Skip to content
Scriptorium Scriptorium

CSL and bibliographic standards

Last updated: 2026-05-17

Synthesis

Section titled “Synthesis”

Scriptorium does not generate bibliographies; it operates on bibliographies the author has already curated. This design choice rests on hallucination evidence (see hallucination-in-llm-citations) and on the practical observation that the scholarly identifier ecosystem already does this work well. The relevant standards split into three layers: a style layer (how citations are rendered) — Citation Style Language (CSL); a storage layer (how references are stored) — BibTeX/BibLaTeX, CSL-JSON, RIS; and an identifier layer (how works, authors, organizations, and funders are pinpointed) — DOI / CrossRef / DataCite, ORCID, ROR, the Open Funder Registry, and CURIE prefixes coordinated by the Bioregistry.

Each layer is independently mature. CSL governs nearly every modern reference manager (Zotero, Mendeley, Paperpile, and most academic word-processor integrations). Persistent identifier coverage is now high enough that “the bibliography exists in machine-readable form with stable identifiers” is a reasonable assumption for any contemporary manuscript. Manubot’s reliance on Bioregistry CURIE prefixes — currently spanning 1,677 identifier types — illustrates the practical reach of this infrastructure beyond pure DOI/ORCID lookups.

Scriptorium’s job is to read whatever bibliography the manuscript ships with, treat it as authoritative, and never silently introduce references that are not in it. That posture only works because the existing standards are good enough to make “the author’s bibliography” a stable artifact to anchor against.

Evidence

Section titled “Evidence”

Citation Style Language (CSL). Originally conceived by Bruce D’Arcus in 2006 for OpenOffice citation rendering, formalized as an XML-based specification. Early CSL adoption was driven by Simon Kornblith’s Zotero integration; Mendeley followed and (with Columbia University Libraries) obtained an Alfred P. Sloan Foundation grant to build the CSL style editor. Mendeley and Elsevier have donated to the CSL project since 2014.1

The current specification (CSL 1.0.2) defines:

  • A style language (XML) describing how to render citations and bibliographies for a given journal or discipline.
  • A data interchange format (CSL-JSON) for individual references.
  • A locale mechanism for language- and region-specific conventions.

The CSL Style Repository hosts thousands of journal- and discipline-specific styles, maintained by community contributions. Zotero, Mendeley, Paperpile, Pandoc (via citeproc), and several journal submission systems consume CSL natively. This makes CSL the de facto interoperability standard for “what should this citation look like in this venue.”

BibTeX / BibLaTeX. The older LaTeX-native ecosystem remains dominant in physics, mathematics, computer science, and statistics. BibTeX’s @article{...} syntax is well-understood and stable. BibLaTeX is the modern successor with richer entry types and Unicode support. Most reference managers can export to BibTeX, and most manuscript workflows in LaTeX-heavy fields use BibTeX as the canonical bibliography format. Scriptorium should accept BibTeX/ BibLaTeX as a first-class input.

DOIs and CrossRef. The Digital Object Identifier system, operated for scholarly content primarily by CrossRef and DataCite, is the most widely deployed persistent identifier for academic works. CrossRef holds metadata for ~175 million scholarly works as of the October 2025 public data file (https://www.crossref.org/blog/2025-public-data-file-now-available/; the corpus continues to grow at ~5M records/year, with the 2026 public data file reporting ~180M records) and exposes structured metadata via a public REST API. CrossRef’s metadata records increasingly carry references to other works, ORCID author identifiers, and ROR institutional identifiers — turning the citation graph itself into queryable infrastructure.2

DataCite. Mirrors CrossRef for datasets, software, and other research outputs. Particularly important in data-citation contexts; DataCite’s metadata schema is closely aligned with CrossRef’s to support interoperability.

ORCID. Persistent identifier for individual researchers. ORCID records can be tied to publications, peer reviews, grants, and affiliations. The ORCID/CrossRef auto-update system propagates publication records to ORCID profiles automatically when an article is registered with CrossRef. This makes ORCID a reasonable disambiguator for “who is the author” in contexts where lastname-only attribution is ambiguous.3

ROR. The Research Organization Registry, launched in 2019, is jointly run by California Digital Library, CrossRef, DataCite, and (formerly) ORCID. ROR provides CC0-licensed open identifiers for research organizations, replacing the historically fragmented landscape of institutional identifiers. ROR is now the de facto standard for affiliation identification in CrossRef metadata.4

Open Funder Registry (FundRef). CrossRef’s funder identifier service. The registry contains over 15,000 funder entries; CrossRef holds funding metadata on more than 1.7 million works. This matters for MANUSCRIPT_STATE.yaml declaring grant numbers and funder identifiers consistently.

Bioregistry and CURIE prefixes. The Bioregistry is a meta-registry that catalogs identifier prefixes across biology, biomedicine, and adjacent fields. CURIEs (Compact URIs) take the form prefix:identifier (e.g., doi:10.1136/bmj.b2680, clinicaltrials:NCT04280705, pmid:19622839). Bioregistry currently catalogs 1,677 prefixes (as referenced by manubot). Manubot — the open scholarly-publishing tool — initially used Identifiers.org for ~700 prefixes, then migrated to Bioregistry for both broader coverage and more rigorous validation (open contribution model; better regex patterns; resolution of namespace inconsistencies).56

This matters for scriptorium because it means a manuscript’s “citations” need not all be DOI-based. A reference to clinicaltrials:NCT04280705 or arXiv:2401.01234 is a fully machine-resolvable citation under existing standards, and scriptorium should treat such references as first-class.

How this informs scriptorium

Section titled “How this informs scriptorium”

Bibliography is input, not output. Scriptorium reads BibTeX, CSL-JSON, or in-manuscript citation keys (Pandoc-style [@doi:10.1136/bmj.b2680] or \cite{...}) and treats the resulting reference set as the canonical “what citations exist in this manuscript.” Any skill operating on the manuscript can ask: is this in-text citation present in the bibliography? That is a syntactic check, not a generative act.

Citation audit operates on the citation set, not on free-text references. Because CSL/BibTeX entries carry stable identifiers, citation-audit does not have to parse rendered prose to figure out which paper a citation refers to. It can ask the structured question: “in-text citation [Smith 2019] resolves to bibliography entry smith2019, which has DOI 10.1234/abcd. Does the manuscript’s claim accompanying this citation match what we know about that DOI?”

No skill writes new bibliography entries. Even when reviewer-simulation recommends “consider adding a recent meta-analysis,” the recommendation does not propose specific bibtex keys, DOIs, or author strings. Adding bibliography entries is the author’s act, using their reference manager and the identifier-resolution services above.

MANUSCRIPT_STATE bibliography pointer. The state schema includes a bibliography field pointing to the canonical bibliography file (e.g., bibliography: refs.bib). Skills resolve in-text citations against this file. If the field is absent, citation-aware skills can still operate on in-text citation keys but should emit a warning that the bibliography link is missing.

Manubot pattern as design influence. Manubot’s three-stage pipeline (Pandoc parses citation keys → manubot-cite filter retrieves metadata via CrossRef / Bioregistry → citeproc renders styled output) is a clean separation of concerns that scriptorium should mirror in spirit: identifier resolution, metadata retrieval, and rendering are three different jobs, and conflating them is where errors enter.

Open questions / weak evidence

Section titled “Open questions / weak evidence”
  • Coverage of non-biomedical fields by Bioregistry is uneven; scriptorium should not assume every domain has a comparable identifier catalog.
  • Some venues still demand plain-text references with no identifiers (older journals, certain humanities outlets). Scriptorium degrades gracefully to “citation key only” in those cases but loses the ability to do identifier-level verification.
  • The interaction between CSL styles and journal-specific pre-submission validation is messier than the standards suggest; journals’ production pipelines often diverge from their published style guides.

References

Section titled “References”

Footnotes

Section titled “Footnotes”

  1. Citation Style Language. About. https://citationstyles.org/about/ ↩

  2. CrossRef. Public REST API and infrastructure documentation. https://crossref.org/ ↩

  3. ORCID. Public-facing documentation. https://orcid.org/ . CrossRef–ORCID auto-update: https://www.crossref.org/community/orcid/ ↩

  4. Research Organization Registry (ROR). About page. https://ror.org/about/ ↩

  5. Hoyt CT, Balk M, Callahan TJ, et al. Unifying the identification of biomedical entities with the Bioregistry. Scientific Data. 2022;9:714. doi:10.1038/s41597-022-01807-3. ↩

  6. Manubot. Usage documentation, citation prefix support. https://github.com/manubot/rootstock/blob/main/USAGE.md ↩