The machine around the writing
Perpetūra has grown since I first wrote about it here. Five interconnected stories now: three public and named (Hyunjae in Joseon Korea, 1742; Reine in Belle Époque Paris, 1893; Liuyan in Tang dynasty China, around 750), plus a fourth, set in the present day, that just launched with its name sealed, and a fifth still to come. I write it, build it, test it and ship it alone, to the web and two storefronts.
Doing that taught me one honest lesson. At this scale, quality on your own is not willpower. Willpower is what you have on a good Tuesday. Quality is a system, a machine built around the writing, and the machine is what shows up on the bad Tuesdays too.
Restated canon rots
Earlier notes covered the canon shelf: story bible, style guide, an outline and a living progress doc per story. What I hadn’t learned yet was the sharper rule. At one point a canon rule, restated in a working note, drifted a few degrees off the original. The pages written against that paraphrase were confident, and wrong. Nothing flagged them, because they agreed perfectly with the restatement. Restated canon rots; only the original binds.
The fix was a gate, not a resolution. Before any drafting session I have to state, in one line, which story I’m on, what its documented next step is, and which canon documents I actually have open. If I can’t write that line, I don’t get to write anything else; I go and read first. It sounds bureaucratic for a studio of one. It is, and it’s also the cheapest quality control I own.
Done means passed, not felt
A chapter is not finished when it feels finished. It’s finished when it survives the done-sequence, a fixed chain of command-line checks, several of which hard-fail CI. The prose linter and the structural audit are old news in these notes. The newer members are the interesting ones. A dead-gate check hunts the nastiest bug in branching fiction: a flag that gets set, but never on any path that reaches the node reading it, so the scene renders for exactly zero players while every naive test passes green. A referent check asks whether a reader could attach a pronoun to the wrong man, which in a romance is not a small bug. And a register profiler measures the narrator’s voice numerically against the story’s own baseline; the engine note covers how that works.
Judgment, structured
The editorial council (several models from different vendors, fixed lenses, a debate round, a verify pass against the actual text) is written up elsewhere in these notes. The part that belongs in this one: none of it replaces editorial judgment. It structures it. The council produces a document, a human decides what the document means, and the human is me. I overrule it more often than I would have guessed.
Shipping is its own discipline
Production is human-gated. Code reaches the live branch when a person promotes it, never automatically, and pushed is not shipped — after every release I verify the change on the live URL before I believe my own tooling. Unfinished stories stay hidden behind a flag even in production, so shipping scaffolding is safe rather than brave.
The flag system got its proper test with the fourth story. It shipped sealed in plain sight: hidden behind the flag and an unlock gate, its name showing on the shelf only as “???” until a reader has finished three of the other stories. Nobody meets her early, not because I asked nicely, but because the code will not allow it.
What the machine is for
It would be easy to read all this as process for its own sake. It’s the opposite. The machine exists so one person can hold five novels’ worth of interconnected story at once and ship without the quality sliding, and so the scarce hours, the human ones, go where nothing else can spend them: the writing itself. Every gate buys me the same thing — the right to sit down with a chapter and think about nothing except the sentence in front of me. Any part of the system that stops earning that right gets removed.
The point of the system is the prose. Everything else is scaffolding, and scaffolding comes down.