Development covered 5 Aug 2026 to 8 Aug 2026 (commit dates).
Numbers are where a long book comes apart. Not plot, not character, numbers: a person's age, how many years since a thing happened, what an account book says, how many words are carved on a stone. Each one is written once in a chapter nobody will reread and contradicted three hundred pages later by an author who is thinking about something else.
This stretch built mechanical checks for that across the tie-in novel, and the interesting result was not the checks. It was what to do with what they found.
Seven checks that pay for themselves
Each of these came from finding the same defect by hand at least twice, which is the only reason I trust the list.
An age that goes backwards. A character is sixty-eight in one chapter and sixty-one forty chapters later. Trivial to check once every mention is extracted, impossible to hold in a head.
A dead character who turns up. One chapter says the shopkeeper passed away two years ago; a later one has her behind the counter at seventy-one. This one is embarrassing in a specific way, because it is the kind of mistake a reader remembers.
Dialogue arithmetic. Characters trade in numbers constantly, and if two of them agree that three lots of nine hundred is two thousand eight hundred, some reader will do the multiplication.
The calendar running backwards. Both the long form, where the world's own year count regresses between chapters, and the short, where a date inside a single chapter goes the wrong way.
An age used as a name. The book often refers to an unnamed person by their age, and that number then has to stay put across every reference.
Counting characters and counting punctuation. When the text says an inscription is fourteen characters long, somebody will count. Punctuation is not a character in that sense and the naive count includes it.
A lifespan that does not increase with rank. The world's own rules say it must; two places said otherwise.
Then the part that made it usable
The first full run produced around three hundred findings. That is not a report, it is a second job.
Checking a sample of them against the actual text gave a ratio I now expect: roughly a third are genuine errors, a third are false alarms from the checker misreading context, and a third are real inconsistencies the author put there deliberately and would be wrong to remove.
That last third is what separates prose checking from code checking. A character can be wrong about a number. Two accounts of the same event can differ because two people remember it differently. A check that flags those is not detecting a defect, it is detecting fiction.
Adding three triage rules that discarded the most common false shapes cut a three hundred item run to fifty eight, with no genuine errors lost. Getting the noise down was worth far more than adding an eighth check would have been. A report nobody reads finds nothing.
The editorial answer is usually fewer digits
Here is the part I did not expect to be a rule and now is.
A wrong figure in the text invites you to work out the true one and set it down. Precision like that belongs in an account book, an inscription, an official notice, a formal record, because in those places precision is part of the texture and a reader feels it.
Everywhere else the better move is to blur it, and the move after that is to remove it entirely. This is an adventure novel, not an arithmetic problem. A reader wants the scale of a thing, not its audit trail, and a sentence saying a fortune was several tens of thousands is both more readable and impossible to contradict later.
So the fix has three settings in order: blur, delete, and only then compute. I had written that down as delete first and immediately lost the sense of scale in a few passages, which is how the middle setting arrived.
A related rule fell out of the same pass. Precise numbers past four digits do not belong in narration at all. A hundred thousand is a quantity a reader can feel; a hundred and seventy four thousand two hundred and six is a number they skip, and it is one more thing that can be wrong in chapter three hundred.
The method behind all of this
None of the checks was designed in the abstract. Every one started as a defect I fixed by hand, followed by one question: is this an instance of a class.
That question is the whole method. Fixing the instance costs ten minutes. Asking whether it is a class costs another ten and sometimes turns into a check that finds thirty more. It also, usefully, sometimes returns no, and knowing which of your fixes were one-offs is worth having.
I also keep a short list of the checks I tried that did not pay: shapes I suspected would find errors and that returned nothing but intentional repetition. Writing down what did not work is the only thing that stops the next pass from building it again.
Everything here ships in Incremental RPG.
Top comments (0)