinct Wave 2 dossier target, pending owner choice of route and disposition. Existing Stage AIAYN articles and their architectural HTML bindings remain separate. Later works and optional citation rankings retain their own dates; no current citation ranking or universal scaling conclusion is asserted.

Analysis

A citation to the paper does not select one result. In the retained arXiv v7 PDF, the abstract and Table 2 report 41.8 English-to-French BLEU while §6.1 prose reports 41.0. The proceedings HTML abstract and its PDF also differ. The reader can identify which passage supports a precise attribution without deciding that all passages describe identical experimental configurations or choosing an unsupported corrected value.

The inspected representations also delimit the experimental account. All seven retained arXiv PDFs include a constituency-parsing experiment; the inspected proceedings PDF does not. This establishes a difference between the selected documents, not what every reviewer saw or which edition a later citer used. Distinct binary hashes establish distinct selected byte objects, not seven different scientific findings.

The paper’s architectural proposal, earlier attention work, individual reviewer judgments and later engineering work remain separate. Earlier work is cited by the paper; the public reviews contain both favorable assessments and questions. A reviewer’s statement about reimplementation is not an independently verified reproduction record. Later success on other tasks needs the works that report it, rather than being attributed retrospectively to the 2017 experiments.

B. SOURCE INVENTORY

AIAYN-S01 — arXiv abstract record and submission history (PC-02 PAPER-S01)

Exact title: Attention Is All You Need
Issuer: arXiv (cs.CL; cs.LG)
Identifier: arXiv:1706.03762
Submission history (as read on the research date): v1 12 Jun 2017 17:57 UTC; v2 19 Jun 2017; v3 20 Jun 2017; v4 30 Jun 2017; v5 6 Dec 2017; v6 24 Jul 2023; v7 2 Aug 2023 (submitter shown: Llion Jones)
Landing: https://arxiv.org/abs/1706.03762
Current abstract (v7) states: 28.4 BLEU on WMT 2014 English-to-German; "a new single-model state-of-the-art BLEU score of 41.8" on English-to-French "after training for 3.5 days on eight GPUs".
v1 abstract: the retained version-specific listing and PDF both report 41.0 for English-to-French. The PDF's embedded v1 date is 12 June 2017; the listing was retrieved on 6 November 2026.
Historical HTML limit: the earlier draft asserted that an October 2022 Wayback page proved the listing abstract had changed by 6 December 2017. No exact snapshot URL or retained Wayback body supports that assertion here, so the inference is withdrawn. Even a 2022 snapshot labeled v5 would establish what that page showed in 2022, not necessarily its state in 2017. Independently, v5 is the first of the seven inspected PDFs whose abstract and Table 2 report 41.8 (S02); that is a PDF finding, not a historical-HTML finding. Any historical listing-change claim remains VERIFY against its own dated artifact.

Authority / posture: repository metadata for version labels and submission dates, not a byte digest or peer-review record. Exact retained PDF bytes are identified separately under S02.

Local research reading: aiayn-v1-html, SHA-256 f9016d878e3f7a21b8b48852a964176029078fc51a2df6fc1f1f11d0698e5baf; aiayn-v7-html, SHA-256 0204ac433469e325e3e23726af11946de5e21dc82433c8682ca3d0910798100f. The v1 and v7 listing pages were acquired on 2026-11-06.

May support

• the existence and dates of seven editions;

• the headline numbers as stated in the v7 abstract.

Cannot establish

• which edition any citing work read;

• that the body table agrees with the abstract (see S02).

Captureability: HIGH
Priority: P0

AIAYN-S02 — Exact arXiv PDF editions (PC-02 PAPER-S02)

Exact title: Attention Is All You Need, per version
Exact PDF locators: https://arxiv.org/pdf/1706.03762v1 through https://arxiv.org/pdf/1706.03762v7

Authority / posture: author-posted PDF editions. All seven were retained, parsed and checked against their embedded version labels on 2026-11-06. Their distinct hashes establish different selected byte objects, not seven different scientific results or accepted governed identities.

Local research reading — English-to-French BLEU: abstract on physical p. 1; Table 2, Transformer (big), and §6.1 prose on physical p. 8 in each file. Each line lists abstract / Table 2 / §6.1 prose:

• v1: 41.0 / 41.0 / 41.0; 15 pages; abstract training time 3.5 days.

• v2: 41.2 / 41.0 / 41.17; 16 pages; abstract training time 4.5 days.

• v3: 41.0 / 41.0 / 41.17; 16 pages; abstract training time 3.5 days.

• v4: 41.0 / 41.0 / 41.0; 15 pages; abstract training time 3.5 days.

• v5: 41.8 / 41.8 / 41.0; 15 pages; abstract training time 3.5 days.

• v6: 41.8 / 41.8 / 41.0; 15 pages; abstract training time 3.5 days.

• v7: 41.8 / 41.8 / 41.0; 15 pages; abstract training time 3.5 days.

Every inspected PDF abstract reports 28.4 for English-to-German. The differing English-to-French statements are not silently reconciled into one result; their model/task context must accompany any comparison. This reading does not establish why the authors changed a value, whether every passage describes identical experimental conditions, or which statement should prevail.

Retained PDF identities:

• aiayn-v1-pdf, SHA-256 0a9bff1be3575cf9fe26a4a3d3a9175bc791ecec90de8b9deba4cff75d2b9476.

• aiayn-v2-pdf, SHA-256 803b95695f178d1804b8cf66ab7ae3d7c8b94e044a045cbff8d9b37435ca4dd4.

• aiayn-v3-pdf, SHA-256 aa7a3201a45567fda42fedf050c1e1a397b51bcd8f5c4a315abf232670d721cd.

• aiayn-v4-pdf, SHA-256 df285e492fc7c0169588493018fe473ee1fe9a22e525e80c1ada7f98535af729.

• aiayn-v5-pdf, SHA-256 bfaaec89262875f927cf1b38b2da2d775f3309b7bea3537f29b606ca67e79065.

• aiayn-v6-pdf, SHA-256 3c3ac946023fe79165ffcc4fb0696a6281da5f8ee32cbbc0a7c0677400891a41.

• aiayn-v7-pdf, SHA-256 bdfaa68d8984f0dc02beaca527b76f207d99b666d31d1da728ee0728182df697.

The retained v6 and v7 PDFs are hash-distinct; the NFKC/whitespace-normalized extracted-text comparison found only their embedded version/date changed. That bounded text comparison is not a complete visual or scientific-equivalence finding. All seven arXiv PDFs contain a constituency-parsing experiment (v1 physical p. 8; v2–v7 p. 9); the proceedings PDF differs on this point (S03).

May support

• the abstract, table and prose values as stated in a named PDF passage;

• the per-edition contribution and affiliation footnotes;

• the architecture, long-sequence and future-work wording at its actual location.

Cannot establish

• that every distinct hash represents a distinct scientific revision;

• the edition or passage read by a later citer, or the cause of an internal discrepancy.

Captureability: all seven selected PDFs locally acquired and read; governed acquisition remains separate
Priority: P0

Publication-candidate retained-reading note: The 7 November readback rechecked the seven PDF bodies against their source-card digests and located the abstract on physical p. 1, Table 2’s Transformer (big) row and §6.1 results prose on physical p. 8. These remain editorial research locators, not minted passage identities or governed PDF bindings.

AIAYN-S03 — NeurIPS 2017 proceedings record (PC-02 PAPER-S03)

Exact title: Attention is All you Need
Issuer: Advances in Neural Information Processing Systems 30 (NIPS 2017), Curran Associates
Identifier: proceedings hash 3f5ee243547dee91fbd053c1c4a845aa; paper ID 3058
Landing inspected: https://papers.nips.cc/paper_files/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html
Exact PDF inspected: https://papers.nips.cc/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
Proceedings abstract page states: 27.5 BLEU English-to-German and 41.1 English-to-French, describing a 165-million-parameter single model.

Authority / posture: official proceedings representations of the refereed work. The acquired HTML abstract reports 27.5/41.1; the acquired 11-page PDF reports 28.4/41.0 in its abstract and Table 2, and 41.0 in §6.1 prose. Unlike all seven inspected arXiv PDFs, this proceedings PDF does not contain the parsing experiment. That absence alone establishes nothing about every document seen by reviewers. Cite the representation and passage.

Local research reading: aiayn-proceedings-abstract-html, SHA-256 42974e1042efb1f9127b6a8bd910eedd8e925b89c4de2b3653ecba00ab2f3cf0; aiayn-proceedings-pdf, SHA-256 d87d482d5ae7960e2e43d7dd6d21377e60e73e8fce1bf2a01aff7aca8a08c537 (abstract p. 1; Table 2 and §6.1 p. 8). Both official representations were retained and read on 2026-11-06; the finding no longer depends on a metadata extraction or university mirror.

May support

• the refereed edition's reported numbers;

• the edition-difference finding when placed beside S01/S02.

Captureability: HIGH
Priority: P0

Publication-candidate retained-reading note: The same retained readback verified the proceedings PDF at abstract physical p. 1 and Table 2/§6.1 physical p. 8. The independently pinned HTML abstract remains a different representation, read in November 2026, not a historical 2017 HTML snapshot.

AIAYN-S04 — NeurIPS 2017 public reviews (PC-02 PAPER-S04)

Exact title: Reviews for paper ID 3058, Attention is All you Need
Issuer: NeurIPS proceedings (public review record)
Landing inspected: https://papers.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Reviews.html

Local research reading: aiayn-proceedings-reviews-html, SHA-256 c2a2be17951249982f109471fa3ae5b593812b4b32e4b05af4fedf6b31a69631; retained and read on 2026-11-06.

Authority / posture: three reviewer texts published with the proceedings; each is one reviewer's judgment and request, not a collective certification of the paper's claims. Reviewer 1 reports reimplementing the work and calls it "a clear accept" while noting the learning-rate schedule "seems to really matter" (a reviewer's statement about reimplementation, not a separate reproduction record); Reviewer 2 notes the architecture advances state of the art by about 1 BLEU and asks about dropout and a dangling reference; a reviewer notes the lack of mathematical definition of multi-head attention and asks for evaluation on long sentences and for stronger validation of the long-range-dependency argument.

May support

• what the reviewers judged established at acceptance;

• the open points the reviewers recorded (long sequences, hyperparameters), which match the paper's own deferrals.

Cannot establish

• anything about later editions or later reputation.

Captureability: HIGH
Priority: P0

AIAYN-S05 — Google Research publication page

Exact title: Attention is All You Need
Issuer: Google Research (publications)
Landing: https://research.google/pubs/pub46201/

Authority / posture: institutional republication page carrying an abstract (reads 28.4 BLEU English-to-German and 41.0 English-to-French as retrieved on 2026-11-06, which matches the proceedings PDF abstract, not the arXiv v7 listing).

May support

• another institutional abstract representation for the comparison, not a new scientific edition inferred merely from a different host.

Local research reading: this page was not retained in the completed pull. Its prior web reading remains distinct from the inspected S01–S04 bytes.

Captureability: HIGH
Priority: P1

AIAYN-S06 — FlashAttention (PC-02 PAPER-S05)

Exact title: FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness
Issuer: arXiv
Authors: Tri Dao et al.
Identifier: arXiv:2205.14135v1 (2022), the selected representation in this pull
Landing inspected: https://arxiv.org/abs/2205.14135v1
Exact PDF inspected: https://arxiv.org/pdf/2205.14135v1

Local research reading: aiayn-flashattention-v1-pdf, SHA-256 be21f20c749ba8443988b479bece93617dd37b214ea9099413f951f756d4104f; version-specific HTML and PDF retained and read. This selected preprint is not automatically the later proceedings edition or the edition intended by another producer.

Authority / posture: later engineering work on attention's memory cost; included as an example of what later work added that readers fold back into the 2017 paper.

May support

• the "later success remembered as original establishment" distinction.

Captureability: HIGH
Priority: P2

AIAYN-S07 — Bahdanau, Cho, Bengio (2014) attention mechanism

Exact title: Neural Machine Translation by Jointly Learning to Align and Translate
Issuer / selected representation: arXiv:1409.0473v1 (2014); the later ICLR 2015 publication is a separate edition
Landing inspected: https://arxiv.org/abs/1409.0473v1
Exact PDF inspected: https://arxiv.org/pdf/1409.0473v1

Local research reading: aiayn-bahdanau-v1-pdf, SHA-256 52391f14a0de04fdcbf2854e0349df5557ae01db0f37845e0aba96f0f1a9a2b1; version-specific HTML and PDF retained and read. No byte-equivalence with the ICLR edition is asserted.

Authority / posture: prior work the paper cites; evidence against the "introduced attention" reading.

Captureability: HIGH
Priority: P1

AIAYN-S08 — Citation-count and ranking surfaces — OPTIONAL

Candidates: Wikipedia article (states more than 250,000 citations and a top-ten-of-century rank, as of its own revision); Semantic Scholar or Google Scholar count; Nature's 2025 analysis of most-cited twenty-first-century papers (VERIFY identity and rank)
Landing: https://en.wikipedia.org/wiki/Attention_Is_All_You_Need

Authority / posture: tertiary and database-dependent; usable only with source and date attached.

Priority: OPTIONAL

C. RECORD-CONTRADICTION / QUALIFICATION MAP

C-1 — Different English-to-French BLEU statements across and within representations

The retained v7 listing, PDF abstract and PDF Table 2 report 41.8; the same PDF's §6.1 prose reports 41.0. The proceedings HTML reports 41.1; its PDF abstract, Table 2 and §6.1 report 41.0. Earlier PDFs show additional differences, including v2's 41.2/41.0/41.17 and v3's 41.0/41.0/41.17 at those three locations (S02). v5 is the first of the seven inspected PDFs whose abstract and table report 41.8. That is not proof of when a historical HTML page changed. Each statement carries its representation, passage and model/task context; the reason for a difference and which result should prevail are not established by reading the numbers.

C-2 — 28.4 vs 27.5 (English-to-German BLEU)

S01: 28.4. S03 abstract page: 27.5 with a 165M-parameter single model. Same qualification.

C-3 — "Introduced attention" vs "based solely on attention"

S07 predates the paper and is cited by it. The paper's novelty claim is architectural (no recurrence or convolution), not the invention of attention.

C-4 — Reviewer caveats vs later reputation

S04 records individual reviewers' open points (long sequences, hyperparameter sensitivity, the long-range-dependency argument) beside favorable judgments. Later reputation treats the architecture as general; the paper deferred that, and its reviewers asked for it.

C-5 — Proceedings PDF vs proceedings web abstract

Two objects at one venue carry different abstract text (confirmed: page 27.5 and 41.1; PDF 28.4 and 41.0). "The NeurIPS version" is itself two surfaces; a citation that names the venue without the surface is incomplete.

C-6 — Title vs claim

"All you need" is a title. The inspected arXiv PDFs report two translation tasks and one English constituency-parsing task; the inspected proceedings PDF reports the translation experiments and does not contain that parsing experiment (S02/S03). Keep the experiment attached to the representation that includes it.

C-7 — Training cost vs generality

In v7, the 3.5-day statement attaches to the big English-to-French model and the PDF names P100 hardware; v2's abstract instead reports 4.5 days. Quote the selected experiment, hardware and edition. Neither value is a general property of the architecture (S02).

D. FIVE-CLAIM DEMONSTRATION MATRIX — PROVISIONAL ONLY

These are the same five local C1–C5 research candidates expanded in the shared-reader grammar, with no authored canonical identity. The design dossier’s AIAYN-C01–C09 references remain navigation only. Full AIAYN-S handles below identify this dossier’s existing source cards; they are not the Stage AIA/AIAE handles or governed Claim-to-SourceEdition bindings.

C1 — Edition chronology and representation identity

Provisional reader wording

The selected arXiv record lists seven versions of Attention Is All You Need, from 12 June 2017 through 2 August 2023; the work also has a NeurIPS 2017 proceedings edition with separate HTML and PDF representations.

Candidate research source cards, not new bindings

• AIAYN-S01

• AIAYN-S02

• AIAYN-S03

What this analysis can establish

The retained version listings and embedded PDF labels identify the selected edition dates. All seven PDFs have different SHA-256 digests; the proceedings PDF and HTML are separately identified bytes. The authorship set is the eight authors named in Part A.

What it cannot establish / reason to qualify a stronger claim

Different hashes do not establish seven substantive scientific revisions. The recorded normalized v6/v7 text comparison found only their embedded version/date changed at that comparison scope. No publication-day state of the later-retrieved HTML or edition used by a later citer is inferred.

Standing and next check

Keep the exact selected representation and date role. No PC-02 or Stage identity is transferred to these local candidates. No governed Claim, SourceEdition, Claim-to-SourceEdition relationship, Check outcome or standing is created by this provisional analysis.

C2 — BLEU values by passage and representation

Provisional reader wording

In arXiv v7, the English-to-French abstract and Table 2 report 41.8 BLEU while §6.1 prose reports 41.0; the proceedings HTML abstract reports 41.1 while its PDF abstract, table and prose report 41.0.

Candidate research source cards, not new bindings

• AIAYN-S02

• AIAYN-S03

What this analysis can establish

For all seven selected arXiv PDFs, the recorded comparison uses abstract physical p. 1 and Table 2’s Transformer (big) row/§6.1 prose on p. 8. In abstract/table/prose order: v1 41.0/41.0/41.0; v2 41.2/41.0/41.17; v3 41.0/41.0/41.17; v4 41.0/41.0/41.0; v5–v7 41.8/41.8/41.0. The proceedings PDF is 41.0/41.0/41.0 at those locations; its separately retained HTML abstract is 41.1. These are attributed WMT 2014 English-to-French passage readings.

What it cannot establish / reason to qualify a stronger claim

The values alone do not establish identical experiment/model conditions, an author-issued correction or which value should prevail. English-to-German is a different task: the PDF abstracts report 28.4 while the proceedings HTML reports 27.5. A model description in one passage is not silently transferred to every other passage.

Standing and next check

The selected local comparison is complete at its stated scope. Preserve representation, internal location and task/model qualifiers; leave corrected-result and experimental-equivalence inferences unresolved. No governed Claim, SourceEdition, Claim-to-SourceEdition relationship, Check outcome or standing is created by this provisional analysis.

C3 — Individual public reviews

Provisional reader wording

The official proceedings review page contains three reviewer texts; Reviewer 1 reports reimplementation and calls the paper a clear accept, while the reviews also record questions and caveats.

Candidate research source cards, not new bindings

• AIAYN-S04

What this analysis can establish

The retained public review page supports attribution of a particular reviewer’s judgment and stated activity. The existing reading records learning-rate sensitivity and other questions about the architecture, evaluation and presentation alongside favorable assessments.

What it cannot establish / reason to qualify a stronger claim

Three reviews are not a collective certification of every claim. A reviewer’s statement about reimplementation is not a separately inspected independent-reproduction record, and the review page does not establish that every reviewer saw the exact selected proceedings PDF.

Standing and next check

Retain each speaker’s attribution. Do not promote reported reimplementation into an independently verified result. No governed Claim, SourceEdition, Claim-to-SourceEdition relationship, Check outcome or standing is created by this provisional analysis.

C4 — Architectural and experimental scope

Provisional reader wording

The arXiv PDFs describe a sequence architecture without recurrence or convolution while retaining attention, feed-forward layers and other components; their selected experiments include WMT 2014 translation and English constituency parsing.

Candidate research source cards, not new bindings

• AIAYN-S02

• AIAYN-S03

• AIAYN-S07

What this analysis can establish

The inspected arXiv versions carry the parsing experiment (v1 physical p. 8; v2–v7 physical p. 9), while the inspected proceedings PDF does not contain it. The paper’s own discussion defers restricted attention for very long sequences to future work. The earlier Bahdanau preprint is separately retained prior work cited by the paper.

What it cannot establish / reason to qualify a stronger claim

The title does not mean attention is the only component or that the paper invented attention. The experiments do not establish performance on every task or modality. Absence of parsing in the inspected proceedings PDF is a representation finding, not an account of every document in the review process.

Standing and next check

Preserve the distinction between prior work, the selected paper’s assertions and demonstrated tasks, and later claims. PC-02 references remain navigation, not claim identity or standing. No governed Claim, SourceEdition, Claim-to-SourceEdition relationship, Check outcome or standing is created by this provisional analysis.

C5 — Later reputation and unsupported generalization

Provisional reader wording

“Attention Is All You Need proved that Transformers are all you need for AI” goes beyond what the selected paper representations and public reviews establish.

Candidate research source cards, not new bindings

• AIAYN-S02

• AIAYN-S03

• AIAYN-S04

• AIAYN-S06

What this analysis can establish

The selected record permits a bounded account of the proposed architecture, reported translation/parsing tasks at their actual representations, and reviewer judgments. The selected FlashAttention v1 preprint is an example of separately dated later work on attention’s memory cost; its findings belong to that work.

What it cannot establish / reason to qualify a stronger claim

The compressed formulation is an editorial question, not a sourced universal claim. Neither the title nor later reputation proves general scaling, every later modality or present citation rankings. The selected later preprint is not automatically its subsequent proceedings edition.

Standing and next check

Attach later results to their own works and editions. Omit unsourced popularity/ranking claims and any universal experimental conclusion from the proposed lead. No governed Claim, SourceEdition, Claim-to-SourceEdition relationship, Check outcome or standing is created by this provisional analysis.

E. BEST "WHY NOT?" SPECIMEN

Popular formulation

"The Transformer paper reported 41.8 BLEU on English-to-French."

Why not

Qualify. The v7 abstract and Table 2 do report 41.8, as does its listing; that documentary statement is not made false by a different number elsewhere. Its §6.1 prose reports 41.0. The proceedings HTML reports 41.1, and its PDF abstract, table and prose report 41.0 (S01–S03). State which passage is being cited and preserve the disagreement rather than asserting one unqualified score for every representation.

So: the paper reports X needs which edition, which representation, which internal passage, and which task/model conditions.

Reopening condition

The local read-and-hash comparison is complete for the selected seven arXiv PDFs and the proceedings PDF/HTML; its readings are in S02/S03. A precise sentence can now identify the passage that reports 41.8 and acknowledge the internal disagreement. A stronger claim that it is the corrected result, or that another value describes an identical experiment, needs additional author-issued or experimental evidence. Governed acquisition and any public display authorization remain separate from this local comparison.

F. SECOND "WHY NOT?" — WHAT WAS ESTABLISHED

Popular formulation

"The paper introduced attention."

Why not

The paper cites earlier attention work (S07); its proposed sequence architecture removes recurrence and convolution while retaining attention and feed-forward layers. Reviewer 1 regarded the combination, rather than striking novelty in each underlying technique, as the achievement (S04). That is the reviewer's judgment, not an independent reproduction result.

Counterpedia reveal The paper's own background section and its reviewers already draw the line the popular sentence erases.

G. THIRD "WHY NOT?" — LATER SUCCESS

Popular formulation

"The 2017 paper showed Transformers scale to everything."

Why not

The inspected arXiv PDFs report two translation directions and an English constituency-parsing task, and defer restricted attention for very long sequences and other modalities to future work (S02; PC-02 AIAYN-C06, C07). The inspected proceedings PDF does not contain the parsing experiment. Later claims about scaling or other modalities require their own works; S06 is one selected example of later work on attention's memory cost, not evidence for every later generalization. Attribute each result to the work and representation that reports it.

H. SOURCE-PAGE / REVERSE-WIKIPEDIA VALUE

1. S01 + S02 — the version history and the seven PDFs

A comparison of seven hash-distinct PDFs whose selected passages differ in reported values. Binary difference, textual change and scientific change remain separate claims.

2. S03 — the proceedings edition

A refereed edition whose public abstract carries different headline numbers from the preprint abstract.

3. S04 — the public reviews

A rare object: what the referees said was established, with their open points, readable beside the paper.

4. S07 — the prior work

The citation inside the paper that defeats "introduced attention".

I. 30–60 SECOND DEMO MOMENT

Open the paper and show four named representations/passages, with paraphrased labels:

arXiv v7 PDF, abstract and Table 2 — embedded version date 2 August 2023

English-to-French BLEU 41.8; English-to-German 28.4.

The same arXiv v7 PDF, §6.1 prose

English-to-French BLEU 41.0 for the big model.

NeurIPS 2017 proceedings HTML abstract — read 6 November 2026

English-to-French BLEU 41.1; English-to-German 27.5; a 165-million-parameter single model.

NeurIPS 2017 proceedings PDF, abstract, Table 2 and §6.1 — read 6 November 2026

English-to-French 41.0; abstract and Table 2 English-to-German 28.4.

Ask:

What BLEU score did "Attention Is All You Need" report?

Counterpedia answers:

The v7 abstract and table report 41.8 for English-to-French, while its results prose reports 41.0. The proceedings HTML and PDF differ too. These are real statements at different locations; identify the representation, passage and experiment instead of selecting one number for "the paper". Earlier versions show further internal differences, recorded in the source inventory.

Then expose the source identity, internal location, model/task conditions, and the difference between a version date and the date a live web page was read. All proposed Entry displays remain subject to the applicable acquisition and publication steps.

J. CAPTURE PRIORITY

All P0 representations below have local, hashed, inspected bodies in the completed pull. The native Acquisition/identity/publication steps remain separate; this list no longer asks a future reader to discover whether those PDFs exist. S05, optional citation surfaces and the asserted Wayback snapshot were not retained.

P0 — essential

1. S01 arXiv abstract page with full submission history

2. S02 all seven arXiv PDFs (v1–v7), exact bytes

3. S03 proceedings PDF and the proceedings web abstract page

4. S04 the reviews page

P1 — important corroboration / contradiction

1. S05 Google Research publication page

2. S07 Bahdanau et al. arXiv:1409.0473v1, the locally acquired selected preprint; any later conference edition remains separate

P2

1. S06 FlashAttention arXiv:2205.14135v1, the locally acquired selected preprint; do not infer another producer's edition selection

OPTIONAL

1. S08 citation surfaces with capture dates

K. INTEGRITY FLAGS

1. Edition identity is the subject

A sentence about reported numbers names the edition, representation and internal passage, with model/task conditions. Edition alone does not resolve an internal discrepancy.

2. Abstract, table and results prose may disagree within one edition

Read all three at named locations. For v7, the abstract and table agree at 41.8 while the prose says 41.0; v2/v3 have additional differences. Preserve those readings without declaring which is corrected.

3. Individual reviews are attributed judgments

Keep the three reviewer texts distinct from one another and from later reputation. A reported reimplementation remains the reviewer’s statement, not a separately verified reproduction record or collective certification.

4. "Introduced attention" is defeated by the paper's own citations

Render the prior work the paper cites.

5. Citation counts are dated and database-specific

Any count carries source and date or is omitted.

6. PC-02 is the claim owner, and a reference is not a mapping

Claim identifiers AIAYN-C01 to C09 are PC-02's; this dossier references them and mints none. The five local candidates the projector derives from Part D carry authored_id: null; a citation in prose does not identify them with PC-02's claims. A producer mapping, where warranted, is explicit and separate.

7. Author footnotes vary by edition

Contribution and affiliation footnotes are to be read per edition, not paraphrased from memory.

8. Locators marked VERIFY are unresolved

Resolve every VERIFY item relied on by the proposed lead or demo before stamping its sidecar; omit an unresolved assertion outside the selected scope. Preserve completed local checks and real remaining gaps separately. A declared sidecar basis array is not a per-sentence proof.

L. PROVISIONAL ATTENTION-IS-ALL-YOU-NEED THESIS FOR REVIEW

A canonical paper is not one byte object or one undifferentiated statement. The selected "Attention Is All You Need" PDFs are hash-distinct, and reported values differ across representations and within several PDFs. Public reviews carry individual judgments; the arXiv and proceedings PDFs do not include exactly the same experiments. Counterpedia can let a reader inspect what a named passage reports, what its experiment establishes or leaves open, and which later claims belong to other works. "The Transformer paper reported X" needs its edition, representation and internal passage, with the model/task conditions; those identifiers make the statement inspectable without deciding which conflicting value is correct.