Where a Novel Loses Track of Its Own Words

The received wisdom is that a manuscript gets sloppier as it goes: that a writer is sharp for the first hundred pages and tired by the last. We measured 392,764 real inconsistencies across 12,602 published novels to see where they sit, and the folklore is wrong. Inconsistency does not accumulate towards the end of a book. It sits exactly where the vocabulary is: 22.7 per cent of a novel's odd spellings first appear in its opening tenth, against 21.5 per cent of the vocabulary itself. The two curves are the same curve.

What is being counted

One row per spelling a book uses that is not the spelling it mostly uses for that word. If a novel writes Sun-god eleven times and Sungod once, the once is the row. A word written three ways produces two rows, because the thing being counted is a departure rather than a word.

Across 12,602 full length novels there are 392,764 of them. Sixteen novels in the corpus carry none at all.

Position comes from the scanner's own mention lists, the same lists the mention heat map is drawn from, so every figure below is measured in the same units a reader sees when they scan their own book.

Most of it is one slip. A third of it is a book that never decided.

These rows point in opposite directions and both are true. Two in five inconsistencies are a single occurrence: a two second fix, and the reason a consistency report is not as frightening as its headline count suggests.

Three in ten are the other thing entirely. A near even split is not a mistake a book made once. It is a decision the book never made, applied inconsistently for eighty thousand words, and it is the kind a reader notices because they see both spellings often enough to register the difference. Those are the ones that take an afternoon.

The practical consequence is that a raw count is a bad way to read a consistency report. Thirty four findings that are all single occurrences is a twenty minute job. Six findings that are all near even splits is a decision you have been putting off since chapter three.

Capitalisation is commoner. Hyphenation is worse.

Nearly two thirds of what a consistency check finds is a capital letter: the Duke and the duke, the Manor and the manor. Only a quarter of those are a one off, which means capitalisation is overwhelmingly the divided kind. The book is genuinely using both forms throughout, usually because the underlying question, is this a title or a description, was never settled.

Hyphenation is the opposite shape. It is half as common and three quarters of it is a single stray occurrence: a book that writes Sunday-school four hundred times and Sunday School once. That single occurrence is the easiest thing in a manuscript to fix and the hardest to see, because there is nothing wrong with the sentence it is in.

Where in the book it happens, and the answer nobody expected

Where a book's first odd spelling of each word falls, one tenth of the book at a time, with the corpus figure for where a novel first uses its vocabulary at all underneath it:

A novel front loads its vocabulary hard, which the benchmark established: 21.5 per cent of every word a book will ever invent arrives in its first tenth. What this study adds is that inconsistency arrives on precisely the same schedule, at 22.7 per cent. There is no late book collapse. The opening is not cleaner than the ending; it simply has more words in it to get wrong, and it gets them wrong at the same rate.

Measured against each word's own life rather than against the book, the symmetry holds again. An odd spelling comes late in that word's life 32.1 per cent of the time and early 34 per cent of the time. A book is no likelier to lose a word it had already settled than to fail to settle one it started wrong.

One caution about a figure that looks like a finding and is not. Averaging every occurrence of an odd spelling rather than taking the first produces a neat bell centred on the middle of the book. That shape is arithmetic: a word used throughout a novel averages out near its midpoint however evenly it is spelled. It is in the published dataset and it means nothing.

Does a longer book get worse?

A 200,000 word novel carries 4.8 times the inconsistencies of a 50,000 word one. Per 10,000 words the five bands are indistinguishable: 2.71, 2.95, 2.92, 2.88, 2.98. There is no length at which a book starts going wrong faster.

That is reassuring and it is not comforting. A long book is not sloppier. It is the same book, more times over, and the work of checking it scales with its length rather than with its difficulty. What does change with length is the cast: a 200,000 word novel names 81 characters against 23 in a 50,000 word one, and it sets up four times as many terms that it then stops using.

The things that are not spelling at all

A consistency surface is wider than its spellings. Three more measurements, medians across all 12,602 novels:

The median published novel introduces six things it then stops mentioning, and 97 things it uses repeatedly with no sentence among their first ten mentions telling the reader what they are. Neither is a mistake in itself. A novel that explained everything would be unreadable. But a term used three times across ninety thousand words and never explained is exactly the term a reader forgets and a writer assumes was covered, and it is worth knowing which of yours they are.

What this cannot tell you

A variant is a difference on the page rather than a mistake. Lady and lady are both correct English and which one a sentence wants depends on the sentence. What these figures show is how often a finished book answers the same question two ways, which is a useful thing to ask about your own draft and not a verdict on anybody's.

Transcription is somewhere in these numbers. Every text is a transcription of a printed edition and a proportion of these differences belong to the compositor or the typist rather than to the author. That is a reason to read the medians as floors.

Position is measured in words rather than pages, because a percentage of the way through a manuscript is what the scanner knows and what an author can act on. And the corpus is weighted towards books first published before 1930, so this describes the books that established the conventions rather than what modern publishing does; the published methodology states that limit and every other one.

A practical next step

The finding that changes what you do is the split between single slips and near even splits. Sort a consistency report by how many times the minority spelling appears, deal with the ones that appear once in a single pass, and then treat what is left as a set of decisions rather than a set of errors.

The Novel Kit's capitalisation checker and hyphenation checker are the two measurements behind the table in section three, run against your manuscript instead of these. They report both spellings, how often each appears and where, which is what turns a count into a decision. One off terms, last seen and term clustering are the three term life measurements in section six, and the fix list is where they all arrive in the order you would actually work them.

Sources checked

Every figure here is generated from the Novel Manuscript Benchmark, a public dataset of public domain books scanned with the tool this site sells, on its default settings. All 392,764 inconsistencies are published below as a CSV with both spellings, both counts and the position of each, and are regenerated whenever the corpus grows, so a later run will move these numbers and the dataset rather than this page is the authority.

No customer manuscript was involved and none could have been. Nothing in this product's storage holds manuscript text or extracted terms, which its own test suite enforces, and the corpus is public domain fiction precisely so that a study like this can exist without going near anyone's work in progress. Each book was scanned by the engine that runs a scan for a reader, so a figure here is comparable with a figure from a scan of your own.

Frequently asked questions