Graphics & Perception

Defines a declarative JSON grammar of graphics that separates data series, encodings, scales, and threshold annotations from the rendering backend.

Four-level model of chart descriptions (encoding, statistics, trends, domain context); readers value levels 2-3 most. A per-card screen-reader description, a table view and keyboard-navigable marks cover levels 1-3; a takeaway title covers 3-4.

Replicates and extends Cleveland & McGill's graphical perception hierarchy on web displays, quantifying log-error across position, length, angle, area, and horizon-graph encodings.

Cleveland-McGill ranking (position on common scale > length > angle/area > colour) justifies the dot/bar forms used throughout the Bundle Insights gallery; the gallery mostly honours it, but adds colour flags on point estimates (fs_infra 'more than twice the fleet rate') that ignore the drawn CI.

HOPs help untrained readers judge trend reliability.

Shows that chart titles strongly frame what viewers recall and conclude, often more than the data. Relevant to dashboard panel titles: a title that states the scope ("P1 within target") without the exclusion lets readers generalise to all urgent tickets. Titles should carry the question and the scope.

Incentivised experiment: CDF and quantile-dotplot displays of a waiting-time distribution beat point estimates and intervals for threshold decisions. Directly applicable to SLA feasibility: an empirical CDF of P1 response minutes with vertical lines at 5/15/30 answers "what % within X" for any X. It does this better than a bar chart of four percentiles.

Crowdsourced color-difference thresholds measured for points, bars and lines show that perceptible difference depends strongly on mark type and size: thin lines and small points need much larger color distances than large fields. A palette validator that applies one fixed deltaE threshold regardless of mark (e.g.

Draco formalizes visualization design knowledge as hard and soft constraints, enabling automated checking of chart specs. A gallery can carry a small hard-constraint set (values in domain, CI contains estimate, title numbers equal data numbers) checked at export time.

Controlled experiments showing truncated axes, area-as-quantity and aspect-ratio tricks measurably change reader judgments. Supports treating a silently out-of-range value (a 250% share, a point estimate outside its own interval) as a misleading-visualization defect, not a cosmetic one.

Visual Information-Seeking Mantra: Overview first, zoom and filter, then details-on-demand; foundational for dual-persona diagnostic interfaces.

Munzner's what-why-how: match idiom to task (lookup, compare, rank, trend). Several gallery bar lists are ranking tasks left unsorted, or mix dimensions and units on one axis (fs_infra, pipeline, fleet, closure), which violates expressiveness.

Correction to my earlier perspective: the abstract says the authors analysed visualization-accessibility papers from the last ~20 years; I did not verify the exact count (56) I stated earlier, so treat that number as unverified. The design-space framing (users, tasks, media) and the gap in interactive/complex-chart access stand.

Azimuth: dashboard-level screen-reader access needs an overview, consistent per-chart summaries and navigation between views. Supports short L1/L2 chart descriptions, named disclosures and a role filter that announces what changed.

Cliff effect around p=0.05 persists across CI, gradient and violin displays; gradient CIs reduce it somewhat. Significance-only colouring in a gallery should carry a redundant continuous cue.

For error signatures vs escalation three cautions apply: use exact binomial limits at small n; adjust for overdispersion (multiplicative or random-effects) when many units fall outside limits, which signals systematic confounding (incident size, customer escalation propensity) rather than many true outliers;

Chartability (POUR+CAF, 50 heuristics) applied to the gallery: passes keyboard navigation of marks, a table alternative and text contrast; fails a bypass link, label-in-name on the theme button, and concise descriptions (aria-describedby median 525 chars, duplicated on article and chart). Touch users get no tooltips.

Survey of uncertainty-vis evaluation: readers misjudge intervals; evaluations should test decisions, not only perception. The gallery draws Wilson CIs widely, but several rate cards (fs_bands, fs_leader) show point values only, and colour flags point estimates whose CIs span the reference.

Synthesis of what works in visual communication. Direct labels beat legends; redundant encoding helps colour-deficient readers; a takeaway title or annotation focuses attention; uncertainty should be shown; dual axes and unaligned comparisons mislead. Used as the general benchmark for the support suite's text, colour and annotation practice.

Two experiments: readers shown 95% confidence intervals overestimate treatment effects and pay more than readers shown 95% prediction intervals; axis rescaling helps less than prediction intervals or animated HOPs, and inferential-uncertainty displays make readers underestimate outcome variability.

Relative-distance judgments over four single-hue and five multi-hue quantitative colormaps: perceptually-uniform multi-hue ramps (viridis family) performed at least as well as single-hue luminance ramps, rainbow-type maps were worse, and combining perceptual-space distance with color-naming predicted performance better than either alone.

Replicates Cleveland & McGill's position/length/angle/area comparison with Bayesian multilevel models and shows the canonical encoding ranking is an average-observer artefact: many individuals deviate from it, so 'position beats area' is a population prior, not a law.

Padilla et al.: choose uncertainty displays for the decision; intervals and frequency framings beat point-only, and deterministic construal errors are common. In the gallery, evidence-grade chips and 'lower bound' wording help, but CI-free bars invite deterministic reading.

VisEval: NL2VIS benchmark of 2,524 queries over 146 databases with automated checkers for validity (code runs, chart renders), legality (data/chart type correct for the query) and readability (layout, overlap, scales).

Visualization rhetoric: framing choices (what is coloured, titled, omitted) steer interpretation. A page that colours significance on some cards and categories on others creates conflicting rhetorical codes; one colour contract is needed.

Analyzes eye-tracking fixations and thousands of free-text recall descriptions across 393 visualizations. Proves that readers fixate first on the title and direct data labels, and that message redundancy (reinforcing the core quantitative contrast in the title, direct bar callouts, and takeaway banner) directly increases accurate recall without…

Background colour changes inferred colour-quantity mappings only when a colormap appears to vary in opacity: without apparent opacity variation people infer dark-is-more; with it, opaque-is-more. On light backgrounds the biases agree; on dark backgrounds they conflict, so the 'right' direction of a sequential ramp in dark mode is ambiguous.

Review of control charts in health surveillance: distinguishes Phase I (establishing a stable baseline) from Phase II monitoring, warns about overdispersion, autocorrelation and the need for risk adjustment, and favours CUSUM/EWMA for small persistent shifts.

Tufte: maximise data-ink, use small multiples on shared scales, keep graphical integrity (proportional encodings, labelled units). Applied: gallery footers (30-35% of card height) and repeated Era/Evidence text are non-data ink at page level; shared-axis small multiples (bias, coverage, audit) are exemplary.

HOPs outperform error bars for judging how reliable an ordering is. Relevant to ranking cards (fs_infra, sigrel) where readers will rank groups whose intervals overlap heavily.

gallery council 9 (dataviz/communication), 2026-10-28

Argues presentation (storytelling) is a distinct visualization goal from exploration: guided sequences, annotation and a stated message. Used in the bundle-insights gallery for takeaway-sentence titles and stakeholder-ordered sections.

Readers prefer charts with substantial integrated text; annotation placement drives takeaways. Supports short on-chart captions over long hover-only panel descriptions.

Explains, from the US Census Bureau's experience, how many consistent published aggregates can be solved jointly to reconstruct underlying records, motivating formal privacy (differential privacy) for 2020. Operational lesson: each extra cross-tab or margin raises reconstruction risk even when every cell passes a threshold rule;

Interviews with visualization designers: accessibility is skipped under time pressure and for lack of tooling. Operationally: put a11y checks (reflow at 320 px, focus-not-obscured, colour-only encodings) into an automated pre-publish script rather than relying on review rounds.

VizLinter encodes visualization design rules (expressiveness, effectiveness, misleading-scale rules) as lint checks over Vega-Lite specs and proposes minimal fixes.

Evaluates 2,070 single-panel visualizations across 261 participants and establishes that intrinsic visualization memorability is consistent across observers (Spearman rank correlation r=0.75).

Physiologically-based stage-theory CVD model that handles anomalous trichromacy and dichromacy in one framework via a severity parameter; matrices for protan/deutan at severity 1.0 are the de-facto standard in colour tooling (colour-science, DaltonLens).

Controlled study of dual-scale charts. Readers make more errors and spurious cross-series comparisons when two quantities share an area with different scales. A separate aligned strip for n is safer.

Analysed 378 responsive visualization pairs from news sites and catalogued the transformation strategies designers use between desktop and mobile (rescale, reposition, transpose axes, simplify/aggregate, remove or move labels and annotations, change encoding).

Interview study of dashboard users: they rarely stop at the dashboard. They export, annotate and add context and narrative to communicate upward ("data voice"). Dashboards that lack explanation, context and story push users into side spreadsheets and slides. Supports adding narrative and annotation panels, and "so-what" text, on executive views.

gallery council 9 (dataviz/communication), 2026-10-28

UpSet: matrix plus bar layout for set intersections, scaling past Venn/Euler. Rationale for not using a Sankey for audit class to sub-group, where cases belong to many sub-groups (overlapping memberships, not a flow partition).

Few: a dashboard is a single-screen, at-a-glance display of the few measures needed to act; failures include exceeding one screen, burying the key measure, excessive detail, and decoration. Applied to Bundle Insights: the 49-card page is a report, not a dashboard; it needs a one-screen executive layer above it.

KMunicate: survey of 1,100+ stakeholders on Kaplan-Meier presentation; recommends confidence bands and a numbers-at-risk table under the curve. Applied to the ticket-to-bundle and bundle-to-analysis curves for business readers.

Surveys 83 dashboards and builds a typology by purpose (strategic, operational, organizational, learning), audience, update cadence and visual features. Main point: "dashboard" spans very different genres.