TRISTAN CHIAPPISI · Field Report · No. 1 · January 2026
I · The Default Lie II · What Tufte Built III · What Few Built IV · The Next Generation V · The 2026 Notebook Three Rules
Field Report · No. 1 · 2026 Published 2026-01-03 · Columbus, Ohio

After the Pie Chart.

A field guide in two halves. Eight before/after pairs grounded in Cleveland, Tufte, and Few, then seven newer charts coming out of the 2025–2026 visualization community that the canonical books were written too early to include.

By Tristan Chiappisi Columbus 8 pairs + 7 modern charts · 19 min read Connect on LinkedIn →
Same data · Different chart · Different argument
From the Editor

The chart you reach for is, almost always, an argument you didn't realize you were making.

I have been making charts for a living for more than a decade. I have made them well and I have made them badly. The single most-recurring observation across that decade is that the default chart is wrong, not occasionally, not in edge cases, but as a base rate. The pie chart your spreadsheet drew the moment you clicked the icon is, in almost every case I have ever seen, the wrong chart for the data underneath it. So is the dual-axis line graph the analyst dropped into the board deck. So is the boxplot in the engineering postmortem. So is the choropleth map in the policy brief. The defaults are wrong, and they are wrong for perceptual reasons that Cleveland, Tufte, and Few already wrote down decades ago.

This issue is not an opinion piece. It is a field guide. Eight chart pairs. The same data, drawn the way the default would draw it, and the way the research says you should. The "before" charts are not strawmen. They are the charts that get made every day in dashboards, decks, and press releases. The "after" charts are the upgrades that, for almost every audience and almost every dataset, make the actual point of the data legible.

Three names show up over and over below. William Cleveland ran the experimental work in 1984 that ranked the perceptual tasks underneath every chart: position more accurate than length, length more accurate than angle, angle more accurate than area. Edward Tufte built the vocabulary (sparkline, small multiple, slope graph) and the data-ink ratio principle that governs any honest chart drawn since. Stephen Few took both bodies of work into the boardroom and wrote out, in plain language, what they meant for your dashboard. Most of what follows is borrowed from one of the three.

I came across Tufte and Few about ten years ago, after I had already spent years in data, and finding them is one of the biggest single shifts in my career I can name. I had been moving the numbers around for a long time at that point. What I had not yet learned was how to take those numbers and turn them into a chart that actually told the story they were trying to tell. Tufte and Few were the answer. Most of what I know about communicating with data, I learned by reading and re-reading the two of them.

Every chart in this issue is hand-drawn in pure SVG. Every "before" is a chart I have, depressingly often, seen in production. Every "after" cites the source that proved it works. If you make charts for a living, or for a deck that decides anything, this is the field guide I wish someone had handed me in 2013.

Tristan Chiappisi, ed. · Inaugural Edition
Six Numbers, One Thesis
1984
Year Cleveland & McGill published their experimental ranking of perceptual tasks. Position on a common scale is the most accurate. Length is next. Angle, the task pie charts ask of the reader, is well behind both.
Cleveland & McGill · JASA 79(387)
1983
Year Edward Tufte introduced the data-ink ratio, the slope graph, and most of the vocabulary that still governs any honest chart drawn since. The Visual Display of Quantitative Information, still in print.
Tufte · Graphics Press
2005
Year Stephen Few published the bullet graph design specification, a deliberate replacement for the dashboard speedometer. Twenty years later, most dashboards still ship the speedometer.
Few · Perceptual Edge
2006
Year Tufte coined and specified the sparkline, a word-sized graphic that compresses a time series into the flow of running text. Twenty years later, most dashboards still don't use them.
Tufte · Beautiful Evidence
2019
Year Allen et al. introduced the raincloud plot, a peer-reviewed replacement for the boxplot that shows the raw data, the density, and the summary statistics in one frame. Open source, three languages.
Wellcome Open Research · 4:63
1
Number of Y-axes a single chart should have. The dual-axis line is the most defensible-looking lie in the modern dashboard. The fix takes two charts, not one.
Few · Now You See It, 2009

I.

What the defaults get wrong

The Default Lie. Two charts you should never draw, and what to draw instead.

The two most common chart types in business dashboards are also the two best-documented offenders against perceptual research. Both are easy to fix. Neither is fixed by accident.

A.

The pie chart. The chart that asks you to compare angles.

Cleveland & McGill (1984) ran the foundational experiment on graphical perception. Their ranking of elementary perceptual tasks placed position on a common scale at the top, length and angle further down, and area further down still. A pie chart asks the reader to compare angles. A Cleveland dot plot, drawn from the same numbers, asks them to compare positions. Same numbers. Different ask. Different chart.

SOURCE: Cleveland, William S., and Robert McGill. Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods. Journal of the American Statistical Association, 79(387):531–554, 1984. Data: illustrative cloud-infrastructure market share, Q4 2025. Numbers approximated from publicly reported quarterly figures and Synergy Research Group estimates of the long tail.
Read both charts. Find the third- and fourth-largest providers. On the dot plot, you read it instantly. On the pie, you cannot, because you cannot reliably distinguish a 12% slice from an 8% slice from a 4% slice without going to the legend, and the legend is in a separate region of the page entirely. This is not a stylistic preference. It is a documented perceptual fact.
B.

The dual-axis line. The chart that lets you draw any correlation you like.

When you put two metrics on a single chart with two different y-axes, you choose the scale of each independently. You can make any pair of variables appear to track, diverge, or cross at whatever moment is convenient. Stephen Few has been arguing against this chart since 2008. The replacement: a connected scatterplot, where one variable is the x-axis, the other is the y-axis, and time becomes a path through the plane. There is only one frame of reference, and the analyst no longer gets to pick it.

SOURCE: Few, Stephen. Dual-Scaled Axes in Graphs: Are They Ever the Best Solution? Perceptual Edge Visual Business Intelligence Newsletter, March 2008. Data: illustrative SaaS company headcount and revenue, 2018–2025, modeled on representative growth curves of late-stage venture-funded software companies.
The connected scatter says exactly one thing: from 2018 to 2022, this company added headcount and revenue at roughly the same rate. From 2022 forward, headcount kept climbing while revenue per head started to flatten. The dual-axis chart could be used to argue almost the opposite. Honest charts have one frame of reference. The dual-axis line has two, and the analyst chooses both.
Headline finding
The chart is an argument before it is a decoration. The default chart is the argument the software wanted to make, not the one the data actually supports.

II.

The compact, honest graphic

What Tufte Built. Two graphics most software still won't draw for you.

Tufte's career is a long argument that the highest-quality chart is the one with the highest data-to-ink ratio you can get away with. Two of the techniques he formalized, small multiples (1990) and sparklines (2006), still get solved worse by most dashboards in 2026.

C.

Six lines on one axis. The spaghetti chart.

A spaghetti chart is what happens when you ask a default chart library to compare a small group over time: half a dozen colored lines on a single axis, all overlapping, all fighting for the reader's attention. The reader cannot follow any single line without losing the others, and cannot compare any pair without working harder than the chart should make them work. Small multiples, the term Tufte formalized in Envisioning Information (1990), solve this by giving each series its own axis on a shared scale.

SOURCE: Tufte, Edward R. Envisioning Information. Graphics Press, 1990. Section: "Small Multiples." Data: illustrative monthly indexed performance of six large-cap technology stocks, January 2024 to December 2025, normalized to 100 at start. Values are constructed to be representative, not to identify any specific security.
The spaghetti version is a chart of one variable: color. The reader's job is to decode which line is which, then trace it. The small-multiples version lets the reader compare six trajectories at a glance because each one occupies its own panel: the same operation, six times, on a shared scale. Same data. The argument is now visible.
D.

The table you would have written anyway, with a sparkline per row.

Tables are a fine chart. Most reports underuse them. What most reports also do not do, even though Tufte specified it in Beautiful Evidence (2006), is add a sparkline: a word-sized graphic that compresses the trajectory behind the number into the line of the table. The sparkline does not replace the number. It explains it. The 2026 dashboard that ships with sparklines in every row is the dashboard the analyst trusts.

SOURCE: Tufte, Edward R. Beautiful Evidence. Graphics Press, 2006. Chapter 2: "Sparklines: Intense, Simple, Word-Sized Graphics." Data: top 8 U.S. metropolitan statistical areas by 2024 population (Census Bureau ACS 5-year), median household income (Census ACS), illustrative trailing 12-month population change.
A sparkline next to a number tells the reader two things at once: the latest value, and whether the latest value is the result of a steady trend, a recent spike, or a recovery from a trough. A bare table tells the reader only the first. The cost of upgrading from one to the other is roughly fifty pixels per row.

III.

The dashboard, drawn correctly

What Few Built. Two upgrades for the chart types that own your dashboard.

Stephen Few's contribution is operational. He took Tufte's principles and turned them into specifications a working dashboard designer could implement on Monday morning. The bullet graph and the strip plot are two of the most useful pieces of that work.

E.

The speedometer gauge. The chart that takes 200 pixels to tell you one number.

Few's 2005 design specification for the bullet graph is the most underused dashboard component in the working world. It replaces a circular gauge (eight square inches of decoration around a single needle) with a horizontal bar that shows the current value, the target, the qualitative ranges (poor / satisfactory / good), and the prior period, in roughly one-fifth the space and with several times the information. There is no good argument for the gauge in 2026. There has not been one since Few wrote the spec.

SOURCE: Few, Stephen. Bullet Graph Design Specification. Perceptual Edge, October 2013 (originally proposed 2005). perceptualedge.com/articles/misc/Bullet_Graph_Design_Spec.pdf. Data: illustrative quarterly KPIs for a representative B2B software company.
Count the bits of information in each version. A gauge gives you one: the current reading. A bullet graph gives you four: actual, target, qualitative range, and (with a marker) prior period. Same horizontal real estate. Four times the data density. Few specified this in 2005, and most dashboards have still not noticed.
F.

The boxplot. The chart that hides the distribution it is supposed to show.

A boxplot summarizes a distribution into five numbers: minimum, lower quartile, median, upper quartile, maximum. That is fine when the distribution is well-behaved. It is dangerously misleading when the distribution is bimodal, skewed, or has a fat tail, because two visibly different distributions can produce identical boxplots. The raincloud plot (Allen et al., peer-reviewed in Wellcome Open Research, 2019 / v2 2021) fixes this by adding back the raw data and the density curve. Same five numbers, plus everything the boxplot threw away.

SOURCE: Allen, M., Poggiali, D., Whitaker, K., Marshall, T. R., van Langen, J., & Kievit, R. A. Raincloud plots: a multi-platform tool for robust data visualization. Wellcome Open Research, 4:63, 2019 (v2 2021). wellcomeopenresearch.org/articles/4-63/v2. Data: illustrative API response-time samples (n=80 per service) constructed to demonstrate distributions a boxplot would render identically.
Look at "search" and "recommendations" in the boxplot. They look like the same kind of distribution. Now look at them in the raincloud. Search is bimodal: there are two latency regimes. Recommendations is right-skewed: most requests are fast, with a long tail. The boxplot summary made the two indistinguishable. The raincloud makes them obviously different. That difference is your incident.
Pull quote
If your chart can be made by clicking one button, your chart is the chart that wins arguments your software wanted to win, not the chart that wins yours.

IV.

What practitioners built on top

The Next Generation. Two charts that the working data community has slowly made standard.

The slope graph and the hex tile cartogram are two of the post-Tufte/Few chart types that the contemporary data-visualization community (not the chart libraries) has standardized. Both are simple to draw. Both are vastly clearer than what they replace.

G.

The grouped bar chart. The chart that asks you to compare twenty things by length.

When your question is "what changed between two time points across many categories," the default chart is a grouped bar: two bars per category, one per period, lined up. The reader's task is to compute, mentally, the difference between each pair of bars. The slope graph (specified by Tufte in The Visual Display of Quantitative Information in 1983, popularized by Few and Cole Nussbaumer Knaflic) is a single line per category, drawn between two y-axis positions. The reader's task becomes: look at the slope. Up means up. Down means down. The eye does the math.

SOURCE: Tufte, Edward R. The Visual Display of Quantitative Information. Graphics Press, 1983 (slope graph specification, p. 158). Knaflic, Cole Nussbaumer. Storytelling with Data. Wiley, 2015. Data: illustrative ranks of the top 12 programming languages, ordered by working-developer survey mention rate, 2020 vs 2025. Ordering modeled on the public Stack Overflow Developer Survey series.
The grouped bar chart shows you 24 numbers and asks you to compute 12 differences in your head. The slope graph shows you 12 lines and asks you to do nothing. Your visual system does the differencing for you, instantly. The same data takes roughly the same number of pixels and yields a totally different cognitive cost. That is the trade you are making every time you ship the bar chart.
H.

The state choropleth map. The chart that reports geography instead of data.

A U.S. state map shaded by some metric (the choropleth) is one of the most-shipped charts in American journalism and policy, and one of the most distorted. Wyoming and Montana take more visual real estate than Massachusetts and Connecticut combined while containing a fraction of the population. The reader's eye reads "size = importance," but the geography is uncorrelated with whatever the data is. The hex tile cartogram, used widely by NPR's visuals team, FiveThirtyEight, and the Bureau of Transportation Statistics among others, gives every state the same area. The metric becomes the only thing the eye can encode.

SOURCE: NPR Visuals team and FiveThirtyEight on hex tile cartograms. Datawrapper Academy, "When to use tilegrid maps." Harrower & Brewer, ColorBrewer.org: An Online Tool for Selecting Color Schemes for Maps, Cartographic Journal 40(1):27–37, 2003 (sequential colormap selection). Borland & Taylor, Rainbow Color Map (Still) Considered Harmful, IEEE CG&A 27(2), 2007. Data: illustrative state-level "weekly AI-tool use among adults, 2025," constructed for visual comparison and not a published metric.
On the choropleth, the empty Mountain West dominates. On the hex tile, every state is the same size, and the metric is the only signal. If your question is geographic (where is the wildfire), use a real map. If your question is statistical (which states differ on this metric), use a hex tile. Confusing the two is the most common error in policy graphics in 2026.

V.

Charts Tufte didn't have a name for

The 2026 Field Notebook. Seven charts coming out of the working community right now.

The first four parts are grounded in research that pre-dates the iPhone. This part is grounded in research and practice that mostly post-dates GPT-3. These are charts the canonical books were written too early to include. The 2025–2026 working data-visualization community has, in the last few years, made them part of the default vocabulary. Almost none of them ship in your charting library yet.

I.

The embedding scatter. The chart of the LLM era.

Take a few hundred high-dimensional vectors (sentence embeddings, image features, customer reviews, agent traces) and project them into two dimensions with UMAP (McInnes, Healy & Melville, 2018) or t-SNE (van der Maaten & Hinton, 2008). The output is a scatterplot with no interpretable axes. Distance carries the signal. Five years ago this was an ML research chart. Today every interpretability dashboard, every retrieval-augmented system, and every embedding atlas (Nomic Atlas, latentscope, Anthropic's Clio) uses it.

SOURCES: McInnes, L., Healy, J., & Melville, J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426, 2018. van der Maaten, L., & Hinton, G. Visualizing Data using t-SNE. Journal of Machine Learning Research 9:2579–2605, 2008. Anthropic, Clio: Privacy-preserving insights into real-world AI use. 2024. Data: 250 synthetic support-ticket embeddings across five hand-labeled clusters.
The axes are not labeled because they cannot be. UMAP coordinates are not interpretable in the way a regression coefficient is. What the chart shows is structure: points that are close together share semantic similarity in the original high-dimensional embedding space. The clusters are the argument. The labels are the punchline.
J.

The attention heatmap. The chart you draw to ask a model what it was looking at.

Every transformer-based language model produces, as a side effect of inference, a matrix of attention scores: for each output token, how much weight did it place on each input token? Visualizing that matrix as a heatmap (one row per output token, one column per input token, color by attention weight) is the foundational technique of mechanistic interpretability. The chart is a 2017 paper (Attention Is All You Need) made visible. It has become a working tool for prompt debugging, alignment research, and model auditing in 2024–2026.

SOURCES: Vaswani et al. Attention Is All You Need. NeurIPS 2017. arXiv:1706.03762. BertViz (Jesse Vig, ACL 2019 demos) and TransformerLens (Neel Nanda, 2022 onward) as canonical references for production attention visualization. Anthropic's transformer-circuits.pub for ongoing interpretability work. Data: synthetic attention pattern for an English-to-French translation, single attention head, illustrative and not from a real run.
Read the diagonal. Most output tokens attend strongly to their semantic counterpart in the input: chat looks at cat, noir looks at black, verts looks at green. The off-diagonal cells are where the model is doing something interesting (or wrong). Five years ago you needed a research codebase to render this chart. Today it is one line in TransformerLens.
K.

The Sankey, repurposed. The chart that finally has its decade.

Sankey diagrams have existed since 1898. The original was a steam-engine energy-efficiency diagram. They have always been the right chart for compositional flow: how much of this ended up as that. What changed in 2024–2026 is the use case. AI agent runs are flow problems: a thousand triggers fan out into a few planning strategies, fan out into a few tools, fan back into outcomes. Modern agent observability platforms (LangSmith, Datadog LLM Observability, OpenLLMetry) lean on flow diagrams to summarize what a thousand runs actually did.

SOURCES: Sankey, Matthew Henry Phineas Riall. The Thermal Efficiency of Steam-Engines. Minutes of the Proceedings of the Institution of Civil Engineers, 134, 1898. Modern application: LangSmith, Datadog LLM Observability, OpenLLMetry. Data: synthetic trace summary of 1,000 agent runs through trigger, planning strategy, tool call, and outcome.
The Sankey answers compositional questions a bar chart cannot. "Of the runs that triggered a multi-step plan, how many ended in error?" "How much of the success rate is concentrated in the search-only path?" These are the questions agent operators ask every day in 2026. The chart was waiting for the use case for 128 years.
L.

The annotated beeswarm. The strip plot the New York Times has built a house style around.

A beeswarm or jittered-strip plot shows every observation as an individual dot, packed laterally so they don't overlap, with summary statistics drawn on top. The technique is older than the term. The annotated beeswarm, with labeled outliers, quantile bands, and direct callouts to specific points, has in the last three or four years become a modern editorial style for showing distributions in the New York Times Upshot, the Pudding, the Financial Times, and Bloomberg. It is what the boxplot wishes it were.

SOURCES: Wilke, Claus O. Fundamentals of Data Visualization. O'Reilly, 2019, ch. 9 (jittered strip charts). D3 force-based collision packing, Bostock, 2014. Editorial precedent: The New York Times Upshot, The Pudding, and Financial Times graphics across 2022–2026. Data: synthetic distribution of one-week task completion times for 220 contributors on an open-source project.
This is the chart Tufte would have drawn if Tufte had had D3. Every observation is visible. The summary statistics are still drawn on top. The outliers are labeled. The eye gets the distribution shape, the central tendency, and the specific points that justify a closer look, all in one frame, with no abstraction.
M.

The horizon chart. The technique that lets you put a hundred time series on one screen.

A horizon chart compresses a time series by folding the y-axis: the area above a baseline is split into bands of increasing color intensity, and the area below the baseline is mirrored above and rendered in a contrasting hue. The result is a chart that uses a fraction of the vertical space of an ordinary line graph and stacks cleanly into a tight matrix of dozens of series. Heer, Kong & Agrawala (CHI 2009) formalized the technique and showed empirically that horizon charts preserve perceptual accuracy at much smaller chart sizes than ordinary line graphs. Modern observability tooling has been picking it up steadily ever since.

SOURCE: Heer, Jeffrey, Nicholas Kong, and Maneesh Agrawala. Sizing the Horizon: The Effects of Chart Size and Layering on the Graphical Perception of Time Series Visualizations. ACM CHI 2009. idl.cs.washington.edu/papers/horizon. Data: synthetic 60-day p95 latency series across eight production services.
The horizon chart trades a little accuracy on any single value for a lot of accuracy on cross-series comparison. When you are paged at 3 a.m. and need to know which of forty services is the offender, this is the chart you want: eight services on the screen at once, the bad ones obvious without scrolling. Few would have approved.
N.

The calendar heatmap. The chart GitHub taught the world to read.

A 365-day metric, laid out as a grid of 52 weeks by 7 days, with each cell colored by intensity. GitHub put it on every developer profile in 2013. Strava, Wakatime, Duolingo, and most modern habit trackers followed. It is now the default chart for any daily-cadence quantity over a year, and it has a property no other chart shares: the eye can read weekly periodicity, monthly clusters, and "the dip when I went on vacation" simultaneously, without any axis labels at all.

SOURCES: GitHub contribution graph, introduced 2013. d3-calendar, Mike Bostock, ObservableHQ. Cal-heatmap.js, Wan Qi Chen, 2014. Datawrapper Academy, calendar heatmap guide. Data: synthetic 365 days of "minutes spent writing" for the year 2025, with a deliberate two-week dip in late August and a productive sprint in November.
Notice you read the chart without any axis labels. Weeks are columns; days are rows; months are visible as horizontal clusters of darker cells. The chart works because the calendar is already a familiar 2D layout. That is also why it does not generalize. The calendar heatmap only works on the calendar.
O.

The railroad diagram. The chart that lets you read a grammar in one glance.

A railroad diagram, formalized by Niklaus Wirth in 1977 and made famous by the SQLite documentation, visualizes a grammar as a track. The reader follows the line from left to right. Required elements sit on the main line. Optional elements have a skip branch arcing above them. Repetitions have a loop arcing below. The chart was originally invented for programming-language reference manuals, where it remains the cleanest way to render a syntax. The 2025–2026 angle is that the same chart is now the cleanest way to render a tool-call schema, a JSON contract, a prompt template, or a regex pattern. Tools like railroad-diagrams.js (Tab Atkins) and regexper.com have made it cheap to draw one.

SOURCES: Wirth, Niklaus. What Can We Do About the Unnecessary Diversity of Notation for Syntactic Definitions? Communications of the ACM, 20(11):822–823, 1977. SQLite documentation, syntax diagrams. sqlite.org/syntaxdiagrams.html. Tab Atkins, railroad-diagrams.js, used in W3C and WHATWG specifications. github.com/tabatkins/railroad-diagrams. Regexper.com (Jeff Avallone) for regex visualization. Diagram: structure of an LLM tool-call schema, modeled on the OpenAI / Anthropic function-calling format.
Read the track from left to right and you have read the grammar. Required nodes sit on the main line. Optional nodes have a skip arc above them. The reader can see, in one glance, that name and parameters are required, description is optional, and that parameters has its own internal grammar drawn underneath. The same chart works for prompt templates, JSON schemas, and regex. Anywhere you have a grammar, this is the diagram.
Three rules to take with you

If you make charts for a living, three things to keep on the wall.

1. The chart is an argument.

The chart you draw is a claim about which dimension of the data is most important. A pie chart claims the angles matter. A dot plot claims the rank order matters. A slope graph claims the change matters. Pick the chart that makes the claim you are willing to defend. Defaults pick a claim for you.

2. Data-ink ratio is a budget.

Every gridline, every drop shadow, every legend entry, every 3D bevel, every "smooth" interpolation between points is ink the reader has to subtract before getting to the data. Tufte's rule still holds: the chart you ship should be the chart you cannot delete any further pixel from without losing meaning.

3. Direct labeling beats legends.

Few is right. The reader's eye should never have to leave the chart to find out what a color means. Label the line where the line is. Annotate the point you want them to see. The legend is a mechanical device for charts where direct labeling is impossible, and it is almost never impossible.

Sources

Every claim, every chart, where to check it.

Cleveland, William S., and Robert McGill · Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods. Journal of the American Statistical Association, 79(387):531–554, 1984. tandfonline.com/doi/10.1080/01621459.1984.10478080
Tufte, Edward R. · The Visual Display of Quantitative Information. Graphics Press, 1983; 2nd ed. 2001. Original source of the data-ink ratio, the slope graph, and the small-multiples typology. edwardtufte.com
Tufte, Edward R. · Envisioning Information. Graphics Press, 1990. Section: "Small Multiples."
Tufte, Edward R. · Beautiful Evidence. Graphics Press, 2006. Chapter 2: "Sparklines: Intense, Simple, Word-Sized Graphics."
Cleveland, William S. · The Elements of Graphing Data. Wadsworth, 1985; revised Hobart Press, 1994. Source of the modern Cleveland dot plot.
Few, Stephen · Show Me the Numbers: Designing Tables and Graphs to Enlighten. Analytics Press, 2004; 2nd ed. 2012. The single most-useful book on dashboard chart selection in print. perceptualedge.com/library
Few, Stephen · Now You See It: Simple Visualization Techniques for Quantitative Analysis. Analytics Press, 2009. Argument against dual-scaled axes, definitive treatment.
Few, Stephen · Bullet Graph Design Specification. Perceptual Edge, October 2013 (originally proposed 2005). perceptualedge.com/Bullet_Graph_Design_Spec.pdf
Allen, M., Poggiali, D., Whitaker, K., Marshall, T. R., van Langen, J., & Kievit, R. A. · Raincloud plots: a multi-platform tool for robust data visualization. Wellcome Open Research, 4:63, 2019; v2 2021. Peer-reviewed software-tool article. wellcomeopenresearch.org/articles/4-63/v2
Knaflic, Cole Nussbaumer · Storytelling with Data: A Data Visualization Guide for Business Professionals. Wiley, 2015. The contemporary practitioner's translation of the Tufte/Few canon. storytellingwithdata.com
Borland, David, and Russell M. Taylor II · Rainbow Color Map (Still) Considered Harmful. IEEE Computer Graphics and Applications, 27(2):14–17, 2007. The peer-reviewed argument against rainbow colormaps. ieeexplore.ieee.org/document/4118486
Harrower, Mark, and Cynthia A. Brewer · ColorBrewer.org: An Online Tool for Selecting Color Schemes for Maps. Cartographic Journal, 40(1):27–37, 2003. The peer-reviewed paper behind colorbrewer2.org.
van der Walt, Stéfan, and Nathaniel Smith · A Better Default Colormap for Matplotlib. Matplotlib Project, 2015. Origin of the viridis colormap family. Perceptually uniform and colorblind-safe. bids.github.io/colormap
Healy, Kieran · Data Visualization: A Practical Introduction. Princeton University Press, 2018. Free online edition. socviz.co
Wilke, Claus O. · Fundamentals of Data Visualization. O'Reilly, 2019. Free online edition. The contemporary reference for chart-type selection. clauswilke.com/dataviz
Heer, Jeffrey, and Michael Bostock · Crowdsourcing Graphical Perception: Using Mechanical Turk to Assess Visualization Design. CHI 2010. Modern replication of Cleveland & McGill's perceptual rankings. idl.cs.washington.edu
McInnes, Leland, John Healy, and James Melville · UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv:1802.03426, 2018. The dominant non-linear dimensionality-reduction technique behind embedding scatterplots. arxiv.org/abs/1802.03426
van der Maaten, Laurens, and Geoffrey Hinton · Visualizing Data using t-SNE. Journal of Machine Learning Research, 9:2579–2605, 2008. The original 2D embedding technique that defined the genre. jmlr.org/papers/v9/vandermaaten08a
Vaswani, Ashish, et al. · Attention Is All You Need. NeurIPS 2017. arXiv:1706.03762. The transformer paper that made attention heatmaps the canonical interpretability chart. arxiv.org/abs/1706.03762
Vig, Jesse · A Multiscale Visualization of Attention in the Transformer Model. ACL 2019 (System Demonstrations). The BertViz tool. github.com/jessevig/bertviz
Sankey, Matthew Henry Phineas Riall · The Thermal Efficiency of Steam-Engines. Minutes of the Proceedings of the Institution of Civil Engineers, Vol. 134, 1898. The original Sankey diagram, an energy-flow visualization for a steam engine.
Heer, Jeffrey, Nicholas Kong, and Maneesh Agrawala · Sizing the Horizon: The Effects of Chart Size and Layering on the Graphical Perception of Time Series Visualizations. ACM CHI 2009. Empirical justification for horizon charts at small chart sizes. idl.cs.washington.edu/papers/horizon
GitHub contribution graph · Introduced 2013. Cal-heatmap.js (Wan Qi Chen, 2014) and d3-calendar (Bostock) are the standard implementations. The chart that taught the world to read a calendar heatmap.
Anthropic · Clio: Privacy-preserving insights into real-world AI use. 2024. Production application of large-scale embedding-based clustering and UMAP visualization. anthropic.com/research/clio
Wirth, Niklaus · What Can We Do About the Unnecessary Diversity of Notation for Syntactic Definitions? Communications of the ACM, 20(11):822–823, 1977. The original specification of the railroad / syntax diagram.
SQLite Documentation · Syntax Diagrams. The most-cited modern application of railroad diagrams in production documentation. sqlite.org/syntaxdiagrams.html
Atkins, Tab · railroad-diagrams.js. Used in W3C and WHATWG specifications. github.com/tabatkins/railroad-diagrams
Avallone, Jeff · Regexper. Browser tool that renders any regular expression as a railroad diagram. regexper.com

About this field guide.

Compiled and published from Columbus, Ohio. I have been working in data professionally since the early 2010s. About ten years ago I came across Edward Tufte and Stephen Few, and their work has been the largest single influence on how I think about charts, and on how I tell stories with data, ever since. This issue is one long thank-you to the two of them.

The writing is set in Source Serif 4 and Inter Tight; the charts are drawn in pure SVG by hand. There are no advertisements, sponsored sections, or affiliate links. There are also no pie charts, except the one in figure A.

Fifteen charts in total. Eight before/after pairs grounded in the canonical perceptual research, plus seven newer charts coming out of the 2025–2026 working community: UMAP embedding scatters, attention heatmaps, AI-agent Sankey diagrams, annotated beeswarms, horizon charts, calendar heatmaps, and railroad diagrams. Every numerical claim is sourced. Every "before" is intentionally bad in the specific ways most production charts are bad in the wild.

Reach me on LinkedIn

Other field reports

Tristan elsewhere