The word cloud that pays by the letter

Dark navy consulting slide, masthead “Beverage & Spirits Practice · Regulatory Diagnostic · Working Draft”, headlined “What the wine rulebook actually talks about”. A subhead reads: every word of 27 CFR Part 4 — Labeling and Advertising of Wine — 22,434 words — counted; common English words removed, nothing else touched; each word is sized so that the area it occupies is proportional to how often it appears; no sampling, no weighting, no editorial judgement. Filling most of the slide, under the label “Word frequency · 27 CFR Part 4 · complete text · top 40 terms”, is a word cloud in gold, blue and grey. “wine” is far the largest word at the centre. Around it, “statement”, “designation”, “information”, “appropriate”, “appellation”, “paragraph” and “statements” are drawn large, while “ttb”, “atf”, “use” and “fruit” are among the smallest words in the picture. Down the right-hand side, a gold-bordered card reads 22,434 words over the line “The complete regulation, counted end to end — nothing sampled, nothing set aside”. A card headed “Method” says font size is set to the square root of the count, so a word’s area rises in proportion to how often it appears, and that every word is horizontal, with no rotation, no shape mask, no hand-placement, and decorative colours. A gold card headed “What leaps out” lists statement 2nd largest, designation 3rd largest, information 5th largest, appropriate 6th largest. A card headed “Read-out for the client” concludes that the drafting emphasis sits on designations, statements and information — the vocabulary of a disclosure regime — and that the agency’s own names, TTB and ATF, barely register, so procedure is not where this rulebook spends its attention. Three ticked boxes run along the bottom: “No axis to truncate — there is no axis, no baseline and no scale break anywhere on this chart”, “Nothing cherry-picked — the whole part, every word of it, in one picture, not a window, not a sample”, and “No colour scale to gerrymander — hue carries no meaning here; only size encodes the data”. A footnote records that the word counts were computed by us from the full XML text of the part as served by the eCFR API for the issue date of 18 August 2026, gives the tokenising rules and the totals of 22,434 tokens, 11,822 after removals and 2,108 distinct, and states that the slide is a demonstration drawn by us in the manner of a consulting deck and was never published anywhere. — Showing the misleading chart

The same forty words on a cream background, kicked “The same forty words · counted, then measured · 27 CFR Part 4” and headlined “The word cloud pays by the letter.” A subhead explains that every word was sized so that its area would match how often it appears, but that the shapes the cloud actually drew track how often a word appears times how many letters it has — r = 0.99 against that product, 0.89 against the count on its own — and that “TTB” appears 126 times and “designation” 74, yet the cloud draws “designation” 2.6 times bigger. On the left, a slopegraph headed “Rank by frequency → rank by area on the page” lists all forty words in frequency order down one side, each annotated with its count and letter count, and in area order down the other, each annotated with its measured footprint in square pixels; lines connect them and cross heavily. Gold lines mark words that rose in the picture, blue lines words that fell, and heavy lines those that moved ten places or more: “atf” at 87 occurrences and three letters drops from 8th to 37th, “ttb” at 126 drops from 4th to 28th, while “statements” at 44 occurrences and ten letters climbs from 35th to 10th and “appellation” at 43 climbs from 36th to 12th. In the middle, a zero-based bar chart headed “What the counts actually look like · the fourteen most common words” shows wine at 515, far ahead of name 151, statement 136, ttb 126, label 120, type 104, part 102, atf 87, class 80, health 77, alcohol 74, designation 74, percent 73 and used 72, with words of eight letters or more picked out in gold. Beneath it, a strip headed “Four of those words, at the size the cloud gave them” draws four tinted rectangles at the cloud’s own scale: ttb, 126 times, 3 letters, 9,676 px²; designation, 74 times, 11 letters, 24,971 px²; atf, 87 times, 3 letters, 6,395 px²; information, 66 times, 11 letters, 21,830 px² — the long words visibly occupying far more of the page than the more frequent short ones. A gold-ruled callout reads: a word’s footprint is its em square repeated once per letter, so sizing the square by the count makes the footprint letters × count — only one of those two is the data, the other is orthography. Five cards run down the right. “The cloud used the charitable rule” notes that font size was set to the square root of the count, the choice that intends area to be proportional to frequency, and that it still fails because the area a font size governs is one em square and a word is as many of those as it has letters; the other rule, size set straight to the count, makes the footprint letters × count squared. “What moved, and by how much” lists the four biggest movers and records that two words out of forty kept their place, twelve moved ten places or more, and rank agreement between the orderings is ρ = 0.53. “And the leader is undersold” notes that wine appears 3.8 times as often as statement, 515 against 136, and covers only 1.8 times as much of the page — four letters against nine — because the mechanism inflates long words and deflates short ones at the same time. “The knobs you cannot see” covers the stop-word list, which here dropped shall, may and must, and the layout seed, rotation and colour. “The check, in ten seconds” says to find the longest word in any cloud and look up its count, then the shortest word still readable, and look up that one. A footnote gives the eCFR source and issue date, the tokenising rules and totals, states that the footprints are advance width times font size measured in the browser that drew the cloud at font size 176 × √(count ÷ 515) in Liberation Sans Bold, notes that the correlations and rank statistic were computed by us with words tied on the count ranked alphabetically, and says both drawings are our own. — Showing the misleading chart

Be the first to star this exhibit

Every word of the federal wine-labelling regulation, counted, then drawn as a word cloud with each word sized so its area would match how often it appears. Measure the shapes it actually drew and they track frequency times letter count instead: “TTB” occurs 126 times and “designation” 74, and the cloud draws “designation” two and a half times bigger. Of forty words, two kept their place.

Title 27 CFR Part 4, Labeling and Advertising of Wine — Alcohol and Tobacco Tax and Trade Bureau, Department of the Treasury, via the eCFR API, issue date 18 August 2026, 2026Added August 20, 2026

01The claim

Here is what the federal wine-labelling rulebook is really about. Every word of 27 CFR Part 4 — 22,434 of them — has been counted, common English words taken out, and the rest drawn as a word cloud. And the slide can say something most charts cannot: there is no axis on it to truncate, no baseline to move, no window to crop, no colour scale to gerrymander. The whole part is in the picture, every word of it, and hue carries no meaning at all — only size encodes the data. Better still, the size rule is the careful one: font size set to the square root of the count, so that each word’s area, not its height, rises in proportion to how often it appears. Read the picture and the finding is plain. “Statement” is the second-largest word, “designation” the third, “information” the fifth, “appropriate” the sixth. This is the vocabulary of a disclosure regime — a rulebook preoccupied with what has to be stated. Meanwhile the agency’s own names, TTB and ATF, sit among the smallest words on the slide. Procedure, apparently, is not where this regulation spends its attention.

02The trick

The counts are right, the text is complete, and the sizing rule is the better of the two in common use. The trouble is that the rule can only reach one of the two dimensions the reader is measuring. Setting font size to the square root of the count controls the em square — the box one character sits in. But a word is not one character. Its footprint on the page is that square laid down once per letter, so the area the eye actually takes in comes out as letters × count. Length is doing half the work, and length is not data about anything; it is spelling. Measure the forty rectangles the cloud drew and they correlate at r = 0.99 with letters × count, and only 0.89 with the count on its own. What that does to the ranking is not subtle. “TTB” appears 126 times, the fourth-commonest word in the part; three letters put it 28th by area. “ATF” appears 87 times, eighth-commonest; it lands 37th of 40, the fourth-smallest thing on the slide. Coming the other way, “statements” appears 44 times — 35th — and lands 10th; “appellation” appears 43 times, 36th, and lands 12th. Of the forty words, two kept their place, twelve moved ten places or more, and rank agreement between the two orderings comes to ρ = 0.53. So the read-out is backwards on its own evidence. The rulebook says “TTB” and “ATF” 213 times between them and says “designation” 74 times; the picture reports the reverse, because one set of words is short and the other is long. The same arithmetic quietly flattens the leader: “wine” occurs 3.8 times as often as “statement”, 515 against 136, and covers only 1.8 times as much of the page. Everything else on the slide is true, and none of it helps — there is indeed no axis to truncate, and that is precisely the difficulty, because with no axis there is nothing to check the sizes against. (This slide is our own demonstration, drawn in the manner of a consulting deck. The word counts are real, computed from the regulation’s published text.)

03The fix

Put the counts on a zero-based axis, sorted, one length per number, and the picture stops arguing with itself: “wine” at 515 with everything else in a long flat tail — which is what free text almost always looks like, and what a cloud can never show, because it has no axis on which a tail could be flat. The slopegraph is worth drawing once for the education: rank the words by frequency down one side, by measured area down the other, join them up, and watch the three-letter words dive while the eleven-letter words climb. Then print the numbers, because a cloud’s central problem is that it makes a quantitative claim and supplies nothing to check it with. Say which stop-word list you used, too. Ours removed “shall”, “may” and “must” as ordinary English — in a regulation those are the three words that do the work, and taking them out is an editorial act, not a technical one. “Shall” alone occurs 170 times in this part, more often than any word left in the picture except “wine”; the list deleted the second-biggest word in the cloud before the cloud was drawn. Rotation, the layout seed and the colours are the same kind of undeclared choice: rotate a word and it reads smaller, re-run the layout and the picture rearranges, and the hues here mean nothing at all, which readers reliably refuse to believe. The habit worth keeping is a ten-second check that works on any cloud you meet. Find the longest word in it and look up its count. Then find the shortest word you can still read, and look up that one. If the picture has been ranking by letters, those two numbers will tell you straight away. And if the real question is which words travel together rather than how often each appears, it was never a frequency question — reach for phrases, co-occurrence, or a dozen quoted responses. Where a cloud is wanted as ornament, let it be ornament, sitting beside the chart that carries the numbers rather than standing in for it.