Tufte Integration Of Modes

Words, Numbers, Images — Together

Overview

Evidence bearing on any question of complexity arrives in several forms at once: prose, measurements, pictures, diagrams, tables. Tufte's argument in Beautiful Evidence ch. 3 (printed pp. 82–121) is that the separation of these forms is an artifact of production convenience, not a property of the evidence — and that the separation is actively enforced by publishing technology, software architecture, and design convention, usually against the intentions of the author and always against the interests of the reader.

His framing claim: evidence is evidence, whether it arrives as words, numbers, images, diagrams, still or moving. It is all information. For readers and viewers, the intellectual task is constant regardless of mode — understand the material, reason about it, and appraise its quality, relevance, and integrity. Nothing in that task licenses sorting evidence into a text section, a figures section, and an appendix.

The chapter's evidentiary spine is a 300-year natural experiment: Newton's Opticks, whose figures were severed from its text in 1704 and stayed severed across 23 editions in 5 languages. If a work of that stature can be rendered substantially incoherent by reproduction technology for three centuries, no work with visual elements is safe.

Scope of this skill: document architecture, figure and table placement, caption policy, cross-reference elimination, and the decision of whether a report's structure serves the evidence or the toolchain.

Naming convention. Failure-mode labels here (Mode Segregation by Production, Book Operating System, Deferred Integration) are descriptive teaching names, not Tufte's own terminology, except where explicitly quoted or page-cited. "Book Operating System" is his phrase (p. 114).

§1. The Baseline: Evidence Is Multimodal Before Anyone Designs It

Tufte opens with measurements rather than assertion (p. 83).

The point of the comparison is not that more pictures are better. It is that the ratio is set by the subject, and that the older, technologically poorer books integrate more completely than the modern journal corpus does. Production capability has improved; integration has not.

Why most display techniques are inherently multimodal. Statistical graphics and maps are visual-numerical fields labeled with words and framed by numbers. Even an austere image evokes other images, remembered narrative, and a sense of scale. Words simultaneously carry semantic and visual content — the nouns on a map both name places and locate them in the two-space of latitude and longitude. The modes are not adjacent; they are interpenetrated. Segregating them requires active effort against the material.

The historical illustration (p. 85). A 9th-century Carolingian manuscript (Harley 647, Aratus, folio 12) draws the constellation Centaurus as a word-picture: Latin text in Rustic script forms wavy tracks whose curvature models the horse's flank and rear. Words begin and end so as to outline the legs and body; the resulting perimeter links 44 star-dots into the southern constellation, while the text itself mentions 24 stars. About 1,320 characters do the work of image, label, and data simultaneously. Tufte's observation on it: the Latin is incomprehensible to nearly every modern reader, yet the visual meaning reads immediately — the universality of images against the stupefying locality of languages.

Leonardo as the working standard (pp. 82–85). Leonardo's anatomical study of the arm combines text, drawings, engineering diagrams, annotation of local detail, and scales of measurement — whatever it takes — to show a subtlety like the slight contraction in arm length when the hand turns flat and the radius crosses the ulna. Tufte's rhetorical question is the right one to carry: how could it be otherwise? In manuscripts the common source of all marks is one hand holding one pen, so integration is the default and segregation would require deliberate work.

§2. The Mechanism: Segregation Is Manufactured, Not Chosen

The bond between verbal and nonverbal evidence comes undone in publishing, as the assorted technologies of reproduction segregate information by the accident of its mode of production (p. 83).

When printing integrated by default (p. 87). For the first 800 years of printing with movable reusable type — from the 600s in Korea, Japan, and China to the 1400s in Europe — text and image were produced by the same method. Information to be reproduced was cut in relief in mirror-image reverse on a woodblock; whatever was to print stood proud, was inked, and was pressed onto paper. Because the method is indifferent to whether a mark is a letter or a picture, text and image print simultaneously and land integrated on the sheet without special intervention. The Diamond Sutra (868), the earliest complete dated printed book, shows the result.

Then the processes diverged. Once high-resolution images required a different press pass from type — engraving and intaglio rather than relief — integration became expensive and segregation became free. Every later failure in this chapter descends from that split.

Galileo paid the cost personally (p. 97). Sidereus Nuncius (1610) presents its evidence in 78 images and drawings, tightly integrated with the explanatory text, covering 30% of the printed area of the book's 60 pages. To achieve it, Galileo personally financed the copper engravings, and each sheet had to pass through two different presses — relief for the type, intaglio for the high-resolution engravings. The complexity had consequences: some copies of The Starry Messenger went out with the engravings simply omitted, and auction houses 395 years later still note which copies lack them.

The stakes of that integration. The book's 5 engraved moon images (4 unique plus 1 repeat) were the first astronomical pictures ever printed, and roughly 10 pages of visual observation demolished the Aristotelian doctrine that celestial bodies are perfect, smooth, and unblemished — a claim that had become religious doctrine on no evidence at all. From then on, theories about the universe had to be tested against the visual evidence of observation.

What the integration bought — visible certainty (pp. 100–109). The book's engraved star fields spill past the text block into the margins — the composition breaks the typographic grid because, as Tufte puts it, nature is nowhere rectangular, and the new stars were breaking the frame of past knowledge. Galileo's own warrant for the whole performance is oculata certitudine — certainty delivered to the eye by direct observation. The core evidence is a 24-page serial diary of 65 dated, annotated observations of Jupiter's satellites, integrated with the running narrative (the first printing of about 550 copies sold out immediately). Tufte then re-arranges that evidence twice: once as stacked one-dimensional linelands that expose the satellites' corkscrew orbital structure in space-time, and once as a compressed parallel layout of words beside images. His point is methodological: both arrangements are needed to analyze the evidence, and re-showing the same data in several arrangements is repetition of data, not redundant design — scientific displays should not be afraid to repeat their evidence with visual variations. In journals, letters, and the printed book alike, Galileo draws no line between words, diagrams, and images; a book design that treats all modes of information alike re-enacts the argument.

The endpoint of the argument — Saturn as a sentence element (EI pp. 120–121). Envisioning Information closes its Epilogue on Galileo's 1613 Istoria, where the telescopic images of Saturn are set inline in the running sentence — word and drawing as one, the stunning never-before-seen images just another sentence element. Tufte's coda names Saturn at once a drawing, a word, and a noun: "The wonderful becomes familiar and the familiar wonderful" (EI p. 121). This 1613 inline glyph is the EI seed of Beautiful Evidence's opening argument, and the tightest historical demonstration that mode integration can reach all the way down into the syntax of a sentence.

The rule this yields: ask of any document what its production method makes cheap. Wherever the tooling makes segregation free and integration expensive, segregation will happen regardless of what the author intended — so it must be resisted deliberately, budgeted for, and checked at proof stage.

§3. Newton's Opticks: 300 Years of Segregation as a Natural Experiment

This is the chapter's central evidence and the strongest argument available for the whole thesis. It works because it holds the content constant — the same words, the same 55 figures — while the production decisions vary across 23 editions and 300 years.

The control condition: Newton's Principia (1687), p. 110

The Principia places its hundreds of physics diagrams properly within the text. It even solves the hardest case: where one diagram is needed across a long argument, the first edition repeats the same diagram in position on 4 double-page layouts, and the third edition on 7 consecutive layouts. Descartes' Principia (1644) used the same technique with 11 repeats.

Two causes, both worth naming because they are reproducible:

1. Production method. The book was printed in relief, so words and woodcut images printed simultaneously in one pass through the press.
2. A content-oriented producer. Edmund Halley — an intensely visual scientist and cartographer — financed publication and supervised book production alongside Newton. Their surviving letters (29 June and 14 July 1686) show Halley proposing woodcuts set into the page with the demonstrations, as more convenient for the reader and not much more expensive, and Newton agreeing.

The lesson: integration in 1687 required a specific person with specific authority caring about it. It has never happened by default.

The failure: Opticks (1704), pp. 110–111

Opticks reports fundamental discoveries — interference effects, the color composition of sunlight, Newton's rings, the reflector telescope — and one-third of Newton's text directly addresses its 55 diagrams. Neither Halley nor woodcuts were deployed. The diagrams were printed as free-standing engravings, fully severed from the text they explain.

The resulting collation:

55 drawings, ganged onto 12 pages of flaps bound into 4 clusters, each cluster opening sideways and dangling from the book.

The navigation tax. Linking a passage to its distant diagram takes 5 hierarchical levels of flap label code: the phrase "Prism DEGdeg" in the text resolves to the flap address "Book I. Part II. Plate IV. Fig. 16." The first 80 text pages alone contain 6,300 such letter-codes referring to illustrations stashed away after page 80. Reading it means constant shuttling between one volume of words and what amount to 4 parallel illustration booklets dangling off it.

Tufte's verdict, and the sentence to carry: this is a competent design for comparing or admiring diagrams, and a poor design for learning about light and optics. Two different jobs; the format silently chose the wrong one.

The end-state of segregation. Separated plates in old books are convenient for a particular kind of nonreader: dealers in antiquarian prints slice the engravings out, frame them, and sell them off — the ultimate segregation of image from text. Appreciating physics drawings as art objects is fine; it is finer to know what they mean and who thought them up.

The wear pattern is the verdict (p. 15, adjacent "Mapped Pictures" chapter). Brisson's Ornithologie (1760) split its content into 6 fat text-only volumes (Latin and French in parallel), linked by awkward codes to a separate atlas of 261 hand-colored engravings. Tufte reports the reader behavior his own set records: the text volumes show no signs of use 244 years after publication, while the atlas of plates is well-worn. Segregation does not merely inconvenience readers — the severed half of a work simply stops being read.

The table: 23 editions, 5 languages, 300 years (pp. 112–113)

Tufte tabulates every published edition of Opticks from 1704 to 1998 and classifies each as integrated or segregated. Only 2 of the 23 completely integrate Newton's images with his words.

The failures are not repetitions of one mistake — later editors kept inventing new ways to segregate:

The two successes:

• First German edition (Leipzig, 1898), 2 volumes — completely integrated. All figures correctly placed with their accompanying text, the text often fitted around the figures. The first edition in any language to combine Newton's words and images: 194 years late. It was 19th-century lithography, allowing text and image to be printed simultaneously on the same page, that finally made it possible.
• Reprint of the second English edition of 1717 (Chicago, 1952) — completely integrated. Published as volume 34 of The Great Books of the Western World. The first time Newton's words and images came together in English: 248 years after publication.

A London 1931 reprint (still in print as of 2006 and the only in-print edition for reading Newton, with an Einstein introduction) is only partly integrated: 28 figures placed with their text, the remaining 27 turned sideways on separate pages 1 to 5 pages away.

What to take from it

• Nearly every editor, translator, and publisher seems simply never to have registered how visual and explanatory the content is. The failure was not malice or cost — it was not noticing that the figures were the argument.
• The initial book design tends to persist, whether it is good (Hypnerotomachia, §4) or bad (Opticks). Getting the original right is worth disproportionate effort, because three centuries of reprints will inherit it.
• The generalization, stated plainly: if Newton's great work can be disrupted and rendered substantially incoherent by the technologies of reproduction for some 300 years, then no work with visual elements is safe.

Descartes, the same story (p. 114, fn. 18). The beautifully integrated text and image of the first edition of Descartes' Principia philosophiae (1644) — illustrations incorporated into the main body of the Latin text — was promptly dismantled by the bureaucracies of secondary production. The 1647 French edition prepared by Abbé Picot appended the illustrations to the end of the text; the canonical Adam and Tannery edition of Descartes' works followed Picot's practice, and its editors were consequently compelled to add awkward linking footnotes tying pictures back to the right bits of text. Brian Baigrie's finding is the decisive one: not one reference to the illustrations is offered that is not already in Descartes's original text. The author had done the integration. Editors undid it, then hand-built a worse mechanism to simulate what they had destroyed.

§4. The Positive Cases: What Integration Looks Like When It Holds

Hypnerotomachia Poliphili (1499) — measured integration, pp. 88–94

Tufte does something unusual here: he measures the integration rather than praising it. He read the English text surrounding each of the 170 woodcuts and collated it against the original 1499 edition.

Tufte's term for what this achieves: a cognitive integration alongside the celebrated optical one. The likely cause is the same as Newton's Principia — image-sketches placed within the original manuscript, so the compositor was laying out an already-integrated document.

Three further devices from the same book:

• Storyboard summary (p. 92). Six woodcuts thread through a 40-page love story, compressing a notably wordy account into a crisp visual digest — setting and linking scenes, locating and animating characters. Tufte compares the structure directly to a modern movie storyboard (the cornfield sequence in Hitchcock's North by Northwest, 1959).
• Words as image (p. 93). 57 concocted Latin and Greek inscriptions (2,300 words in English translation) appear on pots, boxes, tablets, architecture, and a garden of letterforms built from marjoram plants. The book is 94% text and 6% images, yet the words themselves are intensely visual.
• Persistence (p. 94). For 500 years, across the 1546 French edition and the 1999 English translation, the integration held — because the original design was right. The reverse of Opticks.

But — the warning attached to the same book (p. 88). Hypnerotomachia has been studied for centuries by design students as an exemplar of printing, layout, and typography, and only that, since almost no one can read its confected Italian-Latin-Greek hybrid. Unreadable classics teach that some books are beautiful as pure art objects; unfortunately they also teach content indifference — the lesson that in text/image relations a designer need only consider aesthetic effects.

The test to carry from this section: for serious work in reasoning about evidence, the essential test of text/image relations is how well they assist understanding of the content, not how perfectly stylish the pages look. — Beautiful Evidence, p. 88

Applying that test requires knowing what the words mean in relation to the images, and what the images mean in relation to the words — which is a content judgment, not a layout judgment, and cannot be delegated to someone who has not read the material.

The sentencing memorandum — integration under adversarial stakes (pp. 94–95)

A short modern case sits right after Hypnerotomachia: a federal prosecutors' sentencing memorandum from a 2006 congressional bribery case, which sets its documentary exhibit — the handwritten bribe schedule itself — directly into the argument's text. Tufte's two observations generalize: the standard format of American legal documents produces thin information densities (leaded-out type, as if productivity were measured by the page), and adroit placement of visual evidence inside the prose measurably intensifies the advocacy. Integration is not an aesthetic nicety; in adversarial writing it is a competitive weapon.

The Golden Field Guide — the task dictates the apparatus (p. 115)

Tufte permits metaphors and analogies to assist reasoning about analytical design, with one binding constraint: the metaphors should come from the content and the reasoning associated with the content — not from computer interfaces.

The worked case is the jay spread from Robbins, Bruun, and Zim's Birds of North America (1966), illustrated by Arthur Singer. The purpose of a field guide is to combine visual recognition of physical objects with new verbal, geographic, and graphical information, and every element is derived from that purpose:

The location maps are the sharpest demonstration, because they prevent a specific, nameable error: confidently identifying a bird in Montreal that in fact resides exclusively in Acapulco. Tufte's verdict on the spread: assembling all the good evidence in one place is precisely what produces its felt craft, detail, and credibility.

The operational rule: derive the apparatus from the task's known failure modes. "Add a tooltip / a legend / an expandable panel" are interface metaphors and answer no question the content posed.

Jaegerman — the designer-as-researcher standard (pp. 116–117)

Megan Jaegerman's "Spotting a Hidden Handgun" (parts in The New York Times, 26 May 1992; expanded and redrawn for this book) choreographs movement in 3-space through sequences, call-outs, motion arrows, mappings, and multiple viewpoints — silhouette, 3-dimensional figure, flatland footprints — with words and images working together to describe an extended causal sequence.

Tufte's production claim is the one that generalizes: she did both the research and the design, breaking their common alienation. The design amplifies the content because the designer created the content. He then reproduces her backup documentation — verification of Detective Robert T. Gallagher's identity and record against a public database, notes from the interviews, his 18 years with the NYPD and roughly 1,200 gun arrests, and the legal sequence from suspicious demeanor to probable cause to frisk to arrest that determines which tell-tale signs matter — as evidence of the care such work requires.

Why it belongs in a skill about integration: integration is not a layout operation performed after the content exists. Placing the right words next to the right image requires knowing which words bear on which part of the image — a judgment only the person who did the reporting can make reliably. Split research from design and you get a document that must be re-integrated by cross-reference codes, which is where Opticks ended up.

§5. The Modern Failure: The Book Operating System (p. 114)

The contemporary equivalent of the 1704 flap-clusters, and the diagnostic that names it.

Tufte's specimen is a designer field guide to the weather whose physical collation is:

176 pages   black-only type + a few diagrams, on bible paper
 96 pages   color pictures, each with a few words, on glossy stock
212 pages   black-only type + a few diagrams, on bible paper

Icon-symbol thumbtabs then link the segregated words and images back together. Tufte names the resulting apparatus a Book Operating System, and it requires 570 words of documentation telling readers how to use the book.

The sample procedure he reproduces sends a reader chasing 3 links across 3 separate clusters of 6, 24, and 12 pages to answer one question: turn to the thumb tab guide, find the vortex-shaped silhouette, follow it to color plates 207–230, read a caption that refers you to pages 511–522, and only then reach the text.

He also notes what the computer-interface metaphor drags in with it: a presentation style emulating an inept technical manual — second person cloying from overuse (11 uses of you and your in 9 sentences), lame humor, a tone implying the reader is an idiot, clunky typography with double spaces after periods, and narrow unhyphenated columns producing a hyperactive right rag. Like a technical manual, the book's actual subject — the weather — is not mentioned until the last word of its own title.

The diagnostic, stated as a rule: when a document requires instructions for its own navigation, mode segregation has already won. The 570 words are not a helpful feature; they are the receipt.

How to apply it. Count the navigation instructions in your own artifact — "see the appendix," "refer to the dashboard," "full data in the companion deck," onboarding tours explaining where things live. Each is evidence that content the reader needs in one place has been stored in another for production reasons. The remedy is never a better index; it is co-location.

§6. The Cleveland Debate: Labels Are Data, and Clutter Is a Design Failure (pp. 118–121)

The chapter's hardest case, because the opposing argument is strong and Tufte concedes it before resolving it. This is the model for how to handle a genuine conflict between integration and legibility.

The challenge

William Cleveland (The Elements of Graphing Data, 1985, p. 46) holds that word-labels on data may interfere with assessment of the overall pattern of quantitative data, and prescribes: do not allow data labels in the data region to interfere with the quantitative data or clutter the graph; keep notes, keys, and markers outside the data region; put notes in the legend or in the text (pp. 44–47).

The test case is a noisy, cluttered scatterplot of body mass against brain mass for 26 animals (from Carl Sagan's The Dragons of Eden, after Jerison), every dot labeled with its animal's name.

Tufte's five counter-arguments

1. Labels are data — often intriguing data. Only by seeing dots linked to names does the finding emerge that the really big prehistoric animals — tyrannosaurus, brachiosaurus, diplodocus, stegosaurus — have relatively small brains. And only naming the dots exposes plotting errors: why is the hummingbird shown as heavier than the mole, the wolf heavier than humans? Where would a gnat, mosquito, cat, hammerhead shark, or centaur appear? Just like numbers, nouns are evidence.
2. Abandoning labels forces a code. When labels leave the data points, a key is needed to relink names to numbers — and such codes, keys, and legends are impediments to learning that make the reader's brow furrow.
3. The split has no substantive basis. Segregating nouns from data-dots breaks up evidence on the basis of verbal versus nonverbal, a distinction lacking substantive relevance. The separation is uncartographic, and contradicting the methods of map design causes trouble for any graphical display.
4. It moves the display the wrong way. Design strategies that reduce data-resolution take evidence displays in the wrong direction.
5. What clutter? Even the supposedly cluttered graph shows the main ideas plainly: brain and body mass are roughly linear in logarithms, and the linearity loosens as both variables increase.

The concession — and why it matters methodologically

Tufte then does the thing most arguments skip: he refuses to settle a visual question verbally — design disputes are decided by looking at visual evidence, not by argument. He grants that Cleveland has a strong point, and tests it by deleting all 26 labels.

The stripped version reveals something the argument alone could not: without the dark typography of the labels, we see very differently — the big blob of words in the top half of the original inflated the visual variability of body mass for heavier animals. Cleveland was right about the effect.

Interim practical finding: one workable solution for label-clutter, especially in exploratory analysis, is to examine both scatterplots — with and without labels.

The resolution: layering, not deletion (pp. 120–121)

Good design can dispose of clutter and show all the data points and their names. The repair has two moves:

1. The data-dots gather onto a different visual level from their labels — in the revised graph, red clusters the 26 dots.
2. The labels calm down into a quiet, grayed-down field.

The result: label clutter has vanished, but the labels are still there.

The governing formulation: like good maps, statistical graphics should have a layered depth of reading — not a hierarchy of importance of verbal over quantitative information, but a pluralism of distinctions. — Beautiful Evidence, p. 119

The same move on a second chart (p. 120). A Roger Hayward chart of glass viscosity against temperature (in John Strong, Procedures in Experimental Physics, 1938) has an overall sameness of texture and color: administrative elements — frames, grids, pointer lines, tick marks — are as visually active as the evidence curve itself. The redesign pulls the curve out of the graphic debris in red while maintaining a unity of text and linework, by the cartographic strategy of layering and separation. Note what is not done: nothing is deleted, and no annotation is moved out of the field.

The three redesign principles (p. 121)

Stated as the sequence of three graphs — original, labels-stripped, redesigned — with the first two struck through by a red X:

1. Clutter is a failure of design, not an attribute of information.
2. Visual problems should not be fixed by reducing content-resolution (for example, by discarding the words that label data).
3. Instead, fix the design.

The last step (p. 121). Words and data-dots are themselves abstracted representations of actual animals and actual masses. In the spirit of seeking visual solutions to visual problems, the final version lets each animal represent itself at its own two-space location — the mark becomes data point, label, and illustration at once. Integration taken to its limit: not words placed near images, but a single element serving every mode. (Tufte states the honest limit on the graph itself: the animal images are sized to the space available, not to mass — only the brachiosaurus-versus-human pairing approximates true relative scale.)

§7. Do / Don't

§8. Named Failure Modes

MODE SEGREGATION BY PRODUCTION

What happens: Words, numbers, and images are separated because the production process handles them differently — separate press passes, separate paper stocks, separate applications, separate teams — and the separation is then rationalized as a design decision. Tell: The document's structure maps onto the toolchain's structure rather than onto the argument's structure. Fix: Name the production constraint out loud, then decide whether to pay for integration or change the argument. Never let the default stand unexamined. (Opticks, pp. 110–113)

THE DREADED LINKAGE CODE

What happens: Text and image, separated, are reconnected by an addressing system — "Book I. Part II. Plate IV. Fig. 16," "see Figure 3," "per Eq. 12." The codes multiply: 6,300 in the first 80 pages of Opticks. Tell: The reader must hold a place with a finger, or keep a second window open, to read one argument. Fix: Co-location. Repeat the figure rather than address it. (p. 111)

THE BOOK OPERATING SYSTEM

What happens: Segregation grows severe enough to need its own navigation apparatus — thumbtabs, icon systems, cross-cluster procedures — plus documentation explaining the apparatus (570 words, in Tufte's specimen). Tell: The artifact contains instructions for its own use. Fix: Treat the instructions as the diagnosis, not the remedy; restructure so they become unnecessary. (p. 114)

DEFERRED INTEGRATION

What happens: An author integrates correctly; a later editor, publisher, translator, or platform undoes it, then hand-builds linking footnotes to simulate what was destroyed. Tell: The linking apparatus contains no information not already present in the original text — Baigrie's finding on Descartes. Fix: Retain control of layout through reproduction, and check every reissue, translation, and format migration against the original placement. (p. 114, fn. 18)

CONTENT INDIFFERENCE

What happens: Text/image relations are judged on aesthetic effect alone, because the judge cannot or does not read the content. Beautiful, unreadable books teach this as a norm. Tell: The layout could not be evaluated differently by someone who understood the subject. Fix: Apply the p. 88 test — does this arrangement assist understanding of this content? It requires a reader who knows the material. (p. 88)

RESOLUTION REDUCTION AS A CLUTTER FIX

What happens: A crowded display is improved by removing content — labels, annotations, series, categories — rather than by redesigning it. Tell: The "cleaner" version answers fewer questions than the messy one. Fix: The three redesign principles: clutter is a design failure; do not fix visual problems by cutting content-resolution; fix the design, using layering and separation. (p. 121)

RESEARCH/DESIGN ALIENATION

What happens: The person who gathered the evidence and the person who arranged it are different people with no shared understanding, so placement decisions are made on visual grounds. Tell: Annotations mark what looks important rather than what is diagnostic; captions restate rather than explain. Fix: The Jaegerman standard — the designer does the research, or the two work as one unit with shared backup documentation. (pp. 116–117)

§9. Application Checklist

Run before shipping any document that mixes modes.

Placement

• Every figure, table, and diagram sits within one eyespan of the text that discusses it
• No figure is more than a page-turn from its argument; none is in an appendix, plates section, or companion file
• Figures needed at several points are repeated, not cross-referenced
• The document survives its own format migrations (print, PDF, web, mobile) with placement intact

Linkage

• Count the cross-reference codes ("see Fig. N," "per §X," "in the appendix"). Target: zero
• Count the words of documentation explaining how to navigate the artifact. Target: zero
• No thumbtab, icon key, or index exists to reconnect parts that could have been co-located

Labels and layering

• Data points are named directly, not coded to a legend
• Where labels crowd the data, they are layered down (grayed, set on a separate visual level), not deleted
• Both labeled and unlabeled versions were examined during exploratory work
• Administrative elements — frames, grids, tick marks, pointer lines — are visually quieter than the evidence

Provenance of the design

• The apparatus is derived from the content's own task and failure modes, not from an interface pattern library
• The person who did the research participated in the placement decisions
• Backup documentation exists and could be shown: sources, interviews, verifications
• Someone who understands the subject has judged whether the arrangement assists understanding of it

Production

• The production method's cheap path has been identified, and where it favors segregation, the cost of integration is budgeted
• Proofs have been checked for placement drift introduced downstream of the author

Source book: Beautiful Evidence (Cheshire, Connecticut: Graphics Press, 2006) Relevant chapter: "Words, Numbers, Images — Together," printed pp. 82–121. Closely related: "Mapped Pictures" (pp. 12–45) for annotation on images; "The Fundamental Principles of Analytical Design" (pp. 122–139), whose Principle 4 states the integration requirement in one line.