The invisible seam
Part 1 of The Proof, a series about typesetting’s history and the checks that proof this site.
A proof is a visual contract. It gives maker and reviewer something to judge together.
In printing, it was the invisible seam between designer and printer. They could look at the same set page, mark what was wrong, and agree on the correction. The reader saw the result, not that exchange.
The proof tested whether the page we could make matched the one we meant to make.
That agreement is part of the design work. Maker and reviewer need to decide what success looks like, and the set page gives them something specific to discuss. This is where a team’s design culture takes shape: how people give feedback, resolve a disagreement, and learn from the correction.
When a proofreader doubted a word or a fact, they underlined it and wrote “Qy.” in the margin. The author had to answer every query. (Britannica, 1911; Smith, 1929.) Leave a query when the answer belongs to someone else.
On the web, a different window width, font, theme, or browser setting can change what the next reader sees, so the page I approve is only one rendering. Reviewing that page can’t cover every form it will take. The words still need to carry the writer’s meaning and voice.
The reviewer who refused
In 1976, Donald Knuth revised the second volume of The Art of Computer Programming. His publisher had sold its Monotype machine, so the galleys came back phototypeset, and they “looked awful.” Retuned fonts were “much improved from the first attempt, but still unacceptable.” He listed what was still wrong: “The ‘N’ in ‘NOAM’ was tipped; the ‘ff’ in ‘effect’ was much too dark; the letters ‘ip’ in ‘multiple’ were too close together; and so on.”
In his 1996 Kyoto lecture, he said: “I had spent 15 years writing those books, but if they were going to look awful I didn’t want to write any more.”
The proof gave him something specific to refuse. The words could be correct while the type fell short of what he and his publisher intended.
Counting the loss
The galleys failed on type. In a different publication, Knuth and Michael Plass counted bad line breaks.
In their 1981 paper, they checked the first 1,000 line breaks in two volumes of the Journal of the ACM, 20 years apart. In 1960, set by a Monotype operator, they found 13 bad ones. In 1980, set by a leading computer system, they found 55.
They called the count “admittedly subjective,” and the journal’s style had changed, so the comparison wasn’t conclusive. To them, the increase suggested that a machine choosing breaks by fit alone could lose something of the meaning.
To the reader, the second word after a bad break “often seems like an anticlimax, or isolated from its context.” They likened the disruption to turning a page between “Chapter” and “8.” They described rules for keeping meaningful groups of words together.
Someone on site at the press
This site pulled its own bad galleys on a development branch. On the Arbor case study, the headline set on three lines at 767 pixels wide and on four at 768, where an indent took two of the twelve columns out of its measure. The extra line persisted below 800 pixels.
Every gate passed.
A sweep of six pages during an audit found it. The existing checks could accept the type and the layout without seeing what happened when the two met. The missing check was the relationship between them.
Example sources
The indent belonged in the design. Its timing was wrong. The fourth line was a side effect, not a decision.
That’s drift: a change in the result that nobody intended. Moving the indent to a wider breakpoint gave the headline its room back. Before making type smaller to rescue a layout, check whether the layout has taken room away from the words.
Example sources
The sweep became a gate: check:wrap. It resizes the site’s pages from 360 to 2,560 pixels and fails the build if a headline gains a line as the window grows. The reader’s text-size setting stays fixed for that comparison. Enlarging the type is a different test, and may need more lines.
Here’s the rule, from scripts/check-wrap.mjs at commit 7cc6bb5:
// Growing the viewport may cost a headline a line — it must never buy one.The same comment puts it plainly: “more room, never more lines.”
I don’t need to prescribe a line count for every headline to hold that relationship. The words, face, and composition can change. The check only asks whether a wider window has made the headline taller. That leaves room to design while keeping one unintended result out.
This figure measures independent example text at widths you choose, and needs JavaScript.
The audit exposed the defect; I chose the rule. The gate repeats the check when copy, fonts, or breakpoints change. Start with the reader’s problem, then choose what to measure.
Knowing I can measure that result gives me more confidence to experiment with the typography. I can change the words or try another treatment and check whether I’ve taken their room away. If I can’t measure what matters, working out why teaches me something about the medium.
What it can’t see yet
Two layouts can have the same line count and read very differently. Knuth and Plass counted breaks that disrupted meaning; my gate counts lines.
It also visits widths in 10px steps, so a narrow failure can fall between them. It checks that body paragraphs use pretty or balanced wrapping, but it doesn’t read their breaks.
I want to make more of that judgment explicit: which breaks weaken a sentence, which shapes make a paragraph harder to follow, and which of those problems a tool can reliably catch. A pass should tell us what was checked and leave the remaining questions visible.
Someone still has to read the page and judge whether the words carry their meaning. A failure needs judgment too. The layout may need correcting, or an intentional design decision may have outgrown the rule. Maker and reviewer need the rendered page in front of them to decide.
This series looks at earlier proofs for decisions worth carrying forward: how to make the words legible, keep their meaning clear, and accommodate the person reading now.
First, the ruled manuscript page. Then the right edge of hand-set type, the bad breaks of hot metal, the sizes phototype lost, and the web’s changing page. Part 4 will try to count bad breaks here.
The page keeps changing. The contract has to hold.