A dictionary nobody could write was written anyway, by strangers, one quotation at a time. The web is waiting for the same trick.
Before 1857, an English dictionary did not try to cover English. Cawdrey's Table Alphabeticall of 1604 ran to about two and a half thousand entries, all of them hard words — foreign borrowings, terms of art, the vocabulary you might be caught out by at dinner. Ordinary words were assumed. You looked up what puzzled you, not what a word meant.
Samuel Johnson changed that in 1755, and set the ceiling at the same time. One man, nine years, forty-three thousand words, a hundred and fourteen thousand quotations. He was the first to treat citation as evidence, which was genuinely new. But the quotations illustrated definitions he had already formed, drawn from books he happened to admire, and the definitions carried his opinion. Lexicographer, he wrote: a harmless drudge. It was one man's reading, one man's library, one man's judgement — magnificent, and bounded by a single skull.
The Philological Society decided in 1857 that this was not good enough. They wanted every word, including the ordinary ones, including the obsolete and the dialect and the vulgar, and they wanted each sense shown changing over time, with dated evidence. No one had attempted it. It was immediately obvious that no one could.
James Murray became editor in 1879 and did the only thing available to him: he asked. He published an appeal to the English-speaking public — read books, and when you find a word used in a way worth recording, send us a slip.
One word. One source. One line showing it actually used, with a date. That is the whole instrument, and its modesty is the point: it asks for no expertise, no permission, and no opinion. It asks what you read and where.
Millions came. Murray sorted them into pigeonholes in a corrugated-iron shed in his garden that he called the Scriptorium. Among the most prolific contributors was W. C. Minor, an American surgeon who sent thousands of slips over two decades from Broadmoor, where he was confined after killing a man. Nobody asked his standing. They checked his citations.
It is tempting to read this as a charming story about volunteers. It was a structural fix, and it solved four problems at once.
Evidence came first. Johnson chose quotations to support definitions he had written. Murray wrote definitions from the quotations that arrived. That inversion is the difference between an opinion with footnotes and a finding.
Contribution and adjudication were separate jobs. Readers supplied evidence; editors decided what it added up to. Neither could do the other's work, and neither was asked to. Projects that blur these two get edit wars. The OED got a dictionary.
Everything carried a date. Not the date it was filed — the date it was written. That is what let the dictionary show meaning moving, and what lets you check an entry a century later against the same line.
Coverage included the unprofitable. Dialect. Obsolete senses. Words no gentleman needed. Nobody was going to fund that, so nobody had done it.
Search today is Johnson at scale. A single authority, selecting by criteria it does not publish, presenting conclusions without the evidence that produced them. The definitions are confident and the basis is private.
Coverage, meanwhile, is pre-Johnson. Hard words only maps almost exactly onto what is commercially interesting. Archives, primary sources, public-domain canon, the genuinely long tail — not absent because they are unimportant, absent because nobody profits from carrying them.
And on one axis the web is worse than either. It has no historical record. Johnson's 1755 quotations are still checkable. A page cited in 2019 may say something different today, or nothing at all, and the citation will not tell you which.
The knowledge is out there. What is missing is a slip to write it on.
It helps to see the whole sequence, because the web has walked most of it — and skipped the one stage that mattered most.
| The dictionary | What it was | The web equivalent |
|---|---|---|
| Hard-word listsCawdrey, 1604 | 2,500 difficult words. Ordinary language assumed. Selective by difficulty. | DirectoriesYahoo, DMOZ Human-picked, notable sites only. |
| Compiled universalsBailey, 1721 | Broader, but assembled from earlier dictionaries rather than fresh reading. | Meta-searchDogpile, and every merge since Built from other people’s indexes. |
| Authored authorityJohnson, 1755 | One mind. Broad coverage. Quotations chosen to illustrate definitions already formed. Opinion embedded, basis unauditable. | The ranked results pagewhere search still is One algorithm, criteria unpublished, snippets illustrating a ranking they did not produce. |
| Evidence-first, historicalOED, 1857–1928 | Definitions derived from dated citations. Distributed reading. Descriptive. Every entry checkable a century later. | — never built —the missing stage No dated citations. No derivation from evidence. No record of what a page said before. |
| Corpus-derivedCOBUILD, 1987 | Built from a machine-readable corpus. Frequency-driven, genuinely descriptive at scale — and you cannot trace a definition back to one line. | Generated answersnow Fluent, corpus-shaped, and unable to say which page any sentence came from. |
The web went from Johnson straight to COBUILD, without ever building the OED.
That is the whole problem in one line. Corpus methods were a real advance for lexicography and they are a real advance for answering — but COBUILD’s weakness is exactly the language model’s weakness. A corpus tells you what is common, not what is true, and frequency is not provenance. Sinclair knew what he was trading; the trade was worth it because the OED already existed underneath, and you could always go and check.
On the web there is nothing underneath. The generated answer is the first and only layer, and the citations it carries were attached afterwards to prose that did not come from them.
Here is the part that should be obvious and somehow is not. There are people who know a subject better than anything currently ranking for it — the machinist, the midwife, the archivist, the person who has kept the same river gauge for thirty years. They are not absent from the index because they lack authority. They are absent because there is no way in.
Publishing a page requires a site, an audience, and a tolerance for optimisation. Correcting a bad answer requires a channel that does not exist. So the expertise sits in heads and inboxes and local societies, and the index fills instead with whatever was written by someone who needed the traffic.
Murray's answer to this was not a platform. It was a form. The slip was small enough that an ordinary person could produce one in a minute, and structured enough that it was worth something the moment it arrived. It asked for evidence rather than eloquence, and it did not care who you were.
Everything above is a description of MoatGoat's design, arrived at from the other direction.
An index you hold rather than query. Every claim carrying its source, so an answer can be checked rather than trusted. A ranking formula published rather than defended. Coverage chosen deliberately, including the parts no advertiser would fund. And an answer that declines when the sources do not support it — which is the modern form of the same discipline, because the alternative to we do not know has always been somebody's confident guess.
The piece we are adding now is the slip. A way for someone who knows a thing to say here is a passage that shows it, and have that land in a shared index where it is checkable by anyone. Not an essay. Not a profile. A quotation, a source, and a date.
Two honest cautions, because the romantic version leaves them out.
It took seventy years. The first instalment appeared in 1884 and the work was not finished until 1928, and it very nearly destroyed Murray, who died with the letters T through Z unwritten. The lesson is not that volunteers will do it. It is that a slowly accumulating corpus needs a spine that does not depend on one exhausted person.
And the reading programme inherited its readers. Volunteers read what volunteers read — literature, mostly, mostly by men — and the first edition under-represented women's writing, non-literary sources, and Englishes outside Britain. Later editions spent decades correcting it. The failure was not bad evidence. It was unrepresentative evidence that looked comprehensive, which is the exact risk any curated index runs, and the one worth measuring from the first day rather than the tenth year.