% ============================================================================
% Part III — Data and Systematic Review Protocol
% Written by Agent L (Part III instance), 2026-11-03. Every corpus number is a macro from tables/numbers.tex
% (generated by tables/make_numbers.py from data/metadata/prisma_counts.json and linkage_report.json).
% Sources of truth: RESEARCH_PLAN.md §2–§3, protocol/PROTOCOL.md, protocol/amendments.csv,
% reports/PHASE0_FEASIBILITY.md, data/raw/README.md, ingestion/*.py, systematic_review/define_corpus.py.
% ============================================================================

\section{Information sources and conditions of access}
\label{sec:data-sources}

A field-level mapping study stands or falls with the bibliographic databases on which it is built. The choice
of sources determines which works can be observed at all, how reliably they can be dated and typed, whether
citation links are available, and under what licence derived data may be redistributed. This part documents
these choices for WP44 and reports, stage by stage, how the mapping corpus was assembled from the raw
sources. The exposition follows the order of the pipeline (identification, linkage, deduplication,
eligibility, scope, dating) and closes with the quality indicators and the research-integrity procedures that
govern every later analysis. The reporting structure is modelled on the PRISMA 2020 items that apply to a
mapping study \citep{page2021prisma,page2021prismaee}; items concerned with risk of bias and effect measures
have no counterpart in a bibliometric design and are replaced, as PRISMA-ScR does for scoping reviews
\citep{tricco2018prismascr}, by indicators of metadata quality and coverage (Section~\ref{sec:data-quality}).
The search-reporting recommendations of PRISMA-S \citep{rethlefsen2021prismas} are followed to the extent
that a database-scope strategy without keyword queries permits.

\subsection{Why a multi-source design}

No single database indexes computer science completely, and the gaps are not random. The venue-centred
coverage of DBLP, the publisher-centred coverage of Crossref, the preprint-centred coverage of arXiv and the
aggregating coverage of OpenAlex each miss a distinct part of the literature and type the part they do cover
in distinct ways. Comparative studies of bibliographic data sources have documented large differences in
coverage between Web of Science, Scopus, Dimensions, Crossref and Microsoft Academic
\citep{visser2021comparison,martinmartin2021comparison}, and the journal-level coverage of the two commercial
indexes is known to under-represent conference proceedings, which carry a large share of computer-science
output \citep{mongeon2016coverage}. For the open successor of Microsoft Academic, OpenAlex, the first
independent assessments report strong coverage of works with DOIs together with weaker and uneven metadata on
affiliations, document types and references \citep{alperin2024openalex,culbert2025openalex,
scheidsteger2023openalex,delgadoquiros2024completeness}. These findings motivate the design adopted here: one
source supplies the backbone (OpenAlex), a second supplies the definition of what counts as computer science
(DBLP), a third supplies the preprint layer and its dates (arXiv), and a fourth checks the coverage of a
subfield with a dense, curated venue structure (the ACL Anthology). Crossref is used for verification only.

\subsection{OpenAlex}

OpenAlex \citep{priem2022openalex} is the backbone of the corpus. It supplies, for every work, the
identifiers (OpenAlex~W identifier, DOI, arXiv identifier where present), the publication date and type, the
title and an inverted-index abstract where available, the venue (source) and its type, the authorships with
institutions and countries, the topic assignment at three levels (field, subfield, topic) with a primary
topic, the list of referenced works (the citation graph), the open-access status and licence, and the
field-weighted citation impact together with yearly citation counts. OpenAlex is released under the CC0
public-domain dedication, which permits redistribution of record-level derived data.

Two access routes exist and both were used, for different purposes. The public snapshot is a versioned bulk
release in gzipped JSON Lines, published in a public S3 bucket; the release used here is dated
2026-10-14 and comprises 1{,}898 part files for the \emph{works} entity, 675.3~GB compressed, with a manifest
listing the record count of each file. The snapshot was streamed from the primary compute node, each file's
content length checked against the manifest and a SHA-256 of the compressed bytes recorded in a SQLite ledger
before parsing; every one of the 1{,}898 files parsed without error and the sum of the per-file record counts
equals the manifest total of \LinkSlimTotal{} works. Two outputs were written during this single pass. The
first is the full JSON record of every work carrying \emph{any} topic in OpenAlex field~17 (Computer Science)
with a publication year between 2018 and 2026, stored as Zstandard-compressed JSON Lines and partitioned by
the snapshot's own date partitions (\LinkOaCsRecords{} records, about 20~GB). The second is a \emph{slim}
table of every work in the snapshot (identifier, DOI, year, type, primary topic, subfield, field and domain,
citation count and a few flags), kept in Parquet so that works cited by computer-science papers but published
in other fields can be resolved for the interdisciplinary analyses of Part~VIII (decision D2 of
\texttt{RESEARCH\_PLAN.md}). Records flagged by OpenAlex as XPAC expansion records are retained at this stage
with their flag so that the exclusion decision can be taken, and counted, later (Section~\ref{sec:data-eligibility}).
The smaller entity snapshots (topics, subfields, fields, domains, sources, institutions, publishers, funders,
concepts) were downloaded the same day with a per-file SHA-256 ledger; author entities were not downloaded
because the authorship data embedded in each work record suffice for the analyses planned.

The second route is the REST API. Without an API key the public endpoint operates under a shared daily budget
of USD~0.10, list-and-filter calls are priced at USD~0.0001 each, and cursor paging is unavailable, so that a
single filtered query cannot return more than 10{,}000 results through basic paging. Single-record lookups by
identifier are free. The API was therefore used for two narrowly defined tasks. First, an incremental harvest
of works with any computer-science topic and a publication date between 2026-10-01 and 2026-11-02, intended
to capture records added after the snapshot was cut; the harvest is sliced by publication day and, where a day
exceeds the paging limit, split into records with a computer-science primary field and records without one.
Raw JSON pages are stored verbatim under a dated directory with a ledger of counts and cost, and the job
resumes on the following day when the budget is exhausted. Second, free single-record lookups by DOI serve as
one of the verification routes for the bibliography of this manuscript (Section~\ref{sec:data-integrity}).
A free API key, which would raise the daily budget, requires the principal investigator to act on an e-mail
and was deliberately not obtained without instruction; the mandate forbids assuming credentials that have not
been tested.

\subsection{DBLP}

The DBLP computer science bibliography \citep{ley2002dblp,ley2009dblp} is a curated, venue-centred index that
has for three decades defined, in practice, the set of journals and conference series regarded as computer
science. It is used here in that capacity: a work indexed by DBLP is in computer science by construction, and
DBLP venue streams feed the taxonomy of Part~IV. The complete XML dump (\texttt{dblp.xml.gz}, dated 2026-11-02
22:39~UTC, 1.12~GB) was downloaded together with its DTD and MD5 checksum, the checksum verified, and the file
parsed with a validating XML parser into one row per publication record (articles, conference papers,
proceedings volumes, books, book chapters, doctoral and master's theses, data records; the \texttt{www}
person records are skipped). The whole archive was kept so that per-year counts are available for the
historical baseline and for context. The DBLP dump is released under the Open Data Commons Attribution licence
(ODC-BY~1.0). DBLP contributes, in addition to the scope definition, author person identifiers, venue streams,
and DOI links for most records; the \texttt{ee} fields were also scanned for arXiv URLs, with a result reported
in Section~\ref{sec:data-linkage}.

\subsection{arXiv}

The arXiv e-print repository \citep{ginsparg2011arxiv} is the main preprint channel of computer science and,
for the period under study, the place where a large share of the most cited work first appeared. Its
metadata are harvested through the OAI-PMH interface (\texttt{set=cs}, \texttt{metadataPrefix=arXiv}) in
monthly datestamp windows from January 2018 to 2 November 2026. The harvester waits at least 3.5~s between
requests and honours \texttt{Retry-After} headers; every raw XML page is kept verbatim as the immutable
snapshot and parsed afterwards into a single Parquet table. Two properties of OAI-PMH matter for what follows.
The \texttt{from}/\texttt{until} parameters filter on the record's \emph{datestamp}, the time of its last
update, so that a monthly window contains every record updated in that month, not every record first
submitted then; duplicates across windows are resolved at parse time with the latest datestamp winning. The
\texttt{created} element gives the first submission date, which is the date used throughout this study for
preprints. The harvest returned \LinkArxivCsRecords{} distinct records, of which
\LinkArxivCreatedTwoZeroOneEightTwoZeroTwoSix{} were first submitted in 2018--2026; the remainder are older
e-prints updated during the window. The arXiv metadata are released for reuse under the repository's metadata
terms; e-prints themselves are governed by per-paper licences and are downloaded, for the high-priority set
only, within those licences and at the repository's rate limits. The use of arXiv as a research dataset, and
the caveats attached to its category labels and versioning, are discussed by \citet{clement2019arxiv}.

\subsection{ACL Anthology}

The ACL Anthology is the curated archive of the computational-linguistics community and has itself served as a
bibliometric corpus \citep{radev2013acl}. Its BibTeX export with abstracts (\texttt{anthology+abstracts.bib.gz},
46.7~MB, SHA-256 recorded) was downloaded on 2026-11-02 and parsed into a table of 138{,}796 entries, of which
\LinkAclRecords{} are conference papers or journal articles and \LinkAclTwoZeroOneEightTwoZeroTwoSix{} fall in
2018--2026. The Anthology plays a diagnostic role: because its venue structure is complete and hand-maintained,
the share of its entries that the backbone can be matched to is a direct measure of OpenAlex coverage for a
subfield that publishes overwhelmingly in conference proceedings. It also supplies abstracts for those entries
in which OpenAlex has none. The export is licensed CC~BY~4.0.

\subsection{Crossref, Semantic Scholar and other sources}

Crossref \citep{hendricks2020crossref} is the registration agency behind most DOIs in the corpus and offers
open metadata through a REST API with a polite pool. It is not used to build the corpus, whose backbone
already carries Crossref-derived metadata through OpenAlex, but for three verification purposes: resolution of
every DOI cited in this manuscript, a cross-check of publisher-assigned document types on samples, and a
cross-check of citation counts (\texttt{is-referenced-by-count}) on a sample of highly cited works in
Part~IX.

The Semantic Scholar Academic Graph \citep{kinney2023s2} was listed in the mandate among the priority sources
because of two signals it alone provides, the influential-citation classification and citation intent. At the
time of the feasibility audit, anonymous requests to its API returned HTTP~429 with an instruction to apply for
a key, and the bulk-dataset endpoint requires a key outright. In keeping with the rule that inaccessible data
be recorded rather than circumvented, Semantic Scholar is recorded as unavailable (decision D3) and no analysis
in this manuscript uses influential-citation or citation-intent signals; where Part~VIII discusses the
strength of citation links, it does so with OpenAlex reference lists and the knowledge-graph extraction of
Part~IV, whose limits are stated there. The Microsoft Academic Graph \citep{wang2020mag}, from which OpenAlex
inherited much of its historical coverage, was discontinued at the end of 2021 and is not an independent source
for the period studied. Kaggle credentials for the arXiv snapshot distributed there were absent and were not
needed, OAI-PMH being sufficient. Benchmark repositories and code-hosting platforms enter only in the
reproducibility and benchmark analyses scoped in Part~IV and are not part of the bibliographic corpus.

Table~\ref{tab:data-sources} summarises the sources, the access route, the licence and the role of each.

\begin{table}[htbp]
\centering\footnotesize
\caption{Information sources of the WP44 corpus (all accessed 2026-11-02/03).}
\label{tab:data-sources}
\begin{tabularx}{\textwidth}{>{\raggedright\arraybackslash}p{2.5cm}>{\raggedright\arraybackslash}X>{\raggedright\arraybackslash}p{2.1cm}>{\raggedright\arraybackslash}X}
\toprule
Source & Access and snapshot & Licence & Role \\
\midrule
OpenAlex works snapshot & public S3 bucket, release 2026-10-14, 1{,}898 JSONL files, \LinkSlimTotal{} works & CC0 & backbone: identifiers, dates, types, topics, authorships, references, abstracts, OA status \\
OpenAlex REST API & no key; USD~0.10 per day shared budget; increment 2026-10-01 to 2026-11-02; single lookups & CC0 & post-snapshot increment; reference verification \\
DBLP & XML dump 2026-11-02 22:39~UTC, MD5 verified & ODC-BY~1.0 & computer-science scope; venue streams; author identifiers \\
arXiv & OAI-PMH set \texttt{cs}, 2018-01 to 2026-11, raw pages kept & arXiv metadata terms & preprint layer; first-submission dates; categories \\
ACL Anthology & BibTeX export with abstracts, 2026-11-02 & CC~BY~4.0 & coverage check for NLP venues; abstracts \\
Crossref & REST API, polite pool & open metadata & verification of DOIs, types and citation counts \\
Semantic Scholar & API key required; anonymous calls refused & --- & \emph{unavailable, recorded} \\
\bottomrule
\end{tabularx}
\end{table}

\section{Identification: what each source contributes}
\label{sec:data-identification}

There is no keyword search in this study. A mapping of an entire discipline cannot be defined by a query
string without importing the vocabulary of whoever wrote it; it is defined instead by database scope filters
that are themselves documented and reproducible, which is the strategy recommended for mapping studies in
software engineering and adopted here for computer science as a whole
\citep{petersen2008mapping,petersen2015mapping}. The identification stage (S0 in the notation of the research
plan) therefore consists of four scope filters, one per source, applied to immutable snapshots whose dates and
hashes are recorded in \texttt{data/raw/README.md}.

\subsection{Scope filters}

The OpenAlex filter retains every work with at least one topic, primary or secondary, in field~17 and a
publication year in $[2018, 2026]$. The choice of \emph{any} topic rather than the primary topic is deliberate
and conservative at this stage: a work whose primary topic lies in, say, statistics or electrical engineering
but which carries a secondary computer-science topic is a candidate, and the decision whether it belongs to
the core or to the penumbra is deferred to the eligibility stage, where it is made explicit and counted
(Section~\ref{sec:data-core}). The filter yields
\PrismaIdentifiedOpenalexCsTopicTwoZeroOneEightTwoZeroTwoSixPassOne{} records in the first pass over the
snapshot. The DBLP filter retains the publication records of type article, inproceedings, incollection, book,
phdthesis and data with a year in $[2018, 2026]$: \PrismaIdentifiedDblpTwoZeroOneEightTwoZeroTwoSix{} records.
The arXiv filter retains the records of set \texttt{cs} whose first submission falls in the window:
\PrismaIdentifiedArxivCsTwoZeroOneEightTwoZeroTwoSix{} records. The ACL filter retains the conference papers
and journal articles of the Anthology with a year in the window:
\PrismaIdentifiedAclTwoZeroOneEightTwoZeroTwoSix{} records. The four counts overlap heavily, since most DBLP
and ACL records and a majority of arXiv records also exist in OpenAlex; the overlap is resolved by record
linkage, not by summation, and no total of identified records is reported as if the sources were disjoint.

\subsection{The OpenAlex topic classifier as a scope instrument}

Because the OpenAlex filter relies on an automated topic classifier, its properties bear on everything
downstream. OpenAlex assigns to each work up to three topics from a fixed list of 4{,}516, organised in 252
subfields, 26 fields and four domains; the assignment is made by a model applied to the title, abstract and
venue, and the field label of a work is inherited from its topics rather than assigned directly. Two
consequences follow. A work can be labelled computer science because one of its topics belongs to field~17
even when its content is only marginally computational, and conversely a work in a computer-science venue can
receive no computer-science topic when its abstract is missing or its vocabulary is domain-specific. The first
error inflates the candidate set with what will be called the penumbra; the second removes genuine
computer-science works from the first-pass candidate set. DBLP is used to correct the second error, which is
why the linkage stage has two passes (Section~\ref{sec:data-linkage}). The literature on OpenAlex has so far
examined its coverage and metadata completeness rather than the accuracy of its topic labels
\citep{alperin2024openalex,delgadoquiros2024completeness}; the present study treats the labels as a scope
instrument whose errors are bounded by the DBLP and arXiv memberships and quantified in
Section~\ref{sec:data-quality}, not as ground truth.

\subsection{Verification of the scale against the mandate}

The mandate set planning targets of more than one million candidate records and more than one hundred
thousand deeply classified records, subject to availability. The feasibility audit established, before any
ingestion, that the OpenAlex API alone reported more than six million works with a computer-science primary
field in 2018--2026 and that the DBLP dump held more than four million records in the same years; the targets
were therefore attainable without recourse to any restricted source. The audit also established what was
\emph{not} attainable: a Semantic Scholar key, an OpenAlex key, Kaggle credentials and the AWS command-line
tool were all absent, and the ingestion was designed around their absence (HTTPS streaming of the public
bucket, basic paging of the public API). The record of this audit is \texttt{reports/PHASE0\_FEASIBILITY.md}.

\section{Record linkage across sources}
\label{sec:data-linkage}

The four candidate sets describe overlapping populations with different identifiers. Record linkage (stage S1)
establishes, for each DBLP, arXiv and ACL record, the OpenAlex work it describes, so that one node per work
can be formed with the union of the metadata. The methods are deterministic and ordered by reliability; the
probabilistic framework of \citet{fellegi1969linkage} is not needed because the available keys (DOI, arXiv
identifier) are close to exact, and the residual title-based step is restricted to blocks where it is
unambiguous. The implementation follows the standard decomposition of data matching into key normalisation,
blocking, comparison and classification \citep{christen2012matching}; Part~IV gives the formal statement of the
keys and of the precision-estimation design, and this section reports what was done and what it produced.

\subsection{Keys and matching order}

All keys are normalised before comparison. DOIs are lower-cased. arXiv identifiers are lower-cased and
stripped of their version suffix (\texttt{v2}, \texttt{v3}, \ldots), so that all versions of an e-print map to
one identifier. Titles are reduced to a normalised string by lower-casing and deleting every character that is
not a letter or digit; the normalised title is used only when it has at least 25 characters, is unique within
its source, and is unique within OpenAlex, which is the blocking rule that keeps the title step from
producing spurious links between generic titles.

For DBLP, the order is: (i) DOI equality, (ii) arXiv identifier equality, restricted to records not yet linked
on either side, (iii) normalised title equality with a year difference of at most one, restricted to records
not yet linked on either side. The DOI step linked \LinkLinkDblpOaDoi{} records and the title step
\LinkLinkDblpOaTitleYear{}; the arXiv step linked none, because the extraction of arXiv URLs from the DBLP
\texttt{ee} fields returned no candidate (\LinkDblpWithArxiv{} records with an extracted identifier). The
\texttt{ee} fields of this dump evidently do not carry \texttt{arxiv.org/abs} URLs in the form the extractor
expects; whether the CoRR entries are reached by the DOI step instead depends on DBLP's current conventions for
the \texttt{ee} field, which the audit did not characterise, and the question is left to the linkage-precision
review. The step is kept in the pipeline because other dumps may differ. In total \LinkDblpLinkedDistinct{} of the \LinkDblpTwoZeroOneEightTwoZeroTwoSix{} DBLP
records in the window were linked to an OpenAlex work after the first pass, leaving \LinkDblpUnlinked{}
unlinked. The composition of the unlinked set is informative: \LinkDblpUnlinkedByTypeArticle{} journal
articles, \LinkDblpUnlinkedByTypeInproceedings{} conference papers, \LinkDblpUnlinkedByTypePhdthesis{}
doctoral theses, \LinkDblpUnlinkedByTypeData{} data records, \LinkDblpUnlinkedByTypeIncollection{} chapters
and \LinkDblpUnlinkedByTypeBook{} books. Of these, \LinkDblpUnlinkedWithDoiInSlim{} carry a DOI that
\emph{does} exist in the slim table of all OpenAlex works, which means the work is in OpenAlex but was not
caught by the computer-science topic filter; this is the second error of the topic classifier described above
and the reason for the second pass.

For arXiv, the order is: (i) arXiv identifier equality against the identifier OpenAlex records for the work,
(ii) DOI equality for e-prints that carry a journal DOI, (iii) normalised title equality with a year
difference of at most two, the wider tolerance reflecting the typical lag between a preprint and its published
version \citep{lin2020preprints}. The steps linked \LinkLinkArxivOaArxivId{}, \LinkLinkArxivOaDoi{} and
\LinkLinkArxivOaTitleYear{} records respectively, for \LinkArxivLinkedDistinct{} distinct e-prints linked.
For the ACL Anthology, the order is: (i) DOI equality, (ii) equality between the Anthology identifier and the
identifier embedded in an \texttt{aclanthology.org} landing-page URL among the work's OpenAlex locations,
(iii) normalised title equality with a year difference of at most one. The steps linked
\LinkLinkAclOaDoi{} records by DOI and \LinkLinkAclOaTitleYear{} by title, with the URL step contributing the
remainder, for \LinkAclLinkedDistinctTwoZeroOneEightTwoZeroTwoSix{} distinct 2018--2026 entries linked out of
\LinkAclTwoZeroOneEightTwoZeroTwoSix{}; the residual is examined in Section~\ref{sec:data-quality} as a
measure of OpenAlex coverage of conference-heavy subfields.

\subsection{The second snapshot pass}

A DBLP record whose DOI resolves to an OpenAlex work outside the topic filter is, by the scope rule, a
computer-science work that the first pass did not load. The same situation arises for arXiv e-prints whose
journal DOI exists in OpenAlex without a computer-science topic. The linkage script therefore writes the list
of OpenAlex identifiers that are needed but not loaded: \LinkNeededOaIdsSecondPassDblpDoi{} identifiers from
DBLP DOIs and \LinkNeededOaIdsSecondPassArxivDoi{} from arXiv DOIs. A second pass over the works snapshot,
restricted to this identifier list, retrieves the full records of these works; they are normalised into the
same Parquet tables with a \texttt{\_p2} suffix, the DuckDB views are rebuilt over the union of both passes,
and linkage is re-run so that the DBLP-linked set gains the works that the topic classifier had missed. The
second pass is what makes the DBLP membership rule of Section~\ref{sec:data-core} operational: without it,
``DBLP-indexed'' would silently mean ``DBLP-indexed and labelled computer science by OpenAlex'', and the DBLP
correction would be void precisely where it matters. DBLP records with no OpenAlex counterpart after both
passes (\PrismaLinkedDblpOnlyNoOpenalexRecord{} records before the second pass) are kept in a separate table
with their year and type and are reported as such; they enter the volume counts of Part~IX only in the
sensitivity analysis that adds DBLP-only records, because they lack the abstracts, references and affiliations
on which every other indicator depends. Most of them are journal articles and conference papers with a DOI
that OpenAlex does not hold at all, together with doctoral theses, which DBLP indexes from university
repositories that OpenAlex covers unevenly.

\subsection{Precision of the linkage}

Exact-key links are correct whenever the key is correct in both sources; the known failure modes are DOIs
reassigned by publishers between versions, DOIs registered for a proceedings volume rather than a paper, and
e-prints whose OpenAlex record points to a different version. Title-based links can fail on near-identical
titles of distinct works (extended abstracts and full papers, workshop and conference versions). The protocol
pre-specifies the estimation of linkage precision on a sample of 500 DBLP--OpenAlex pairs stratified by method,
with automatic agreement checks (normalised-title similarity, year difference, first-author surname) and a
200-pair subsample reserved for the principal investigator (checkpoint H-1). The sampling script draws 250 DOI
links, 200 title links and 50 arXiv links with a fixed seed; since the arXiv step produced no links in this
dump, the realised sample has 450 pairs, and the precision estimate for the arXiv step is undefined rather
than reported as perfect. At the time of writing the sampling script exists but the sample has not been drawn and the
human checkpoint is unsigned; the automatic agreement rates and the human verdicts will be reported in Part~IX
and Appendix~B, and every conclusion that depends on the title-based links is labelled as pending human review
until H-1 is signed.

\section{Deduplication of preprint and published versions}
\label{sec:data-dedup}

OpenAlex creates distinct work records for an arXiv e-print and for the journal or conference version of the
same paper whenever the two carry different DOIs, which is the normal case since arXiv registers DataCite DOIs
for its e-prints. In a corpus where preprinting is widespread this is the dominant form of duplication, and it
biases every count in the direction of fields that preprint most: a paper in machine learning that appears on
arXiv and then at a conference would be counted twice, a paper in a theory journal once. OpenAlex itself links
some preprint--publication pairs, but the links are incomplete, and prior work on discovering such links has
shown that neither DOIs nor titles alone suffice \citep{cabanac2021preprint}. Stage S2 therefore merges
preprint and published versions into one node per work, in the sense required by the mandate (one node per
\emph{work}, with both dates retained).

\subsection{Pairing rule}

The rule is conservative by design, favouring a small number of certain merges over a larger number of
probable ones. A \emph{preprint-like} record is one whose OpenAlex type is \texttt{preprint} or whose DOI is
an arXiv DataCite DOI (prefix \texttt{10.48550/arxiv}); a \emph{published-like} record is any other record.
Two records are paired when their normalised titles are identical, both titles have at least 25 characters,
and their publication years differ by at most two. When a published record matches several preprints, the
earliest preprint is kept as its preprint; when a preprint matches several published records, the earliest
published record is kept; all other pairs are discarded, so that the pairing is one-to-one. The pairing is
confined to the loaded OpenAlex records; preprints that are not computer-science candidates cannot be paired,
which is immaterial since they would not enter the corpus. No fuzzy title similarity is used at this stage, so
that a preprint whose title changed at publication is not merged; the consequence, a residual double count, is
discussed below.

\subsection{What is kept from each member of a pair}

The published record is the surviving node. The preprint record receives the flag \texttt{duplicate\_of} with
the identifier of the published work and is excluded with reason code E5 (Section~\ref{sec:data-eligibility}).
The published node receives \texttt{preprint\_work\_id} and, through the arXiv linkage, the first-submission
date of the e-print, so that for every work the \emph{first date} is defined as the earliest of its OpenAlex
publication date, the publication date of its preprint record and the arXiv \texttt{created} date. This first
date is the date variable of all monthly series in Part~IX; the publication date of the version of record is
kept alongside it. The distinction matters for structural-change analysis: dating a paper by its journal issue
would displace much of the research activity of 2023 into 2024 and 2025 and would blur any monthly series
around the November 2022 boundary, whereas the first date places the work at the moment it entered the public
record. Citation counts are those of the surviving published node; OpenAlex attributes to the preprint record
the citations that cite the preprint DOI, and these are not added to the published node, a known
under-count for works whose preprint remained the cited version. Part~IV states the resulting measurement
assumption and Part~IX reports, for the high-priority set, the share of citations that accrue to preprint
records.

\subsection{Outcome and residual duplication}

The pairing identified \PrismaExcludedEFiveDuplicate{} preprint records as duplicates of a published work,
of which \PrismaExcludedSequentialEFiveDuplicate{} remained after the earlier exclusions had been applied.
Set against the \PrismaEligibleByTypePreprint{} preprints that remain in the eligible corpus as works with no
published counterpart, the figure shows that most e-prints in the candidate set were not matched to a
published version within OpenAlex. Three explanations coexist and cannot be separated without manual review:
many e-prints are genuinely unpublished, consistent with the publication rates reported for computer-science
preprints \citep{lin2020preprints}; some published versions carry a modified title; and some published
versions are not in OpenAlex or lack a computer-science topic. The first explanation is benign; the second and
third leave a residual double count whose size is bounded in Part~IX by re-running the pairing with a title
similarity threshold rather than exact equality and reporting the difference as a sensitivity, not as a
correction. Because arXiv-linked works carry their e-print identifier, the sensitivity analyses can also
remove every preprint node whose e-print has a journal DOI in the arXiv metadata, whether or not the pairing
found the published record.

\section{Eligibility criteria and exclusions}
\label{sec:data-eligibility}

Stage S3 applies the eligibility criteria of the preregistered protocol (\texttt{protocol/PROTOCOL.md}, P2) to
every loaded record and records, for each, a set of Boolean exclusion flags with reason codes. The flags are
computed independently so that each code can be counted on its own and also sequentially, in the order
E1--E7, so that the PRISMA flow (Figure~\ref{fig:prisma}) sums correctly. Reporting both views is deliberate:
the independent counts show how large each problem is in the raw data, the sequential counts show how many
records each criterion removed given those before it.

\subsection{Reason codes}

\textbf{E1, outside window.} The publication date is missing, earlier than 2018-01-01 or later than
2026-11-02. Because the first-pass filter already restricted the publication year, this code catches only
records with a missing date or a date beyond the extraction date. Forward-dated records, that is works whose
publication date lies in the future relative to the snapshot, are a documented feature of OpenAlex, where
publishers register articles for forthcoming issues; the window end removes those dated after the extraction
date, and the remaining 2026 records are discussed in Section~\ref{sec:data-dates}.
\PrismaExcludedEOneOutsideWindow{} records are flagged.

\textbf{E2, non-research document type.} The protocol lists the document types that constitute research
output: article, conference paper, preprint, book chapter, book, dissertation, report, review and software
paper. The exclusion operates asymmetrically. A record that is \emph{not} DBLP-linked is excluded when its
OpenAlex type is outside the list. A record that \emph{is} DBLP-linked is kept whatever its OpenAlex type,
unless that type belongs to a hard-exclusion list of types that are never research outputs (editorial,
erratum, retraction notice, peer review, paratext, conference abstract, reference entry, dataset, software,
supplementary material, library guide, letter, grant, standard). The asymmetry reflects amendment A-001 and
is explained below. Independently of order, \PrismaExcludedETwoType{} records carry this flag. The largest
contributors are records typed as datasets (\PrismaExcludedETwoByTypeDataset{}), followed by the
residual type \texttt{other} (\PrismaExcludedETwoByTypeOther{}), library guides
(\PrismaExcludedETwoByTypeLibguides{}), software (\PrismaExcludedETwoByTypeSoftware{}), conference
abstracts (\PrismaExcludedETwoByTypeConferenceAbstract{}) and peer reviews
(\PrismaExcludedETwoByTypePeerReview{}); paratext, reference entries, editorials, supplementary materials,
errata, standards, book reviews, retraction notices, data papers and letters make up the remainder. The
prominence of datasets and software among records carrying a computer-science topic is itself a finding about
the backbone: OpenAlex indexes the DataCite registry, where repositories mint DOIs for research objects that
are not publications, and the topic classifier labels them by their descriptions.

\textbf{E3, XPAC expansion records.} OpenAlex marks as XPAC the records it added from non-Crossref sources in
its coverage expansion; these carry sparser metadata and a higher share of non-research objects. The
protocol excludes them, and \PrismaExcludedEThreeXpac{} records carry the flag independently. Amendment A-001
introduced one exception: an XPAC record that is linked to a DBLP record is a computer-science publication
whatever its provenance in OpenAlex, and it is kept with the flag \texttt{xpac\_kept\_dblp}
(\PrismaExcludedXpacKeptDblp{} records). Sequentially, after the type criterion,
\PrismaExcludedSequentialEThreeXpac{} records are removed under E3, far fewer than the independent count
because most XPAC records are also non-research types.

\textbf{E4, no title.} Records with a missing or empty title cannot be linked, deduplicated or embedded and are
excluded: \PrismaExcludedEFourNoTitle{} independently, \PrismaExcludedSequentialEFourNoTitle{} sequentially.

\textbf{E5, duplicate version.} The preprint member of a merged pair (Section~\ref{sec:data-dedup}):
\PrismaExcludedEFiveDuplicate{} independently, \PrismaExcludedSequentialEFiveDuplicate{} sequentially.

\textbf{E6, retracted.} Works flagged \texttt{is\_retracted} by OpenAlex, which draws on the Retraction Watch
data now distributed through Crossref, are \emph{not} removed from the mapping corpus: they are part of the
published record and of its volume, and their presence is itself an object of study given the record numbers
of retractions reported for 2023 \citep{vannoorden2023retractions}. They are kept with the flag and excluded
from every impact analysis (citation rankings, landmark selection, field-normalised scores).
\PrismaExcludedESixRetracted{} loaded records carry the flag, of which
\PrismaEligibleMembershipRetractedFlagged{} are in the eligible corpus.

\textbf{E7, generic deposit platform.} Introduced by amendment A-002, this code excludes records whose primary
source is a generic deposit platform (Zenodo, Figshare, the Japanese Institutional Repositories DataBase,
LA~Referencia, Dialnet, the DOAJ-as-source entries and the OSTI OAI feed) unless the record is DBLP- or
arXiv-linked. The number of records removed under this code is reported by the macro
\pending{PrismaExcludedSequentialESevenGenericDeposit} once the corpus has been rebuilt under A-002.

One inconsistency of numbering must be recorded. In the frozen protocol, code E7 designates the
penumbra of works with a secondary computer-science topic only; the implementation reuses the number E7 for
the generic-deposit exclusion of amendment A-002 and carries the penumbra as a separate flag,
\texttt{cs\_penumbra}. The penumbra is not an exclusion in the same sense as E1--E7: it is a scope boundary,
reported separately and used in sensitivity analyses, and is treated in Section~\ref{sec:data-core}. The
conflict of numbering is flagged here for correction in the amendments log before the protocol is cited in
its final form.

\subsection{Amendments A-001 and A-002 and their evidence}

Two eligibility decisions were taken after the protocol freeze, on the evidence of the first run of the
corpus-definition script, and both are logged in \texttt{protocol/amendments.csv} with their date, their
reason and their status (executing agent, pending confirmation by the principal investigator at
checkpoint H-0). They are reported here in full because an eligibility rule changed after preregistration is
exactly the kind of decision a reader must be able to audit \citep{nosek2018preregistration}.

\emph{A-001, document types.} The protocol's type list was written from the OpenAlex documentation available
at the time; the 2026 snapshot uses the types \texttt{conference-paper} and \texttt{software-paper}, which the
list did not contain. The first run of the eligibility stage excluded under code E2 a very large number of records typed
\texttt{conference-paper}, close to half of them DBLP-linked; the figures of that superseded run are recorded in
the amendment entry and are not repeated here, since the run was not retained. Conference papers are the principal publication form of several areas of
computer science, and their exclusion would have invalidated the composition analyses of Part~IX. The
amendment adds the two types to the list and, more consequentially, makes DBLP linkage decisive: a
DBLP-linked record is a computer-science research output whatever OpenAlex calls it, unless its type is one
that can never be a research output. The same amendment keeps DBLP-linked XPAC records. The corrected run
yields the counts reported above, in which \PrismaEligibleByTypeConferencePaper{} eligible works are
conference papers.

\emph{A-002, generic deposit platforms and date precision.} Inspection of the eligible corpus by source
(\texttt{systematic\_review/inspect\_sources.py}) showed that a single generic repository, Zenodo, contributed
several times more eligible records dated 2026 than dated 2025, that almost none of them carried a reference
list, and that random samples were slide decks, datasets, software releases and project reports typed as
articles because the depositor had chosen that type; the counts are recorded in the amendment entry. Left in the
corpus, such records would have produced a spurious surge of 2026 output concentrated in a repository that
does not publish research articles, and would have done so in the months where right truncation already
makes the series fragile. The amendment excludes records whose primary source is a generic deposit platform
unless DBLP or arXiv vouch for them, and introduces the \texttt{date\_precision} flag discussed in
Section~\ref{sec:data-dates}. The rule is a source-level rule, not a content judgement: a genuine research
report deposited only on Zenodo is lost, which is accepted because the alternative, keeping every deposit,
contaminates the primary estimand of RQ1. The sensitivity analyses of Part~IX report the monthly series with
and without the generic-deposit exclusion.

The amendments illustrate a general point about preregistered bibliometrics. The protocol can freeze the
\emph{rules} but not the data model of a database that changes between the time of writing and the time of
extraction; when the data model changes, the honest course is an amendment with its evidence, not a silent
adaptation.

\section{The computer-science core, the penumbra and the strict core}
\label{sec:data-core}

The decision of what counts as computer science is the single most consequential modelling choice in a
discipline-level mapping, and the protocol makes it with a three-part membership rule (decision D4) rather
than with any one source. A record that survives the exclusions E1--E7 belongs to the \emph{core} if at
least one of the following holds:

\begin{enumerate}
\item it is linked to a DBLP record (venue-based membership);
\item its OpenAlex primary field is Computer Science (classifier-based membership);
\item it carries a secondary computer-science topic \emph{and} is linked to an arXiv e-print whose primary
  category is in the \texttt{cs} archive (author-declared membership).
\end{enumerate}

A record that survives the exclusions but satisfies none of these, which means it carries a secondary
computer-science topic without venue or author support, belongs to the \emph{penumbra}. The three conditions
correspond to three independent judgements about disciplinary membership: the judgement of DBLP's editors
about venues, the judgement of OpenAlex's classifier about content, and the judgement of the authors who chose
a primary arXiv category. Requiring agreement of any one of them with the first-pass topic filter removes the
over-inclusion that a secondary topic alone would produce, while keeping works that any single source would
have missed.

\subsection{Counts}

After the sequential exclusions, \PrismaAfterExclusions{} loaded records remain. Of these,
\PrismaCsCoreEligible{} satisfy the membership rule and constitute the eligible mapping corpus, and
\PrismaCsPenumbraEligibleOtherwise{} form the penumbra. The penumbra is thus a substantial fraction of the
records that a naive ``any computer-science topic'' filter would have called computer science; its exclusion
from the primary analyses is the main difference between the present corpus and what an OpenAlex query alone
would return, and it is the reason why no count in this manuscript should be compared directly with the counts
that the OpenAlex interface displays for field~17. Within the eligible corpus, the three membership
conditions overlap: \PrismaEligibleMembershipInDblp{} works are DBLP-linked,
\PrismaEligibleMembershipCsPrimary{} have a computer-science primary field, and
\PrismaEligibleMembershipArxivLinked{} are linked to an arXiv e-print; \PrismaEligibleMembershipInAcl{} are
ACL Anthology papers. The eligible corpus divides by period into \PrismaEligibleByPeriodBaseline{} works whose
first date falls in the baseline window, \PrismaEligibleByPeriodPost{} in the post-boundary window, and a
small residue of \PrismaEligibleByPeriodPre{} whose first date, taken from a preprint or an e-print, precedes
2018 although the version of record falls in the window; the residue is kept in the corpus and assigned to
the ``pre'' period, so that it enters no monthly series. Figure~\ref{fig:prisma} renders these counts as a
PRISMA-style flow, drawn from the same JSON file that generates the macros of this section
\citep[cf.][]{haddaway2022prismaflow}.

\begin{figure}[htbp]
\centering
\includegraphics[width=0.9\textwidth]{fig_prisma_flow.pdf}
\caption{PRISMA-style flow of record identification, exclusion and the final mapping corpus. Identification
counts are per source and overlap; exclusions are sequential in the order E1--E7; the core and the penumbra
partition the records that survive the exclusions. Source: \texttt{data/metadata/prisma\_counts.json},
rendered by \texttt{figures/make\_figures.py}.}
\label{fig:prisma}
\end{figure}

\subsection{The strict core}

The membership rule is a disjunction, and a sceptical reader may object that a disjunction of three imperfect
judgements inherits the errors of each. The protocol anticipates the objection with a \emph{strict core},
defined as the conjunction of the first two conditions: works that are both DBLP-linked and labelled computer
science by the OpenAlex classifier. The strict core contains \PrismaStrictCoreDblpAndCsPrimary{} works. Every
headline result of Part~IX is re-estimated on the strict core, and the manuscript reports a result as robust
to scope only when the sign and the order of magnitude agree between the eligible corpus and the strict
core. The strict core is not a better corpus; it is a narrower one, biased toward established venues and
against preprints, dissertations and the newer conference series that DBLP has not yet adopted, and its role
is to bound the influence of scope decisions, not to replace them.

\subsection{What the design does not do}

Three limits deserve statement. First, DBLP membership is venue membership: a paper in a DBLP-indexed journal
of bioinformatics or operations research is in the core even when its content is not computational, and the
taxonomy of Part~IV, not the scope rule, is what distinguishes such works. Second, the rule has no language
component; works in any language are eligible, and the share of non-English records, which is concentrated
in dissertations and in regional journals, is reported among the quality indicators. Third, the rule is
applied to the records loaded by the two snapshot passes; a computer-science work that OpenAlex does not
index at all, or indexes without any computer-science topic and without a DBLP or arXiv DOI, is not observed.
The DBLP-only table and the unmatched ACL entries bound the size of this blind spot for two sources with
complete venue lists, and Section~\ref{sec:data-quality} reports them.

\section{Dates, date precision, indexing lag and right truncation}
\label{sec:data-dates}

Monthly series are the primary temporal unit of this study, and the November 2022 boundary falls inside a
year. Everything in the structural-change analysis of Part~IX therefore depends on the quality of dates at
the month level, which is considerably worse than at the year level in every bibliographic database. This
section documents three distinct problems, their diagnosis in the corpus, and the rule adopted for each.

\subsection{Which date}

Each eligible work carries up to three dates: the OpenAlex publication date of the surviving record, the
publication date of its merged preprint record where one exists, and the arXiv first-submission date where the
work is linked to an e-print. The \emph{first date} of a work is the earliest of the three
(Section~\ref{sec:data-dedup}), and the \emph{first month} is its calendar month. The first date measures
the entry of the work into the public record, which is the quantity relevant to questions about the pace and
composition of research activity; the publication date of the version of record measures the pace of the
publishing system and is kept for the analyses of venues and of preprint-to-publication lags. The two
coincide for works without a preprint, which are the majority outside the machine-learning areas.

\subsection{Date precision and the January artefact}

A large share of publication dates in OpenAlex are year-only dates stored as the first of January. The
phenomenon is visible without any model: the raw monthly series of first appearances shows a spike in every
January that no publishing calendar could explain (Figure~\ref{fig:volume-monthly-raw}), and inspection of the
eligible corpus by year found that between about a fifth and somewhat over a quarter of records carry a date of the form
\emph{YYYY}-01-01, with the proportion varying by source and type; dissertations, book chapters and records
from repositories are the most affected, journal articles with publisher-registered issue dates the least.
Amendment A-002 introduced the flag \texttt{date\_precision}, equal to \texttt{year} when the publication
date is the first of January and \texttt{month} otherwise. The rule is a heuristic with a known cost: a work
genuinely published on 1 January is classed as year-precision, and since the rule is applied to the
publication date rather than to the first date, a year-precision work with an arXiv first date is correctly
dated by its e-print. The counts of eligible works by precision class will be reported by the macros
\pending{PrismaEligibleDatePrecisionMonth} and \pending{PrismaEligibleDatePrecisionYear} once the corpus has
been rebuilt under A-002.

All monthly series in Part~IX use month-precision records only. Year-precision records enter the annual series,
where their date is correct, and the composition analyses that do not depend on the month. The exclusion of
year-precision records from monthly series is not innocuous: if the share of year-only dates differs between
categories or between periods, the monthly composition is biased relative to the annual one. Part~IX
therefore reports, for each category, the share of year-precision records by year, and treats a divergence
between the monthly and the annual growth estimates as a warning rather than as two independent results.

\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\textwidth]{fig_volume_monthly_raw.pdf}
\caption{Monthly first appearances in the eligible corpus before the precision filter. The January spikes
reflect year-only publication dates stored as 1 January and motivate the month-precision rule. Source:
\texttt{results/bibliometrics/volume\_monthly.csv}.}
\label{fig:volume-monthly-raw}
\end{figure}

\subsection{Indexing lag and why \texttt{created\_date} cannot measure it}

A work is observable only once the database has indexed it, and the delay between publication and indexing
determines how incomplete the most recent months are. The research plan proposed to estimate the delay
profile from the OpenAlex \texttt{created\_date} field, the date on which the record was created in OpenAlex,
and to use the profile to correct the right tail of the monthly series. The diagnostic run of this estimator
(\texttt{results/bibliometrics/indexing\_lag.csv}) refuted the proposal. For works published in 2018, the
median difference between \texttt{created\_date} and the publication date is close to eight years, and the
median falls by almost exactly one month for each later publication month, which is the signature of a
constant creation date: the field records the date on which the record was created in the current OpenAlex
data model, apparently during a re-ingestion in the second half of 2025, not the date on which the work was
first indexed. For the great majority of works published before that re-ingestion, \texttt{created\_date}
carries no information about indexing delay, and for works published after it the field measures the delay
only for the records created since. A lag profile estimated from it would be an artefact of database
maintenance. The field is therefore not used, and the plan's lag-corrected estimates are replaced by the
approach below. The episode is recorded in detail because a reader who reproduces the estimator on a later
snapshot may find the field meaningful again, and because it shows why the mandate's requirement to preserve
immutable raw snapshots is not a formality: the only way to measure indexing lag honestly is to compare two
snapshots of the same database.

\subsection{Right truncation}

The last months of the window are incomplete for three distinct reasons that must not be conflated: works
published but not yet indexed (indexing lag), works indexed but not yet linked to a DBLP record because DBLP's
own curation lags the publishers, and works whose citation counts have had little time to accumulate. The
first two shrink the counts of the recent months; the third affects only impact indicators and is treated in
Part~IV through fixed citation windows, which are the standard remedy for the age dependence of citation
counts \citep{wang2013window,waltman2016review}.

For the counts, the diagnosis uses the sources against one another. The DBLP dump of 2026-11-02 holds, for
2026, a count well below its 2025 count although ten months of the year have elapsed, which shows that DBLP's
own coverage of 2026 is incomplete; and the eligible corpus for 2026 is of the same order as that for 2025
(\PrismaEligibleByYearTwoZeroTwoSix{} against \PrismaEligibleByYearTwoZeroTwoFive{}), which on its face
suggests that OpenAlex is \emph{ahead} rather than behind. The two observations are reconciled by the
composition of the 2026 records: a much lower share with a reference list, a much lower mean number of
references, a higher share of preprints and, before amendment A-002, the Zenodo contribution described
above. In other words, the recent months are not simply truncated; they are differently composed, with the
fast channels (preprint servers, repositories) over-represented relative to the slow ones (journals,
curated indexes). A correction factor estimated from totals would be wrong for every category whose channel
mix differs from the average.

The rule adopted is conservative. The ``complete'' observation window for monthly inference ends in
June~2026; the months from July to November 2026 are reported descriptively, separately, and never enter a
breakpoint test. Within the complete window, the structural-change models include the last months with
their observed values and the robustness section of Part~IX re-estimates every break with the window ending
in December 2025, so that a result that depends on the 2026 months is identified as such. The API increment
harvested after the snapshot (Section~\ref{sec:data-sources}) adds the works registered between the snapshot
date and the extraction date, which narrows but does not close the gap. Finally, forward-dated records,
works carrying a 2026 publication date for an issue not yet released, are retained when their date precedes
the extraction date because they are, by the definition of the first date, in the public record; the
current build does not count them separately, and a count by month of the records whose publication date lies
after the snapshot date is listed among the quality indicators to be added before Part~IX is finalised.

\section{Quality indicators and known biases}
\label{sec:data-quality}

PRISMA 2020 asks for an assessment of the risk of bias in the included studies; a mapping study has no
studies in that sense, and the protocol replaces the item with indicators of the quality of the bibliographic
record itself, reported per year and per category in Appendix~B. This section defines the indicators,
reports the corpus-level values that are available at the time of writing, and sets out the known biases of
the design, each with the direction in which it is expected to push the results of Part~IX.

\subsection{Metadata completeness}

Four completeness indicators are computed for every work: presence of an abstract, presence of a non-empty
reference list, presence of at least one affiliation with an institution identifier, and presence of a
country for at least one author. In the eligible corpus, \PrismaEligibleMembershipHasAbstract{} works have
an abstract and \PrismaEligibleMembershipHasReferences{} have at least one resolved reference, out of
\PrismaCsCoreEligible{}. Both shares are lower than those reported for the DOI-bearing journal literature in
comparisons of OpenAlex with the commercial indexes \citep{culbert2025openalex,delgadoquiros2024completeness},
which is expected in a corpus that includes dissertations, book chapters and preprints, and both fall sharply
for the most recent year (Figure~\ref{fig:types-openness}). The consequences are concrete. A work without an
abstract cannot be embedded or classified by its content and is classified from its venue, its arXiv
categories and its OpenAlex topics alone, with a lower confidence score; a work without references does not
contribute out-citations to the knowledge graph, so that the citation flows of Part~VIII are estimated on the
subset with references and are weighted accordingly. The ACL Anthology abstracts fill part of the first gap
for one subfield; no source fills the second.

\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\textwidth]{fig_types_openness.pdf}
\caption{Document-type composition and metadata completeness of the eligible corpus by publication year
(2026 incomplete). Source: \texttt{results/bibliometrics/volume\_yearly.csv}.}
\label{fig:types-openness}
\end{figure}

\subsection{Coverage overlap between sources}

The overlap between sources is the study's main instrument for measuring its own coverage, because each
source has a complete list of what it holds. Three overlaps are reported. The DBLP--OpenAlex overlap, after
both passes, is the share of DBLP records in the window with an OpenAlex counterpart; its complement, the
DBLP-only table, is tabulated by year and type in Appendix~B. The arXiv--OpenAlex overlap is the share of
\texttt{cs} e-prints with an OpenAlex counterpart (\LinkArxivLinkedDistinct{} of
\LinkArxivCreatedTwoZeroOneEightTwoZeroTwoSix{} e-prints first submitted in the window). The ACL--OpenAlex
overlap is the share of Anthology papers linked (\LinkAclLinkedDistinctTwoZeroOneEightTwoZeroTwoSix{} of
\LinkAclTwoZeroOneEightTwoZeroTwoSix{}). An unmatched ACL paper is a documented computer-science publication
that the backbone either does not hold or holds under a title or identifier the linkage could not resolve;
the residual is therefore an upper bound on the sum of backbone non-coverage and linkage failure for one
conference-heavy subfield, and a manual review of a sample of unmatched entries, planned with checkpoint
H-1, separates the two.

Figure~\ref{fig:volume-yearly} shows the eligible works by publication year and by source membership. The
widening of the arXiv-linked layer and the relative stability of the DBLP-linked layer are the first
descriptive facts of the study, and they are also a warning about right truncation: in the most recent
years, the fastest source dominates.

\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\textwidth]{fig_volume_yearly.pdf}
\caption{Eligible works per publication year by source membership (DBLP-linked, OpenAlex computer-science
primary field, arXiv-linked, ACL Anthology); memberships overlap. 2026 is incomplete. Source:
\texttt{results/bibliometrics/volume\_yearly.csv}.}
\label{fig:volume-yearly}
\end{figure}

\subsection{Known biases and their expected direction}

\emph{DBLP coverage bias.} DBLP indexes established venues and adopts new conference series with a delay;
within the core, the DBLP-linked subset over-represents long-standing journals and conferences and
under-represents workshops, regional venues and the newest series. Because the strict core is defined
through DBLP, every strict-core estimate inherits this bias, which runs against emerging topics and against
the most recent years. Results that are stronger in the strict core than in the eligible corpus are therefore
unlikely to be artefacts of over-inclusion; results that appear only in the eligible corpus may be.

\emph{OpenAlex topic noise.} The topic classifier mislabels in both directions, and the membership rule
bounds only the over-inclusion, through the penumbra. Under-inclusion is bounded by DBLP and arXiv for the
venues and preprints they cover and is unbounded elsewhere. Within the core, the topic labels are used as
one of several signals in the taxonomy, never alone, and the taxonomy's validation on human-reviewed samples
(Part~IV, checkpoint H-2) is the only direct measure of how much label noise survives.

\emph{XPAC records.} The expansion records excluded under E3 are not a random sample of the literature: they
are concentrated in repositories, in non-English regions and in recent years. Their exclusion removes some
genuine research, mostly outside the venue system, and biases the corpus toward the formal literature; the
DBLP exception restores the part that the venue system vouches for. The sensitivity analysis that reinstates
XPAC records of research types reports the difference.

\emph{Generic-deposit platforms.} The exclusion under E7 removes, besides the deposits that motivated it, any
research output whose only home is a generic repository; the effect is largest in 2025--2026 and in the
areas, such as scientific computing and computational biology, where data and software deposits accompany
papers. The direction of the bias is toward under-counting recent output in those areas, which is the
opposite of the bias the rule corrects.

\emph{Missing abstracts.} Abstract availability depends on publishers' metadata deposits and on OpenAlex's
sources; it is lower for conference papers of some societies and for older records, and it is strongly
correlated with the DOI registrant. Every content-based analysis is conditional on abstract availability, and
the appendix reports the abstract share by category and year so that the reader can judge where the
conditional population departs from the corpus.

\emph{Forward-dated and year-dated records.} Both displace works along the time axis in a direction that is
not symmetric: year-only dates pull works toward January, forward dates push them toward the future. The
month-precision rule removes the first; the second is bounded by the window end and will be counted
explicitly.

\emph{Retractions.} Retracted works are in the corpus and out of the impact analyses. The retraction flag
depends on the coverage of the Retraction Watch data through Crossref, which is incomplete for publishers
that do not deposit retraction notices; some retracted works are therefore in the impact analyses
unflagged. The direction of the bias is small but known: retraction is concentrated in a few venues and in
a few years, and the landmark selection of Part~V checks each candidate against the current retraction
status at the time of review.

\emph{Preprint duplication.} The residual double count after the exact-title pairing inflates the areas that
preprint most, which are also the areas expected to show the strongest growth. The sensitivity with a
similarity threshold and the removal of preprints with a journal DOI bound the effect; a growth result for
machine-learning areas that does not survive these two variants is reported as not robust.

None of these biases is corrected by reweighting. The study reports them, bounds them where a second source
permits, and re-estimates its results on alternative corpora (strict core; corpus with penumbra; corpus with
XPAC research types; corpus with DBLP-only records; corpus with generic deposits) so that the reader can see
how far each conclusion depends on each scope decision. This is the approach that the science-of-science
literature recommends for databases whose coverage is known to be uneven \citep{fortunato2018science}, and it
is the only one compatible with a mandate that forbids the fabrication of corrections.

\section{Research-integrity procedures}
\label{sec:data-integrity}

The mandate for this working paper lists integrity rules that are unusual for a bibliometric study: no
fabricated reference, identifier, count or result; no paper content invented from metadata; no abstract
treated as a full text; no model-generated summary treated as evidence; no temporal association read as
causation; no preprint described as peer-reviewed; inaccessible data recorded explicitly; and human review of
every scientific conclusion. The rules are a response to the way the study is conducted, by a principal
investigator directing autonomous software agents, and this section describes the procedures by which they
are enforced. The procedures are reported here, in the data part, because most of them operate on data and
artefacts rather than on prose.

\subsection{Preregistration and the amendments log}

The protocol (\texttt{protocol/PROTOCOL.md}) was frozen on 2026-11-02 before any analysis of the corpus. It
fixes the research questions, eight directional hypotheses H1--H8 with the sign of the expected effect
stated so that null and opposite results must be reported as such, the eligibility criteria and their codes,
the sources and extraction dates, the data items, the quality indicators, the synthesis methods, the
human-review checkpoints and the data-availability commitments. The preregistration is internal, in the
repository rather than in a registry, which limits its evidential value relative to a time-stamped public
record \citep{nosek2018preregistration}; the repository's version history and the dated amendments log are
the available substitute, and the release of Phase~9 will include both. Every departure from the frozen
protocol is an entry in \texttt{protocol/amendments.csv} with an identifier, a date, the phase, the change,
the reason and the approver, which until the principal investigator signs is ``executing agent (pending
PI)''. At the time of writing the log holds five entries: A-000 records that the mandate's acceptance
criteria were completed by the executing agent after the pasted text was truncated; A-001 and A-002 are the
eligibility amendments of Section~\ref{sec:data-eligibility}; A-003 and A-004 fix the budgets and the models
of the LLM-based taxonomy validation and knowledge-graph extraction. Analyses added after the freeze are
labelled exploratory in Part~IX.

\subsection{Immutable raw snapshots and provenance}

Nothing in the corpus can be traced to a result unless the raw data from which it was built can be
re-obtained or re-verified. The OpenAlex release is a public, versioned object; the study keeps its manifest,
the per-file content lengths and SHA-256 hashes, and the filtered subset, rather than a copy of the full
bucket (decision D6), so that a reader with the same release and the same scripts obtains byte-identical
inputs, and a reader with a later release can measure the difference. The DBLP dump is kept with its MD5,
the arXiv harvest as the raw OAI-PMH pages, the ACL export with its SHA-256, and the API increment as the raw
JSON pages with a ledger of calls and cost. Each ingestion job writes a ledger with attempts, timings, hashes,
counts and errors, which is the job-level provenance the mandate requires. The processing chain from raw
snapshot to PRISMA counts is a sequence of scripts with measured wall times in
\texttt{reproducibility/RUNBOOK.md}, executed in a locked Python environment whose package versions are
recorded; the chain was executed on a single node, deliberately, because it is network-bound and because a
single-node execution is easier to reproduce than a distributed one.

\subsection{Numbers that cannot be typed}

Every count, share and estimate quoted in this manuscript is a LaTeX macro generated by
\texttt{tables/make\_numbers.py} from the result files (\texttt{prisma\_counts.json},
\texttt{linkage\_report.json}, the bibliometric and econometric result tables). The macros cannot be edited
by hand, a macro whose source file is absent is undefined and stops the compilation, and a number the authors
need but the pipeline has not produced appears in the text as a visible pending marker rather than as a
figure. The procedure is the mechanical form of the mandate's requirement that every quantitative claim
trace to a script, an input artefact and a result file. It has a cost in prose, since the text cannot
anticipate a number, and a benefit in auditability, since a reader who regenerates the result files
regenerates the manuscript's numbers.

\subsection{Verified references}

The bibliography of this manuscript is built under a rule that no reference may be cited unless it has been
resolved to a bibliographic record by a single programmatic lookup against Crossref, OpenAlex or arXiv, with
the returned title compared with the intended one and the comparison recorded in a verification table
(\texttt{references\_verification*.csv}: key, identifier, service, returned title, match). References that
fail the lookup are removed from the bibliography and kept in the table with a failed status, and a second
agent re-verifies a sample in the audit phase. The rule exists because the model-based agents that draft the
prose are known to produce plausible but non-existent references when unconstrained; it is also why
preprints are cited as preprints, with their arXiv identifier and without a venue, and why the bibliography
distinguishes a technical report from a peer-reviewed article in its entry type.

\subsection{Agents, artefacts and the register}

The study is executed by a research-director agent that spawns specialised agents with written briefs;
each brief, the artefacts it produced and the agent that verified them are recorded in
\texttt{agents/REGISTER.md}. The rule that governs the register is that no agent approves its own findings:
an artefact enters the manuscript only after a different agent, the scientific critic or the reproducibility
auditor, has checked it, and the register records the checker. Agents exchange structured artefacts (tables,
JSON, Markdown with declared schema) rather than conversational summaries, so that what one agent passes to
another can be inspected after the fact. The writing agents, including the one that produced this part, work
from the source-of-truth files named in their briefs and are instructed to describe what the code does
rather than what the plan intended; where the two differ, as with the indexing-lag estimator of
Section~\ref{sec:data-dates}, the difference is reported.

\subsection{Human-review checkpoints}

Six checkpoints are defined (H-0 to H-5: plan and protocol; linkage sample; taxonomy validation sample;
landmark list and specialist reviews; structural-change results and their interpretation; manuscript
conclusions), each with a named artefact that the principal investigator reviews and signs
(\texttt{reports/HUMAN\_REVIEW\_CHECKPOINTS.md}). Until a checkpoint is signed, every conclusion that
depends on it is labelled ``pending human review'' in the text, with the same visible marker as a missing
number. At the time of writing no checkpoint is signed. The reader should take this literally: the
conclusions of this manuscript are, in its present version, the conclusions of a procedure, reviewed by its
own audit agents and awaiting the judgement of its human author.

\subsection{What the procedures do not guarantee}

The procedures guarantee traceability, not truth. A count traced to a script is still wrong if the script is
wrong; a verified reference is still miscited if the prose misreports what it says; a preregistered
hypothesis can still be tested with an inappropriate model. What the procedures make possible is detection:
every number, reference and decision has a place where an auditor can look. The remaining risks are the
ones this part has tried to state plainly: the coverage and labelling errors of the sources, the residual
duplication, the fragility of the recent months, and the dependence of every scope decision on the
judgement, recorded but not infallible, of the agents that took it.