Shape: research paper
The shape of an empirical research article: IMRaD with a structured abstract, the baseline named in baseline, plus the few additions that have evidence elsewhere.
When to use it
- An empirical study (experiment, observational study, computational evaluation) reported to a journal, a conference or a preprint server.
- Not for a proof-centred mathematical article: mathematical-article.md covers that.
- Not for a review, an essay or a humanities article, whose structure follows the argument rather than the study.
- Inside a publisher class (elsarticle, llncs, IEEEtran, acmart), keep the class for layout and metadata and take the section order from here; the class does not supply one.
How each part maps to the evidence
| Part | What it does | Evidence | Grade |
|---|---|---|---|
| Title | states the finding or the question, since it is read first | Keshav 2007 | convention, untested |
| Structured abstract | Background, Methods, Results with numbers, Conclusions | Taddio 1994, Budgen 2008, Hartley 2004 | measured (information content) |
| Highlights, graphical abstract | optional extras some journals require | Elsevier highlights, Oska 2020 | convention, untested (attention only) |
| Introduction | territory, niche, this paper, then contributions and a map of the sections | Swales 1990, Hill 1982 | supported with caveat (describes practice) |
| Methods: plan and deviations | if the study was preregistered, the protocol and each departure from it with its reason; if not, that it is exploratory | Chambers 2013, Kerr 1998, Scheel 2021 | measured (for the format) |
| Methods: design, measures, analysis | enough detail to repeat the study, item by item | Turner 2012 | measured (completeness) |
| Results | one subsection per claim, a headline sentence first, a figure whose caption stands alone | Hubbard 2017, Keshav 2007 | supported with caveat |
| Discussion | the principal finding first, then comparison, limitations, implications | Hill 1982, Boutron 2010 | supported with caveat |
| Data and code availability | where the data, code and environment are, so the results can be re-run | Buckheit 1995, Pimentel 2019; the audits of availability statements (Gabelica 2022, Danchev 2021, Hamilton 2023) find "on request" weak, so give an archive and a version | supported with caveat |
| Reporting checklist | names the EQUATOR guideline followed, if one exists for the design; at computing venues, the venue's own checklist (NeurIPS, ACL) | Turner 2012, NeurIPS 2026, ACL 2023 | measured for EQUATOR (completeness); venue checklists convention, untested |
| Language-model use declaration | says whether and how a language model was used, worded to the venue's rule | ICMJE 2023, Nature 2023, ACL 2023, ICML 2023, NeurIPS 2025 | convention, untested |
| Declarations, references, appendix | funding, conflicts, contributions; the venue's reference style | ICMJE 2026 | convention, untested |
Each key (Author Year) is listed in references. No part of this shape has been tested against IMRaD on reader outcomes; position says what would test it.
The principles it applies
| Part | Principle |
|---|---|
| Title that states the finding; structured abstract | overview first |
| Contributions and a map of the sections, closing the introduction | overview first, signal the structure |
| Terms and measures defined in Methods before Results use them | define before use |
| Results: one subsection per claim, its headline sentence first | signal the structure, given before new |
| A figure whose caption stands alone, placed with the text that cites it | draw relations, figure next to its text |
| Headings on at most three levels | shallow outline |
Skeleton
Copy this into a new file, or start from the LaTeX version, research-paper.tex.
# Title that states the main finding
First Author (affiliation, contact), Second Author (affiliation)
## Abstract
**Background.** The problem and the gap, in one or two sentences.
**Methods.** Design, sample and main measure.
**Results.** The main result with its number and interval.
**Conclusions.** What the result means, and its main limit.
**Keywords:** three to six terms a searcher would type.
## Highlights (optional)
> Three to five bullets of at most 85 characters each, only if the venue asks for them.
## 1. Introduction
> Establish the territory, show the gap, say what this paper does, list its contributions, and give a one-line map of the sections.
## 2. Methods
### 2.1 Plan and deviations
> If the study was preregistered, link the protocol and list every departure from it with its reason; if not, say that it is exploratory.
### 2.2 Design and materials
> Who or what was studied, how it was chosen, and with what, in enough detail to repeat it.
### 2.3 Measures and analysis
> Each outcome, how it was measured, and the analysis planned for it.
## 3. Results
### 3.1 First claim
> Open with the headline sentence of the result, then the number and its interval, then the figure.
Figure 1. A caption that says what the figure shows and what to conclude, readable without the text.
### 3.2 Second claim
> One subsection per claim, in the order of the contribution list.
## 4. Discussion
### 4.1 Principal finding
> The main result restated in one paragraph, without new data.
### 4.2 Comparison with prior work
> Where the result agrees or disagrees with earlier studies, and why.
### 4.3 Limitations
> What the design cannot show, and how large each threat is.
### 4.4 Implications
> What follows for practice or for the next study, kept within what the data support.
## 5. Conclusion
> Two or three sentences a reader could quote; no claim absent from the results.
## Data and code availability
> Where the data, code and computing environment are, under which licence, and how to re-run the analysis.
## Reporting checklist
> Name the reporting guideline followed (for example CONSORT or PRISMA) and attach the completed checklist.
## Use of language models
> Whether and how a language model was used, and for what, worded to the venue's rule; "none" is an answer.
## Declarations
> Funding, conflicts of interest, author contributions, acknowledgements.
## References
> In the venue's style; every entry with a DOI, an arXiv id or a stable URL.
## Appendix A. Supplementary material
> Material a specialist needs and a general reader can skip.Worked example
A short descriptive paper on a count made across this family on 2026-10-29, filled to the skeleton: each part says what the skeleton asks, with real numbers.
# Eighteen figures in six documentation repositories were typed in characters, half of them on front pages
The maintainers of the reader-ergonomics family.
## Abstract
**Background.** WCAG counts text art as non-text content, and no study compares a typed figure with a drawn one.
**Methods.** One script counted the figures typed in characters in six repositories, on their main branch of one day.
**Results.** 18 typed figures, in five of the six repositories; 9 of the 18 were on a front page or a standard's card.
**Conclusions.** Typed figures were the family's default where readers arrive first; a count says nothing of harm.
**Keywords:** text art, documentation, figures, accessibility.
## 1. Introduction
A figure typed in box-drawing characters reads well in a terminal but reaches a screen reader character by character.
The family had no rule on how to draw a figure. How many were typed, and where?
## 2. Methods
### 2.1 Plan and deviations
The count was planned before it ran and ran once. No deviation.
### 2.2 Design and materials
Six repositories, each read as an archive of its main branch on 2026-10-29. Files read by a model or a terminal
(AGENTS.md, CLAUDE.md) were left out.
### 2.3 Measures and analysis
A typed figure is a fenced block with no language, or marked text, that holds a box-drawing character
(U+2500 to U+257F) or a block arrow. The count is per repository and per kind of page.
## 3. Results
### 3.1 Eighteen typed figures, in five of six repositories
reader-ergonomics 7, topic 5, presentations 4, software-design 1, academic 1, learning 0.
### 3.2 Half were on a front page or a card
9 of the 18 were in a README or a standard.md, the pages a reader opens first.
## 4. Discussion
### 4.1 Principal finding
Typing was the default, and most of all on the first pages.
### 4.2 Comparison with prior work
No study of readers compares the two forms; the standard's rule rests on access, not on measured reading.
### 4.3 Limitations
One day; a figure typed with plain ASCII (`-->`) is not counted; the count shows presence, not harm.
### 4.4 Implications
Each figure was redrawn and offered to its repository's owner; files read by a model keep their typed figures.
## 5. Conclusion
18 typed figures, half on the first pages; the family now draws them from Graphviz sources.
## Data and code availability
The count is the typed-figure check of reader-ergonomics-presentations, checks/local/typed_figure.py.
## Reporting checklist
None applies: a descriptive count, with no participants.
## Declarations
No funding. An AI assistant ran the count and drafted the text; the maintainers decided the rule it informed.