Review

This file reviews the sources on system diagrams by strand: existing reviews first, then theory, controlled studies, layout, and practice. Numbers are from the primary source unless marked (secondary: read in an abstract or in another paper). Each source's quality is in sources.

0. Existing reviews (Garousi 2019, Guideline 2)

Sixteen reviews were found. None includes grey literature and none is about reader outcomes of system diagrams.

On UML evidence: - Pretorius 2008: mapping study of 33 papers; the largest group studies comprehension. - Budgen 2011: 49 publications to end 2008; researchers take the UML as given (secondary). - Fernández-Sáez 2013: mapping of 38 papers (63 experiments, 3 case studies; 66 empirical studies in all) on UML in maintenance. The abstract says only 3 papers test help for code, but the body names two (P5, P8). Generalisation doubtful (full text read). - Guo 2017: a further-evidence review on UML; found, not read. - Koç 2021: 128 papers, 2000-2019; UML used mostly for design, class diagrams most (secondary). - Baldassarre 2019: belief that models help maintenance is scarcely tested, mostly UML (secondary). - Alhazmi 2026: 26 studies of sequence diagrams in education; comprehension is the main topic (secondary).

On notation and understanding: - van der Linden 2019: 70 applications of the Physics of Notations (a design theory for visual notations, strand 1); new notations adopted it with few verifiable details. - Dikici 2018: 1066 retrieved, 45 analysed, process-model understandability; findings inconsistent; thresholds for size unknown (full text read: 1066 retrieved, 480 duplicates, 45 analysed). - Dikici describes Störrle 2011, 2012 and 2014 (S40 to S42) as a series of experiments with 78 participants. - The Störrle papers report separate groups (see the Störrle lines below), so that description is Dikici's, not a fact about the data. - Houy 2012: no consensus on what model understandability means (secondary).

On visualisation: - Shahin 2014: 53 of 23,056 papers, 1999-02-01 to 2011-07-01; graph, notation, matrix and metaphor techniques; few adopted in industry (full text read). - Chotisarn 2020: 105 of 1440 entries, 2013-2019, sorted by structure, behaviour, evolution. - Merino 2018: 387 SOFTVIS/VISSOFT papers, 181 studied; 62 of them lack a strong evaluation (abstract; secondary). - Filho 2019: mapping of visual solutions for architecture evolution; found, not read. - Obaidellah 2018: 63 eye-tracking studies in programming; UML is the most used representation after Java code (secondary). - Rahmanian 2026: 64 studies of language models drawing UML and ER diagrams; evaluation is by metrics and rater scoring of the generated diagrams, not by reader tasks (primary abstract).

How this review differs: - It isolates reader outcomes of diagrams of a system (layout, size, key, labels, notation). - It adds 21 scored grey sources and a practice baseline. - It rechecks the primary papers: the layout and size results rest on one group's data (strand 2).

1. Theory of notations

  • Moody 2009a: the Physics of Notations is a design theory of perceptual properties of notations (secondary: abstract).
    • Its nine principles: semiotic clarity, perceptual discriminability, semantic transparency, complexity management, cognitive integration, visual expressiveness, dual coding, graphic economy, cognitive fit (secondary: Guerreiro 2023 tabulates them).
  • Moody 2009b applies it to the UML family and Moody 2010 to i*; both are analyses with no participants.
  • Störrle 2013 applied it to UML use case diagrams and found it too imprecise to apply objectively (secondary: abstract).
  • Santos 2018: 57 novices, standard versus more transparent i* symbols, with a language key; no gain in accuracy or speed, less visual effort. This is the one test found that has a key present, with no key-absent arm.
  • van der Linden 2018: 104 practitioner responses rank notation requirements differently for experts and non-experts (secondary).
  • Green 1996 and Blackwell 2003: cognitive dimensions, a vocabulary without scores.
  • Grade: convention; experimental support is indirect.

2. Controlled studies of UML and layout

  • Arisholm 2006: for complex maintenance tasks, UML documentation improved correctness and design quality, with no time saving.
  • Scanniello 2018: a meta-analysis of 12 experiments (333 observations, in Italy and Spain; full text read); design-phase models help source-code comprehension, analysis-phase models hurt.
  • Fernández-Sáez 2014: a family of experiments on level of detail; result not read.
  • Störrle 2011: 77 BSc and MSc students in three experiments (78 took part, 77 completed); each participant saw a given model in one layout only.
    • Right answers 6.35 versus 6.76 of ten, +6.5%; time per right answer 38.37 s versus 31.68 s, 17.4% faster; clarity 5.61 versus 7.81, +28.2%.
  • Störrle 2012 reports experiments D to F (class, sequence and state machine diagrams, about 78 participants); not read, described in Störrle 2014 and 2016b.
  • Störrle 2014 and Störrle 2016 re-analyse the 2012 experiments D to F (78 participants, over 1200 measurements), not the 2011 data; they are not new experiments.
    • Performance falls with diagram size (slope -0.014 per element); correlation exceeds -0.7 for poor layouts; experts are less affected.
  • Störrle 2016b pools the 2011 and 2012 data into 156 participants, 60 diagrams of five UML types, about 14,000 data points; the participants again come from earlier experiments of the same group.
    • Size correlates with mean score at r = -0.270 (medium); layout flaws at r = -0.419 (large).
    • Under its counting rules the centre size is 44 elements (all data), 49 (2011 data) and 65 (2012 data); under the 2014 rules it was about 50.
    • Its guidelines: at most 15-20 flaws, a flaw rate of at most 0.5, and fix flaws before cutting elements.

3. Layout aesthetics and eye tracking

  • Purchase 2002: users' preferences among aesthetics on UML differ from those on abstract graphs (secondary).
  • Eichelberger 2009 is quoted in Störrle 2014 as finding no significant impact of quality on class diagrams (secondary; the paper itself is behind a paywall).
  • Ware 2002: on path finding, after path length, continuity and edge crossings cost most (secondary).
  • Yusuf 2007: layouts that carry design information, and stereotypes, helped class-diagram tasks (secondary).
  • Sharif 2010: 15 participants, 8 role-detection questions, eye tracker; canonical layout improved time for four patterns and accuracy for two.
  • Found, not read: Sharif 2009a and 2009b (layout of class diagrams), Soh 2012 (expertise), Byelas 2006 (areas of interest on architecture diagrams).
  • Cepeda Porras 2010 compared pattern representations in class diagrams (result not read; paywall).

4. Practice of drawing

  • Cherubini 2007: whiteboard diagrams are transient and decisions drawn there are often lost (secondary).
  • Baltes 2014: 394 respondents; most recent sketches had some UML elements but were informal, and over half were on paper or whiteboard.
  • Petre 2013: 50 interviews, 5 patterns of UML use (secondary). Mangano 2015 studies sketching at the whiteboard (found, not read).
  • Dobing 2006 and Störrle 2017 survey industrial use of UML, and how widespread it is stays contested.
  • Ozkaya 2020: 109 practitioners from 34 countries, 35 questions; class diagrams lead for structure, deployment diagrams for physical structure (full text read).
  • Hebig 2016: 21,316 UML diagrams in 3,295 of the 1.24 million projects scanned (10% of GitHub), mostly made at project start (secondary). Ho-Quang 2017: collaboration is the main motive (secondary).
  • Koning 2002: guidelines for IT-architecture diagrams, tested with three groups of architects (secondary).
  • Nicacio 2022: 5 architecture diagrams from company blogs shown to 34 respondents (132 invited); their descriptions matched the author text poorly; 41% disagreed that all expected elements were present. Five guidelines follow: identify elements, define notation, give detail, mark start and end of reading, use technology symbols.

Q/A sites (Software Engineering Stack Exchange and Stack Overflow, searched 2026-10-28; 30 threads, scores in sources). Opinion, no measurement with readers.

  • How practitioners use UML:
    • SESE 2010g: formal UML is never used, "just enough" is. SESE 2012i: a practitioner with a UML certification stopped using it except for sequence diagrams.
    • SESE 2017g: class, sequence and state diagrams are still used on whiteboards, selectively. SESE 2023a: diagrams do not run, diff or show in pull requests.
    • SESE 2015e: free software rarely uses UML; Fowler's four modes of use (notes, sketch, blueprint, programming language) explain it. SESE 2015d: UML as blueprint belongs to a building analogy that fails, and diagrams without code to validate them mislead.
    • They agree with Baltes 2014, Petre 2013 and Arisholm 2006 as reported above. The finding that practice does not draw the UML baseline now rests on consistent sources of several kinds. They are opinions, not counts.
  • Abstraction against generation (a new framing): SESE 2011i (70 votes on the answer). A diagram easy to read lacks the detail for code generation, and one usable for generation helps no more than the code.
  • Audience: SESE 2011j gives diagram types by audience; SESE 2013c says a slide of boxes is "marketecture", not UML, with no rule; SESE 2011g says a diagram for executives depends on the audience. The marketecture picture against the technical diagram is new to the pool and supports a rule that names the audience of a diagram.

5. Practice: standards, notations, tools and guides

  • OMG 2017 (UML 2.5.1), Annex A: a diagram is a graphical representation of part of a model; frame and heading are optional. This reading found no layout rule there. OMG 2024 (SysML v2) is the systems successor; only its landing page was read.
  • The C4 model (Brown 2026b, Brown 2026d): a title with type and scope; a key; element type and short description; technology on containers; one-way labelled arrows; a 21-question review checklist.
    • It is its owner's guidance, cites no study, and scores below the quality threshold.
    • Brown 2026c offers adoption claims (10,000 people taught) and advises keeping UML if it works.
  • Other producers state similar rules.
    • Label every line, add a key and a title (IcePanel 2023).
    • Name the product beside each icon (Microsoft 2026 icons).
    • One concept per diagram, with alt text (Fuchsia 2025).
    • Same shape, same meaning; shapes, not colour (GitLab 2026).
    • Text contrast of 4.5:1 (AWS 2026b).
  • Data-flow diagrams: symbol sets (Wikipedia 2026b, secondary); trust boundaries drawn as lines or boxes and flows labelled with protocol (Microsoft 2023); high-level overview plus detail (OWASP 2026).
    • Schneider 2024: 24 participants; a DFD traced to code raised correctness of security answers by 41%. Mbaka 2024 is a planned experiment with no results.
  • Industrial use of C4: Jongeling 2025, 48 respondents (about 30% response) after two years of C4 in PlantUML beside code; most felt successful; 7 named limited layout control. The figures were not checked: the chapter was not reached, because the university has no access; the abstract confirms the 48 respondents, not the 30% or the 7. Mavrogiorgou 2025 and Jokste 2024 describe use without measuring readers.
  • Cloud icons: AWS 2026 publishes icon sets; its icon page gives no drawing rules.
  • Diagrams as code: GitHub renders Mermaid since 2022-02-14 (GitHub 2022). Brown 2021 names drift as the main failure.
  • On 2026-10-28 Mermaid had 90,911 stars, PlantUML 13,383 and D2 25,750 (Mermaid 2026, PlantUML 2026, D2 2026), and 69.4 million npm downloads in a month (npm 2026): adoption counts, not reader benefit.

Q/A sites (searched 2026-10-28), diagrams as code and drift. Opinion, no figure.

  • SESE 2017e, SESE 2019b, SESE 2020a, SESE 2017d and SESE 2012g: diagrams go stale. Generate them where possible, put the update in the scope of the work, and keep the diagram source in version control with a header block. A model fits on a whiteboard where code does not. SO 2012b: text-defined component diagrams exist (PlantUML, yUML). These agree with Brown 2021 on drift and with Jongeling 2025.
  • SESE 2015f: functional programmers design with type signatures instead of UML (one answer of 27 votes; example of practice).