Annotated research bibliography grounding Buena Mono's design decisions · ← all reports
Generated against Buena Mono v1.228 · last updated 2026-07-26
An annotated bibliography of academic work relevant to Buena Mono: the science of legibility and reading, monospaced/code typography, syntax coloring, typefaces as an interface between humans and machines (OCR → LLMs and AI agents), AI models that understand and generate type, and typography in human-in-the-loop AI writing workflows.
Compiled 2026-07-07 for Buena Mono v1.220; last verified against v1.228 (2026-07-22). Entries are ordered oldest→newest within each section. Quotations are verbatim from the cited work or its abstract.
Frutiger, Adrian (1967). "OCR-B: A Standardized Character for Optical Recognition." Journal of Typographic Research (now Visible Language), 1(2). journals.uc.edu/index.php/vl/article/view/4992 Frutiger's own account of designing OCR-B for Monotype — a monospaced typeface engineered so machines could read it and humans could tolerate it (accepted as world standard, ISO 1073/II, in 1973). The founding document of "fonts designed for machine readers," and a direct ancestor of every disambiguated-glyph decision in modern coding fonts (0 vs O, 1 vs l vs I).
Arditi, Aries & Cho, Jianna (2005). "Serifs and font legibility." Vision Research, 45(23), 2926–2933. Classic controlled study isolating serifs from other typeface variables; found at most small legibility effects of serifs at acuity-limited sizes. Grounds the choice of a sans (neo-grotesque) skeleton for screen-first type.
Beier, Sofie & Larson, Kevin (2010). "Design Improvements for Frequently Misrecognized Letters." Information Design Journal, 18(2), 118–137. ingentaconnect.com Tested letter variants within the same typeface using distance and short-exposure methods: "narrow letters benefit from being widened, and x-height characters benefit from using more of the ascending and descending area"; a double-storey a is misread far less than a single-storey one. Directly applicable to glyph disambiguation in a code font.
Legge, Gordon E. & Bigelow, Charles A. (2011). "Does print size matter for reading? A review of findings from vision science and typography." Journal of Vision, 11(5):8. doi:10.1167/11.5.8 A vision scientist and a type designer reconcile psychophysics with typographic practice; defines the "fluent print size" range within which reading speed is maximal. Reference point for UI/default-size recommendations.
Beier, Sofie & Larson, Kevin (2013). "How does typeface familiarity affect reading performance and reader preference?" Information Design Journal, 20(1), 16–31. PDF Separates familiarity effects from intrinsic letterform legibility — relevant when a new mono departs from the shapes developers have internalized from Menlo/Consolas.
Rello, Luz; Pielot, Martin & Marcos, Mari-Carmen (2016). "Make It Big! The Effect of Font Size and Line Spacing on Online Readability." Proc. CHI 2016, 89–96. PDF Eye-tracking study, 104 participants, Wikipedia articles at 10–26pt and line spacings 0.8–1.8: "up to a font size of 18 points, subjective and objective readability as well as comprehension improved continuously" — while line-spacing effects were comparatively weak. The strongest size-specific evidence behind Buena Mono's recommended settings (size matters more than leading; err larger).
Sawyer, Ben D., et al. (2020). "The great typography bake-off: comparing legibility at-a-glance." Ergonomics, 63(4). doi:10.1080/00140139.2020.1714748 MIT AgeLab-lineage glance-legibility research: measured how typeface, weight, size, case, and polarity change the milliseconds needed to correctly read a word at a glance. Found "positive polarity text had significantly lower legibility thresholds compared to negative polarity text" and that perceived size (x-height) is a crucial mediator — relevant to terminal/dark-theme reading and status-line UI.
Bylinskii, Zoya, et al. (2021). "Readability Research: An Interdisciplinary Approach." arXiv:2107.09615 Agenda-setting survey uniting vision science, typography, and HCI; the best single overview of what is and isn't known about on-screen readability.
Wallace, Shaun; Bylinskii, Zoya; Dobres, Jonathan; Kerr, Bernard; et al. (2022). "Towards Individuated Reading Experiences: Different Fonts Increase Reading Speed for Different Individuals." ACM Transactions on Computer-Human Interaction, 29(4), Article 38. doi:10.1145/3502222 352 participants, 16 fonts: matching font to reader yielded a 35% average increase in reading speed with comprehension maintained — and "personal preference for fonts did not necessarily predict reading speed." The strongest modern argument for variable fonts and per-user typographic tuning (wght/slnt axes as personalization, not decoration).
Rello, Luz & Baeza-Yates, Ricardo (2013). "Good Fonts for Dyslexia." Proc. ASSETS '13 (15th ACM SIGACCESS Conference on Computers and Accessibility). Eye-tracking study (48 readers with dyslexia): font choice measurably changes reading performance for dyslexic readers; recommends sans-serif, roman, monospaced styles.
Rello, Luz & Baeza-Yates, Ricardo (2016). "The Effect of Font Type on Screen Readability by People with Dyslexia." ACM Transactions on Accessible Computing, 8(4). PDF Extended study, 97 subjects (48 with dyslexia), 12 fonts: "sans serif, monospaced, and roman font styles significantly improved reading performance over serif, proportional, and italic fonts." The strongest peer-reviewed evidence that monospaced type carries an accessibility benefit for a substantial group of readers.
McChesney, Ian & Bond, Raymond (2019). "Eye tracking analysis of computer program comprehension in programmers with dyslexia." Empirical Software Engineering, 24, 1109–1154. doi:10.1007/s10664-018-9649-y Found that programmers with dyslexia "do not exhibit over-proportional deficiencies compared to non-dyslexic individuals" when reading code (unlike prose) — suggesting code's spatial, columnar layout (a monospace property) changes the reading task itself.
McChesney, Ian & Bond, Raymond (2021). "Eye Tracking Analysis of Code Layout, Crowding and Dyslexia — An Open Data Set." Proc. ETRA '21. doi:10.1145/3448018.3457420 Open eye-tracking dataset on code layout and visual crowding — the raw material for future studies of letter spacing and line length in code fonts.
GitHub Next / Lettermatic (2023). "Monaspace: texture healing." monaspace.githubnext.com (industry research, not peer-reviewed)
Introduced "texture healing" — locally redistributing widths inside a fixed grid to fix monospace's worst texture problems (cramped m, gappy i). The most significant recent design-research contribution to monospace typography; no independent empirical evaluation published yet.
Butterick, Matthew. "Ligatures in programming fonts: hell no." Practical Typography. practicaltypography.com (essay, counterpoint) The canonical argument against coding ligatures (ambiguity, semantic distortion). Included deliberately: no peer-reviewed study on coding ligatures and comprehension exists as of this compilation (see §6 Research gaps) — vendor claims and this counterargument are the current state of the debate.
Sarkar, Advait (2015). "The impact of syntax colouring on program comprehension." Proc. PPIG 2015 (26th Annual Conference of the Psychology of Programming Interest Group). Eye-tracking study: syntax highlighting significantly reduced comprehension time for novices, with the benefit shrinking as experience grows. Foundational evidence behind the "Show Syntax" / colored-code presentation on the Buena Mono site.
Hannebauer, Christoph; Hesenius, Marc & Gruhn, Volker (2018). "Does syntax highlighting help programming novices?" Empirical Software Engineering, 23. Larger replication (n≈390): found no significant effect of syntax highlighting on novices' program comprehension — a caution against overclaiming; color schemes are a UX preference with mixed evidence, not settled science.
(How typefaces function when the reader is a model — the modern descendant of Frutiger's OCR-B problem, and the section most relevant to "fonts for LLMs / AI agents.")
Goh, Gabriel; Cammarata, Nick; Voss, Chelsea; Carter, Shan; Petrov, Michael; Schubert, Ludwig; Radford, Alec & Olah, Chris (2021). "Multimodal Neurons in Artificial Neural Networks." Distill. doi:10.23915/distill.00030 Discovered that CLIP contains neurons responding to concepts whether drawn, photographed, or written as text — and thereby introduced the "typographic attack": a handwritten label reading "iPod" stuck on an apple flips the model's classification. The paper that made rendered text a first-class attack surface for AI.
Noever, David & Noever, Samantha E. M. (2021). "Reading Isn't Believing: Adversarial Attacks On Multi-Modal Neurons." arXiv:2103.10480 Early systematic exploration of text-in-image attacks on CLIP-class models.
Cheng, Hao; Xiao, Erjia; Gu, Jindong; et al. (2024). "Unveiling Typographic Deceptions: Insights of the Typographic Vulnerability in Large Vision-Language Models." ECCV 2024. arXiv:2402.19150 Quantifies how typographic factors — font size, opacity, color, position of rendered text — modulate attack strength on LVLMs; their mitigation dataset reduces the impact by 28%. Literally: measurements of which typographic parameters make text more or less "loud" to a model.
Cheng, Hao; Xiao, Erjia; et al. (2025). "Exploring Typographic Visual Prompts Injection Threats in Cross-Modality Generation Models." arXiv:2503.11519 Extends typographic prompt injection to image-to-image generation pipelines — typed words in an input image steer generative agents.
Westerhoff, Justus; et al. (2025). "SCAM: A Real-World Typographic Robustness Evaluation for Multimodal Foundation Models." arXiv:2504.04893 Large real-world benchmark (photographed text, not just synthetic overlays) showing current VLMs "remain susceptible" to typographic attacks.
Waseda, Futa; et al. (2025). "Read or Ignore? A Unified Benchmark for Typographic-Attack Robustness and Text Recognition in Vision-Language Models." arXiv:2512.11899 Frames the core tension for AI agents operating on screens: the model must read legitimate text (UI labels, code) yet ignore adversarial text. Directly relevant to computer-use agents reading terminals and editors set in monospace type.
Hufe, Lorenz; Venhoff, Constantin; et al. (2026). "Dyslexify: A Mechanistic Defense Against Typographic Attacks in CLIP." arXiv:2508.20570 Locates the "typographic circuit" — specific attention heads that read rendered text — and ablates it, trading text-reading for robustness.
Chen, Tianle & Ghadiyaram, Deepti (2026). "A Systematic Study of Cross-Modal Typographic Attacks on Audio-Visual Reasoning." arXiv:2604.03995 Typography as attack vector generalizes across modalities (audio-visual LLMs).
Liu, Bohan; et al. (2026). "Towards Robustness against Typographic Attack with Training-free Concept Localization." arXiv:2607.02494 Training-free interpretability defense; state of the art in the typographic-robustness arms race as of mid-2026.
Wang, Zhangyang; et al. (2015). "DeepFont: Identify Your Font from An Image." ACM Multimedia 2015. arXiv:1507.03196 The first large-scale CNN font-recognition system (Adobe) — the start of machine perception of typefaces.
Shirani, Amirreza; Dernoncourt, Franck; et al. (2020). "Let Me Choose: From Verbal Context to Font Selection." ACL 2020. arXiv:2005.01151 Models that pick a fitting font for a text's meaning — machine judgment of typographic voice (cf. Buena Mono's "Calm / Precise / Warm / Editorial" presentation tags).
Wang, Yizhi; et al. (2020). "Attribute2Font: Creating Fonts You Want From Attributes." ACM Transactions on Graphics / SIGGRAPH 2020. Generates fonts from human-readable attributes (weight, contrast, warmth) — the generative counterpart of fontquant-style measurement.
He, Haibin; et al. (2022). "Diff-Font: Diffusion Model for Robust One-Shot Font Generation." arXiv:2212.05895 One-shot font generation via diffusion; representative of the stroke-aware generation line of work.
Iluz, Shir; et al. (2023). "Word-As-Image for Semantic Typography." SIGGRAPH 2023. Semantically deforms glyph outlines while preserving legibility — AI operating directly on vector letterforms.
Yang, Yukang; et al. (2023). "GlyphControl: Glyph Conditional Control for Visual Text Generation." NeurIPS 2023. arXiv:2305.18259 Glyph-conditioned control for rendering correct text in diffusion images — treating letterforms as structural constraints, not textures.
Thamizharasan, Vikas; et al. (2023). "VecFusion: Vector Font Generation with Diffusion." CVPR 2024. arXiv:2312.10540 Cascaded raster→vector diffusion producing usable Bézier control points — machine-drawn outlines in the same medium type designers work in.
Shimoda, Wataru; et al. (2024). "Type-R: Automatically Retouching Typos for Text-to-Image Generation." arXiv:2411.18159 Post-hoc correction of rendered-text errors in generated images.
Koo, Myungkyu; et al. (2025). "FontAdapter: Instant Font Adaptation in Visual Text Generation." arXiv:2506.05843 On-the-fly adoption of unseen fonts in generative models.
Huang, Xiaoke; et al. (2026). "VecGlypher: Unified Vector Glyph Generation with Language Models." arXiv:2602.21461 An LLM that emits SVG path tokens for high-fidelity vector glyphs — as of 2026, language models draw type directly. The clearest sign that "fonts for LLMs" now runs in both directions: models read type, and models make type.
(Supports the site's "Human & machine, in color" section: typographically distinguishing borrowed machine words from one's own.)
Kim, Jiho; et al. (2024). "Towards Full Authorship with AI: Supporting Revision with AI-Generated Views." arXiv:2403.01055 Interface research on keeping the human author in control while LLMs contribute — the design space Buena Mono's provenance-coloring idea lives in.
Russell, Jenna; Karpinska, Marzena & Iyyer, Mohit (2025). "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text." arXiv:2501.15654 Experienced readers can spot machine prose via lexical clues — but only with effort; visual/typographic provenance marking (color, style) is the interface-level answer to the same need.
Mysore, Sheshera; et al. (2025). "Prototypical Human-AI Collaboration Behaviors from LLM-Assisted Writing in the Wild." arXiv:2505.16023 Empirical taxonomy of how people actually interleave their own words with LLM output — the behavioral evidence that mixed human/machine text is now the default writing condition a "writer-first" typeface serves.
calt, would be genuinely novel.