Skip to content

Comparative methodology · 24 selected cases and 2 sentinels

When a funder count changes with the metadata: a DOAB case study

In selected DOAB records, semicolon splitting and raw UUID-value counting do not consistently reproduce embedded source-entity structure. A typed procedure preserves unmatched funding assertions and explains why an identifier inventory is not an organization count.

Published
Share
Share this page
EmailCite & reuse this study
Key findings as text
Reusable imagesWhen a funder count changes with the metadata: a DOAB case study — 24 + 2: Hash-selected development cases and separate sentinels. Six records per CSV state; two purposive structural examples. No population prevalence.Download sharing image (JPG) ↓

Key findings

24 + 2

Hash-selected development cases and separate sentinels

Six records per CSV state; two purposive structural examples. No population prevalence.

1 of 6

Token-count disagreement in the semicolon-name sample

One entire semicolon-bearing label belongs to one embedded source entity; not proof of one real organization.

6 of 6

Grant-only cases with textual funding assertions

CSV names are blank; xoai has text and authorities but no same-record embedded funder entity.

12

Constructed adversarial fixtures tested

Synthetic software tests, not additional observed records or an external funding standard.

The question hidden inside a funder count

A catalogue field can contain a funder's name without providing a resolved organization. An identifier can appear in a relationship field without matching an embedded entity. Conversely, a record can contain a named funding assertion even when the convenient CSV column for funder names is blank. These distinctions matter whenever a researcher converts book metadata into a table of institutions, grants or funding relationships. Before asking which organization appears most often, the researcher must determine what each counted item actually represents. A delimiter, an identifier-shaped string and a source entity are different analytical units.

This comparative methodological case study examines those distinctions in a small, fixed set of Directory of Open Access Books records. It compares a retained DOAB CSV with the corresponding fuller xoai metadata for 24 records selected across four CSV field states and two deliberately selected structural examples. It then evaluates three simple counting recipes and a typed decision-and-abstention procedure. The purpose is to identify reproducible operations that these particular representations support, rather than estimate the prevalence of funded books or rank organizations. The broader question of organizational funding concentration remains unanswered.

The central result is practical: a defensible inventory of embedded source identifiers is possible, but it is not interchangeable with a funder census. Splitting a semicolon-containing name can divide one represented entity into several tokens. Counting every UUID-shaped value can produce a larger number than the embedded entity inventory. Restricting attention to successfully joined entities can also hide explicit textual funding assertions. A useful procedure therefore needs more than a single total. It should return the supported inventory, preserve unmatched evidence and explain which proposed attribution operations remain unjustified.

Existing work and the narrower contribution

The limitations of bibliographic funding data are not a new discovery. Grassano and colleagues examined coverage, uses and limitations of funding acknowledgements, including the need to harmonize names and distinguish organizational levels. Their comparison used reference acknowledgements, which provide a kind of validation that this metadata-only study does not possess. We cannot turn an apparent discrepancy between two catalogue representations into a measure of actual-funding accuracy. The relevant reference here is an observable structure inside the retained metadata, not a verified list of the organizations that financed a book.

Work by Kramer and de Jonge distinguishes the availability of names and identifiers across open funding metadata. Kramer's subsequent analysis considers direct and project-mediated funder links and notes that its source databases were captured at different dates. The UKRI and MoreBrains metadata report also distinguishes schema capability from populated fields and includes long-form outputs. These contributions already establish the importance of linkage architecture, incomplete identifiers and book-specific metadata coverage. The Crossref funding-information specification likewise distinguishes funder identity from award numbers and grant identifiers, and requires appropriate structural association. None of those principles is presented here as a new general method.

The narrower contribution is an executable account of what happens when particular counting and matching operations encounter heterogeneous DOAB value, authority and embedded-entity structures. Tsuji's earlier DOAB statistics addressed subjects, languages, publishers, publication years and licences; the operation studied here is instead a same-record linkage decision. The reviewed literature does not establish the exact meaning of every UUID-shaped DOAB value encountered in these files. Rather than invent that meaning, the procedure retains unresolved values as an explicit output. This application-specific contribution should be judged by its concrete decision consequences, not by a claim to have invented conservative metadata practice.

A fixed source frame, not a population estimate

The retained CSV contains 114,618 catalogue rows, including 103,963 book records and 10,655 chapter records. The book records have unique UUIDs and handles, and no nonblank parent-book field under the eligibility check used here. The frame is a set of represented catalogue records, not a deduplicated census of intellectual works. Separate editions or manifestations may have separate records. The original CSV was retrieved on 30 October 2026, but its generation timestamp was not verified. Retrieval time is consequently not treated as the date on which every field was observed or updated.

The selection used four mutually exclusive states of two CSV fields: a nonblank funder-name cell without a semicolon, a nonblank name cell containing a semicolon, a blank name with nonblank grant text, and both fields blank. Their frame sizes were 2,461, 126, 4,280 and 97,096 respectively. Six records from each state were selected by a deterministic hash ordering of the book UUID with a fixed seed. These equal-sized groups were chosen to inspect different representations, not to reproduce the catalogue's distribution. Their results are reported separately, without a pooled prevalence estimate or inferential uncertainty interval.

Two additional cases were selected to inspect potentially difficult labels: a Max Planck Institute and Society combination, and a European Commission, European Research Council and University of Passau combination. They are purposive sentinels, not additional hash-selected observations. The selection manifest was frozen before the 26 retrievals. Each selected record returned a successful response and matched its CSV book UUID and handle; no unavailable record was replaced. The previously inspected positive record that motivated this pilot was not among these 26 cases and is not added to their denominators.

What the retained structures can tell us

The CSV name field is a convenient projection, whereas xoai exposes several distinct kinds of material. A relationship may have a value, an authority identifier and other associated fields. Embedded linked-item metadata may separately identify an entity and supply its name, aliases or asserted DOI. These components can be compared through exact strings, but their roles should not be collapsed. In this study, UUID-shaped means only that a string has the expected syntactic form. It does not establish that the string denotes a funder, an award or any other particular object.

The reference count for the comparison is the number of distinct valid UUID strings in embedded funder-entity blocks within one record. Each block must contain exactly one UUID-shaped identifier for this count to be defined. The number of embedded blocks is retained separately, so duplicate blocks cannot silently inflate a distinct-ID inventory. An empty block list yields zero embedded identifiers, not zero organizations or zero funding. A block with an invalid or multiple identifier makes the distinct-ID reference undefined and triggers abstention from the count comparison.

Other assertions remain available even when they cannot join to that inventory. Text-valued relationship fields are preserved across every language branch. Authority identifiers and unmatched UUID-shaped values are recorded separately. Entity DOI values remain source assertions rather than externally resolved identities, while aliases remain aliases rather than automatic instructions to merge organizations. The retained files do not establish a parent hierarchy, financial amounts or a reliable mapping between each book-level grant string and a particular funder. Those operations are therefore outside the evidence contract of the procedure.

Three counting recipes and an abstention procedure

The first baseline splits the CSV name cell on literal semicolons, trims the pieces, removes empty pieces and counts the remaining token occurrences. It preserves repeated tokens. This is a deliberately simple recipe, not an assertion that a particular earlier publication used it. The second baseline counts UUID-shaped occurrences in relationship value fields across all language branches, including repeated values but excluding authority fields. The third takes the exact union of authority strings and UUID-shaped values, intersects that set with the same record's valid embedded entity identifiers, and counts the distinct matches.

Each baseline is compared with the embedded-ID reference through a signed difference: its count minus the reference count. The finite number of nonzero differences is reported alongside the relevant case denominator. These are count disagreements relative to a source representation, not errors against true organization identities. The comparison can reveal that a recipe does not reproduce the chosen observable, but cannot establish that one recipe correctly measures actual funding. A zero difference is particularly uninformative when both counts are zero and the record still contains unmatched textual evidence.

The typed procedure adds outputs that the three counts cannot express. It marks semicolons inside an embedded entity label, preserves text assertions, distinguishes unresolved UUID-shaped values from unjoined authorities, identifies exact literal grant placeholders and flags disagreement between CSV tokens and embedded names. It also returns explicit abstentions for organization counting, parent attribution, grant-to-funder assignment and causal export-loss explanations. These rules were developed after observing the 26 cases. The selected records are development cases, not a held-out test set, and the protocol is not described as preregistered.

A semicolon can belong to one entity name

The clearest delimiter counterexample is handle 20.500.12854/45768. Its CSV name contains “Ecosystems Services for Poverty Alleviation (ESPA); University of Southampton”. In the retained xoai record, that entire string is the name of one embedded source entity with one identifier. Splitting the CSV produces two token occurrences, while the embedded-ID reference is one. The result does not prove that the string describes one real organization. It proves the more limited, directly inspectable point that semicolon splitting does not reproduce the represented entity structure in this case.

The European Commission sentinel shows the same problem more strongly. Handle 20.500.12854/182950 contains a name string combining the European Commission, the European Research Council and the University of Passau. These three apparent components belong to one embedded entity name. The token baseline returns three against one embedded identifier, a signed difference of two. By contrast, the Max Planck sentinel contains two embedded entities for the Society and Institute. A single delimiter rule cannot distinguish these two representation patterns, and neither pattern supplies an authoritative organizational hierarchy.

Across the six hash-selected semicolon-name cases, the token baseline disagrees with the embedded-ID reference in one case, with a signed total difference of one. It agrees numerically in the other five. Such agreement does not certify complete ascertainment: handle 20.500.12854/157342.2 contains two embedded names corresponding to the CSV tokens but also two additional text-valued funding labels in an English-language branch. Counting the two joined entities is a reproducible operation; treating those two as the exhaustive set of funding assertions would discard other represented evidence.

Extra identifiers and the meaning of a blank name

Raw UUID-value counting disagrees with the embedded-ID reference in four of the six single-name cases and two of the six semicolon-name cases. The signed differences total five and three respectively. Some records contain an embedded entity identifier plus additional UUID-shaped values that have no corresponding embedded entity block in that response. The procedure does not label those additional strings as missing funders, awards, duplicate edges or migration remnants. Each would be a semantic claim requiring evidence that the present comparison does not supply.

The grant-only group makes the distinction especially important. All six selected records have blank CSV funder-name cells but explicit textual labels and authority identifiers in xoai. Two text labels name the H2020 European Research Council, three name Knowledge Unlatched and one names the National Endowment for the Humanities. None of the six contains a same-record embedded funder entity. Each also has a UUID-shaped relationship value different from the authority identifier. Consequently the raw-UUID baseline returns one against zero embedded IDs in each case, while the token and exact-join baselines return zero.

Those zero counts must not become conclusions that no funder is identified anywhere. The records contain textual assertions that a researcher can report as such, even though the chosen embedded-ID reference is empty. The typed output therefore exposes the text and authority evidence while declining to manufacture an organization count. Among the six cases with neither CSV field populated, five have no extracted funding relationship and one has an unresolved UUID-shaped value plus a literal grant placeholder. Neither an absent relationship nor an unresolved value establishes whether the underlying book actually received funding.

Numerical agreement is not an accuracy result

The exact authority-or-value join equals the embedded-ID count in every observed case. This is a property of these retained structures: each embedded identifier appears among the corresponding record's values or authorities. It is not an independent success rate for funder identification. The embedded structure supplies both the reference inventory and part of the evidence used by the join, so the comparison is intentionally close to the representation. Its value is in exposing the operation and its boundary, not in demonstrating superiority over an external truth standard.

The signed differences are kept separate by selection group. Raw UUID-value counting differs in all six grant-only cases and one of the six neither-field cases, as well as in both sentinels. The Max Planck sentinel has four UUID-shaped values against two embedded IDs; the European Commission sentinel has two against one. Across the 24 hash-selected cases, embedded-ID occurrences total 17 when summed over records, while the two sentinels contribute three more. These are record-level source-ID occurrences, not the number of distinct organizations in the catalogue, and are not used to calculate concentration.

Grant strings require an equally modest interpretation. The exact literals “[...]”, “[…]” and “[grantnumber unknown]” appear in the retained source fields. They are source text, not truncation introduced by the display. Placeholder flags occur in four single-name cases, two semicolon-name cases, one grant-only case and one neither-field case; both sentinels also contain placeholders. Other nonblank grant text is preserved but not independently validated as an award identifier. The procedure never pairs grant positions with name positions, because adjacency or equal list lengths would not establish a relationship.

Constructed tests and explicit undefined denominators

Twelve synthetic fixtures exercise the decision rules separately from the observed records. They include a semicolon-bearing entity name, multilingual alias and text assertions, absent embedded metadata, an extra UUID-shaped value, duplicated values, grant placeholders and a renamed CSV label. Further fixtures cover duplicate entity blocks, an embedded entity without a matching value or authority, two source identifiers sharing a label, and invalid or multiple identifiers within one block. These are deliberately constructed software tests. Their frequency and arrangement say nothing about how often such patterns occur in DOAB.

The embedded-entity-without-match fixture is especially useful because its reference count is one while its exact-join count is zero. It demonstrates that the agreement seen in the 26 observed records is not guaranteed by all possible valid inputs. The renamed-label fixture has equal numerical counts but an explicit label-disagreement flag, showing why equal totals can conceal differences in content. The duplicate-value fixture returns two raw occurrences against one embedded identifier. These tests make the operational meaning of each output inspectable without inventing additional observations.

The aggregation tests also check abstention rather than merely testing the record classifier. A synthetic mixed group contains ten defined references and two undefined references; a separate group contains only the two undefined cases. Each summary reports its reference-defined and reference-undefined denominators, together with compared and abstained counts for every baseline. Undefined-only sums are null, not zero. This prevents a failed identification contract from disappearing into an apparently complete denominator. All 26 observed cases have defined references under the specified identifier-shape rule, so the robustness tests do not change their reported comparisons.

Reproduction, timing and a practical stopping rule

The compact accompanying inputs contain only the selected public bibliographic metadata needed for these structural decisions, together with aggregate outputs and separately labelled synthetic fixtures. They exclude whole books, abstracts, full XML responses and the complete CSV. An offline script reproduces the per-case comparisons, grouped summaries and abstention regressions from explicit input and output paths. Original-source hashes and request URLs allow separately obtained source bytes to be checked, but the default reproduction does not contact DOAB or require private files. Metadata reuse follows the documented DOAB CC0 policy; that policy is not a licence for the books described.

The comparison remains a comparison of retained representations at different retrieval times. All 26 record datestamps precede CSV retrieval, but those datestamps do not establish the CSV generation time or synchronization of linked entity metadata. The observed contrasts therefore do not identify an exporter bug or causal projection loss. The literature review is also targeted rather than systematic: the closest 2022 journal full-text route and some official reporting were not fully accessible, while author descriptions and a later analysis were available. These boundaries preclude broad uniqueness claims, not the description of the structures actually retained.

For practical use, the stopping rule is straightforward. If the task is to inventory embedded source identifiers, perform and report that operation, retain unmatched assertions and keep the scope visible. If the task is to count real organizations, attribute parent bodies or connect awards to funders, stop when the required semantics are not established. More nonblank fields do not automatically justify more attribution. This finite case study shows why a useful metadata result can consist of a supported inventory plus explicit unresolved evidence, rather than a single confident number that answers a different question.

Count disagreements against embedded source-ID structure

Finite counts by CSV state with purposive sentinels separate; no population rates or organization accuracy.

Full chart and source
Across single-name, semicolon-name, grant-only and neither-field groups of six, token-count disagreement cases are 0, 1, 0, 0; raw-UUID cases are 4, 2, 6, 1; exact-join cases are all 0. Separate Max Planck and EC/ERC sentinels have token disagreements 0 and 1, raw-UUID disagreements 1 and 1, and exact-join disagreements 0 and 0.

Scroll horizontally to inspect the chart. ↔

Chart and values

Scroll horizontally to see all columns.

Count-disagreement cases relative to embedded source IDs
Selected groupCasesEmbedded ID occurrencesCSV-token disagreementsRaw-UUID disagreementsExact-join disagreements
Single-name hash sample66040
Semicolon-name hash sample611120
Grant-only hash sample60060
Neither-field hash sample60010
Max Planck sentinel12010
EC/ERC/Passau sentinel11110

Four hash-selected groups of six and two separate purposive sentinels. Counts compare recipes with distinct embedded UUIDs, not actual organizations or funding truth. Zero disagreements can coexist with unjoined textual assertions.

Each hash group contains six records; each sentinel contains one. A disagreement compares a recipe with distinct embedded UUIDs, not real organizations. Exact joins agree structurally here, not against an independent funding standard. DOAB CC0 metadata; original analysis and figure: Alisson Bento.

Download chart (SVG) ↓ · Download chart (PNG) ↓

Data & downloads

Scroll horizontally to see all columns.

Count-disagreement cases relative to embedded source IDs
Selected groupCasesEmbedded ID occurrencesCSV-token disagreementsRaw-UUID disagreementsExact-join disagreements
Single-name hash sample66040
Semicolon-name hash sample611120
Grant-only hash sample60060
Neither-field hash sample60010
Max Planck sentinel12010
EC/ERC/Passau sentinel11110

Four hash-selected groups of six and two separate purposive sentinels. Counts compare recipes with distinct embedded UUIDs, not actual organizations or funding truth. Zero disagreements can coexist with unjoined textual assertions.

Scroll horizontally to see all columns.

Signed count differences and defined reference denominators
Selected groupDefined referenceUndefined referenceCSV-token signed sumRaw-UUID signed sumExact-join signed sum
Single-name hash sample600+50
Semicolon-name hash sample60+1+30
Grant-only hash sample600+60
Neither-field hash sample600+10
Max Planck sentinel100+20
EC/ERC/Passau sentinel10+2+10

Sum of recipe count minus embedded-ID reference over defined cases only. All observed references are defined; synthetic undefined cases remain separate. No population rate or attribution-accuracy estimate.

Scroll horizontally to see all columns.

Constructed aggregation regression, not observed data
Synthetic groupCasesDefined referenceUndefined referenceCompared per baselineAbstained per baselineReference sum
Mixed fixtures1210210210
Undefined-only fixtures20202Not defined

These synthetic groups exercise the same summary function as the observed records. A null sum means no defined comparison, not zero disagreement.

Evidence & method

How we calculated it

  1. Frozen selection from 103,963 eligible book records in a pinned DOAB CSV; six hash-selected cases in each of four name/grant states and two separately reported purposive sentinels.
  2. All 26 selected xoai responses matched book UUID and handle. No substitutions. Snapshot timing and linked-entity synchronization remain unverified.
  3. Decision rules and synthetic fixtures were developed after observing the 26 cases; these are development cases, not preregistered protocol validation or a holdout.
  4. Reference observable is distinct valid UUID-shaped identifiers in embedded funder blocks, with exactly one identifier per block. Empty metadata means zero embedded IDs only; invalid or multiple identifiers make the reference undefined.
  5. Three explicit recipes count semicolon token occurrences, raw UUID-value occurrences, or exact same-record authority/value matches to embedded IDs. Disagreement is relative to source structure, not real organizations.
  6. Text, authority, unmatched values, entity aliases, asserted DOIs and grant strings remain separate. All language branches are included. No positional grant matching or semantic inference from UUID syntax.
  7. Public offline reproduction uses compact selected bibliographic structures, twelve synthetic fixtures and shared defined/undefined aggregation regressions; original-source bytes are fingerprinted but not redistributed.

What the data cannot tell us

  • Finite exploratory cases do not estimate population prevalence, actual-funding accuracy, organization concentration or financial amounts.
  • Exact joins agreeing with embedded-ID counts is a structural property of these cases, not independent identification accuracy or superiority.
  • Source entities can carry composite labels; source identifiers are not validated real organizations, awards or ultimate parents.
  • Nonblank grant text can be a placeholder; other grant strings and asserted funder DOIs were not externally validated.
  • Different retrieval times and unknown CSV generation/linked-entity timing prevent causal exporter-loss claims.
  • The targeted prior-art review is not systematic; inaccessible journal full text and official reports limit novelty assessment. No first-ever claim.

Original sources

  1. DOAB catalogue CSV

    Retained 30 October 2026: 114,618 rows, including 103,963 eligible book records. Generation timestamp unverified; whole file is not redistributed. Pinned SHA256 and selected-structure provenance are supplied.

    Accessed October 30, 2026

  2. DOAB fuller xoai metadata: selected record example

    Twenty-six selected responses retrieved on 30 October 2026, matched by book UUID and handle. This example has a semicolon-bearing single embedded entity name. Exact request URLs, byte counts and hashes for all selected responses appear in source-hashes.json.

    Accessed October 30, 2026

  3. DOAB metadata policy and formats

    Retained official documentation identifies metadata as CC0 and describes fuller xoai metadata. This is not a licence for book contents, a guarantee of CSV completeness or proof of synchronized representations; later alternate documentation routes were unavailable.

    Accessed October 30, 2026

  4. Tsuji (2018), Statistics on Open Access Books Available through the Directory of Open Access Books

    Full methods and enumerated results inspected in the targeted literature review: subject, language, publisher, year and licence statistics, not the typed same-record funding-linkage comparison. Difference from this paper is not universal novelty.

    Accessed October 30, 2026

  5. Grassano and colleagues (2016), Funding Data from Publication Acknowledgements: Coverage, Uses and Limitations

    Author manuscript methods address reference acknowledgements, database comparisons, name harmonization and organizational levels. The present metadata-only comparison lacks an actual-funding reference standard and does not estimate funding accuracy.

    Accessed October 30, 2026

  6. Kramer and de Jonge, open funder metadata author dataset description

    Author-deposited description, updated March 2026, describes a 5,004-output known-funder comparison and distinguishes names and identifiers. The closest 2022 journal full-text route was inaccessible; this description is not represented as a full-methods substitute.

    Accessed October 30, 2026

  7. Kramer (2026), open funding information analysis

    Accessible author analysis compares project-mediated and direct funder links and notes the differing snapshot dates of the databases. Linkage architecture and snapshot differences are prior methodological concerns, not newly invented here.

    Accessed October 30, 2026

  8. UKRI / MoreBrains, Metadata to support the UKRI Open Access Policy, revised 2022

    Sections 4, 6.7.1 and 6.7.6 distinguish schema capability from populated metadata and discuss funder names and IDs for long-form outputs. Book-funding completeness is not itself a novel gap.

    Accessed October 30, 2026

  9. Crossref funding-information specification

    Official specification distinguishes funder identities, award numbers and grant identifiers, requires structural association, and describes retrieval limitations for unmatched names. These rules are prior art, not deployment-specific semantics for DOAB UUID values.

    Accessed October 30, 2026

Cite & reuse this study

Alisson Bento. When a funder count changes with the metadata: a DOAB case study. Comparative analysis of selected Directory of Open Access Books metadata retained in 2026. https://alissonbento.com/research/doab-funding-metadata-counting-and-abstention/

Selected bibliographic metadata: Directory of Open Access Books (DOAB), under the documented CC0 1.0 metadata policy. This policy does not license the books described. Only selected structural metadata needed for the finite comparisons is redistributed; no abstracts, whole catalogue, full XML, book files, artwork or logos are included. Original comparative analysis, protocol and figure: Alisson Bento. Prior methodological work by Grassano and colleagues, Kramer and de Jonge, Kramer, UKRI/MoreBrains, and Crossref is credited in sources.json. No agency or source-provider endorsement is implied. No licence for the complete article, third-party prose or book contents is invented.

Corrections or editorial questions

Pass it on.

When a funder count changes with the metadata: a DOAB case study | Alisson Bentoalissonbento.com

WhatsAppFacebookXTelegramRedditThreadsEmailSMS