Discipline conventions beyond biomedicine

Last updated: 2026-05-17

Synthesis

Scriptorium’s knowledge base is biomedical-coded. IMRaD structure, reporting guidelines like CONSORT and STROBE, first-author/last-author anchors, and significance-positioning in the NIH style (nih-significance-patterns) are conventions of the life sciences and medicine. Other disciplines write differently — sometimes radically so. A physics PRL, a NeurIPS conference paper, a single-author mathematics theorem-proof paper, an economics working paper, and a humanities monograph chapter are different genres with different success criteria. A tool that silently assumes biomedical norms will produce confidently wrong critique on any of them.

The defensible move is not to build per-discipline skill variants (high effort, uncertain value, ongoing maintenance) but to state the project’s scope explicitly in DESIGN.md and in each skill’s description. Swales’ 2004 Research Genres [1] gives the cross-disciplinary framework: genres vary along dimensions of structure, citation practice, audience model, and epistemic norms, and the variation is large enough that a single editing tool cannot pretend neutrality. Saying so up front is more honest — and more useful — than pretending otherwise.

Evidence and frameworks

Physics: PRL Letters and arXiv preprint norms

Physical Review Letters enforces a strict 4.5-printed-page limit (approximately 3500 words) and explicitly forbids the standard IMRaD section structure. A PRL is a single continuous argument, with methods compressed into the body or relegated to supplementary material. The implicit reader is assumed to know the field’s prior work and notation; introductions are typically a single paragraph.

The dominant submission norm in physics is to post to arXiv at or before journal submission. The result is that the preprint is the canonical version for the community; journal publication serves as post-hoc validation. Scriptorium skills that assume “submission” is a singular event will model this poorly.

Computer science / machine learning

NeurIPS, ICML, ICLR, and related venues use 8–10 page main paper + supplementary appendix as the canonical format, with double-blind peer review handled in 6–8 week conference cycles. The 8-page constraint is real and load-bearing: it forces a writing style with dense figures, compressed related work, and methods deferred to the appendix.

Specific conventions:

• Authorship is typically contribution-based, with no alphabetical convention.
• Code and data release are increasingly mandatory for camera-ready acceptance.
• Supplementary material has expanded rapidly; some papers’ appendices now exceed the main paper in length.
• The “related work” section serves a different function than a biomedical introduction: it is closer to a positioning argument than a literature review.

A scriptorium audit calibrated for “is the introduction sufficiently contextualised” will misfire on an ICML submission whose introduction is intentionally compressed.

Mathematics: theorem-proof-corollary

Mathematics papers are organised around the definition–theorem–proof–corollary–remark structure rather than IMRaD. Krantz’s A Primer of Mathematical Writing [2] is the standard reference for the genre. Key features:

• Figures are rare; the work happens in notation.
• Citation practice is sparser and more targeted; “see Smith (2018), Theorem 3.2” rather than a paragraph of context.
• Single- or few-author papers dominate; alphabetical authorship is the default outside biomedical/medical-statistics collaborations.
• Significance-positioning (significance-positioning) looks different: results are positioned relative to a known open problem or a structural place in a theory, not relative to population health.

A reviewer-simulation skill that probes for “what’s the clinical relevance” of a result in algebraic topology is not just unhelpful — it is misframed.

Economics

Economics papers are typically long (40–80 pages including appendices), with extensive identification-strategy discussion, robustness tables, and theoretical appendices. Alphabetical authorship by surname is the field’s default convention; deviations are read as informative (a non-alphabetical order signals contribution order). Working papers are first-class objects; journal publication often comes years after the working paper has circulated.

John Cochrane’s “Writing Tips for PhD Students” [3] is the most widely cited internal guide and is worth reading just to see how different the field’s prose norms are: it explicitly tells students that “you are primarily a writer”, emphasises the one-paragraph contribution statement, and advises against travelogue-style methods sections.

Humanities and qualitative social sciences

Humanities scholarship is argument-driven rather than IMRaD, typically organised around an interpretive thesis defended through close engagement with primary and secondary sources. The epistemology is fundamentally different: claims are not validated by replication or effect sizes but by the persuasiveness of the argument and the depth of the engagement with prior scholarship.

Qualitative social science (ethnography, critical theory, interpretive policy analysis) shares this orientation. Citation practice is denser and more conversational (“as Foucault notes…”). Reporting guidelines analogous to CONSORT or PRISMA largely do not apply, though there are emerging quality standards for some qualitative subfields (e.g., COREQ, SRQR).

A reviewer-simulation calibrated on biomedical reviewer archetypes (reviewer-archetypes-evidence) will not approximate a humanities reviewer. The epistemic dispositions are not the same.

Engineering

Engineering bifurcates roughly into design / system papers (novel artifact, characterisation, performance evaluation) and applied papers (methodology applied to a problem). Journal conventions vary by subfield (IEEE Transactions style is its own universe). The genre is closer to ML than to biomedicine — figure- and table-dense, with methods often embedded in implementation details rather than separated.

The cross-disciplinary frame

Swales’ 2004 Research Genres [1] argues that genres differ along at least the following dimensions, all of which matter for editing:

• Structural conventions (IMRaD vs. argument-essay vs. theorem-proof).
• Citation density and integration style (cf. esl-writers-swales-hyland on integral vs. non-integral citations).
• Author–reader relationship (presumed shared knowledge, formality, use of “we”).
• Evidence types (experimental data, mathematical proof, archival sources, computational benchmarks).
• Authorship conventions (alphabetical, contribution-order, first-last anchored).

Treating all five of these as biomedical defaults is a scope choice — scriptorium should be honest about it.

How this informs scriptorium

1. Add an explicit scope statement to DESIGN.md. Something like: “Scriptorium’s knowledge layer is calibrated for biomedical and life-sciences manuscripts. Skills will run on physics, CS, mathematics, economics, and humanities texts, but the heuristics embedded in critique skills assume biomedical conventions. Users working outside that scope should treat the output as advisory, not authoritative.” This is a low-effort/high-value addition.

2. Encode discipline in MANUSCRIPT_STATE.yaml. The schema already has target_venue; add (or document) a discipline / field hint that skills can use to dial down or qualify their feedback. Even a single string (“biomedical” | “ml” | “physics” | “math” | “economics” | “humanities” | “engineering” | “other”) is enough to let skills emit “this check assumes biomedical conventions and may not apply to your field” warnings.

3. Don’t build per-discipline skill variants in v0.1–v0.3. The incremental value of argumentative-flow-physics over a well-scoped general argumentative-flow is small. The cost of maintaining five variants is high.

4. Reviewer-simulation needs disciplinary calibration if it is to generalise. The biomedical reviewer archetypes (reviewer-archetypes-evidence) are not portable. Cross-disciplinary calibration is a v0.4+ problem.

Implementation priority for scriptorium

Verdict: No new skill in v0.1–v0.3. The implementation priority is a documentation change: a scope statement in DESIGN.md and a one-line scope note in each critique skill’s description.

If Yes (later): in v0.5+ (“knowledge layer + platform adapters” phase in DESIGN.md), if scriptorium gains traction outside biomedicine, add discipline-specific knowledge files under knowledge/disciplines/{physics,ml,math,economics,humanities}.md that individual skills reference explicitly. Required data: a discipline hint in MANUSCRIPT_STATE.yaml and discipline-tagged reviewer archetypes.

Why useful context anyway: This document scopes scriptorium’s claims. The biggest reputational risk for an agentic editing tool is overconfident wrong output on a manuscript type it wasn’t designed for. Naming the scope is cheap defence.

Condition that flips to Yes: a non-biomedical user community emerges and complains that scriptorium’s biomedical defaults are miscalibrated for their work. That’s the trigger for knowledge/disciplines/ content, not before.

Open questions / weak evidence

• The discipline boundaries are fuzzy. Computational biology shares ML conventions; biostatistics shares economics conventions; digital humanities shares CS conventions. A flat discipline tag in MANUSCRIPT_STATE.yaml will mislabel boundary cases.
• “Biomedical conventions” themselves vary — basic-science papers in Cell read differently than clinical trials in NEJM. A one-axis discipline scope is a simplification.
• Conference-vs-journal is sometimes a bigger split than field-vs-field. NeurIPS papers and biomedical papers differ less along some dimensions than NeurIPS and Journal of Machine Learning Research papers.

References

[1] Swales, J. M. (2004). Research Genres: Explorations and Applications. Cambridge University Press. ISBN 9780521533348 (paperback; hardback ISBN 9780521825603). [TODO verify hardback ISBN.]

[2] Krantz, S. G. (2017). A Primer of Mathematical Writing, Second Edition. American Mathematical Society. ISBN 9781470436582. (The originally referenced ISBN 9781470414597 appears not to correspond to the published second edition.)

[3] Cochrane, J. H. (2005, updated). Writing Tips for PhD Students. Available at https://www.johnhcochrane.com/research-all/writing-tips-for-phd-studentsnbsp (no DOI; widely cited in economics-graduate-training materials).

[4] Physical Review Letters author guidelines. American Physical Society. https://journals.aps.org/prl/authors. (Style and length norms as of 2024–2026.)

[5] International Committee of Medical Journal Editors. ICMJE Recommendations. https://www.icmje.org. (Provided as the biomedical reference point against which other-discipline conventions are contrasted.)