Development covered 2 Aug 2026 to 9 Aug 2026 (commit dates).
Myriad Immortal Sect ships with a novel attached, five books, one per realm, and the first book runs to 520 chapters. Nobody is going to read all of that looking for the moment a character's age stops adding up. So I pointed a model at it and asked for contradictions across chapters. The book follows the game's own realms, so a chapter like Void Refinement has to agree with both the novel around it and the game.
The first version was worse than useless, and the reason is worth writing down.
A report is a claim, not a finding
The reviewer came back with things like "chapter 148 says the pill takes three days to brew, chapter 300 says a week." That reads like a finding. It is a claim about two chapters, and the model produced it from a summary of those chapters rather than from the chapters.
So the reviewer now has to quote. Every reported contradiction carries the sentence it came from, on both sides, and a separate step goes back to the source text and checks the quote is actually there, word for word. If it is not there, the finding is dropped without a human ever seeing it.
That one rule changed the shape of the whole thing. The report went from a list I had to trust to a list I had to spot-check.
Verification has a direction, and I got it backwards
The second lens I added looks for craft problems instead of facts: a thread introduced and never picked up, a character acting against a motive the book established, a rule the book set for itself and then broke.
I wired it into the same quote check and it returned nothing for hours. The check was asking the source to confirm an absence. "This thread is never resolved" has no sentence to quote, because the whole claim is that the sentence does not exist.
The fix was to verify only the half that can be proved. Quote where the thread starts, then search forward from that chapter for the resolution, whole words only. Two of the first false alarms came from a substring matching inside a longer name.
Measured over the first full pass: 21 findings reported, 3 survived verification, and after I read those 3 against the text myself, none of them were real problems. That is a rough hit rate to publish, and I am publishing it because the alternative is pretending a checking pass is more decisive than it is.
Cleared once should mean cleared
The thing that actually made the pass usable was the least clever part. Findings I had read and dismissed came back on the next run, every run, because nothing recorded that I had looked. A pass that re-reports its own noise is a pass you stop running.
Now a dismissal is written down against the specific chapter and claim, with my reason. The next run subtracts them before it reports. Twenty four entries went in that way, and the backlog of things needing a human went to zero for the first time.
The game side of the same week
The rest of this stretch was the game, and it was the same discipline aimed at something else. I played through the mortal realm as a new player and wrote down everything that answered wrong: a panel showing figures no mechanic reads, a valley turning away a five element root over an element it does not require, a crafting screen saying it was out of herbs to somebody holding two hundred. Roughly two hundred of those, found by playing rather than by reading.
A checking pass and a playthrough turn out to want the same thing from you. Make it prove the claim, write down what you have already judged, and be honest in public about how much of it was noise. If roots and realms are unfamiliar words, the site keeps a list of cultivation vocabulary that settles most of them.
The novel and the game share one home, Cultivation Game; open the mortal realm in your browser right now.
Top comments (0)