Canon and claim levels
Canon, in this project, means the directory of files recording what’s been established about the world, the characters, and the timeline, kept separate from the plans and separate from the prose itself. Every entry also carries an explicit claim level, so a fact’s status is visible without having to guess how firmly it’s meant.
Established means the fact is on the page in an accepted chapter, and the entry cites where. Working means it’s binding until someone overrules it but hasn’t actually appeared in the story yet, which covers design decisions and anything the model proposed and David accepted. Open questions, as of a triage pass partway through the project, are no longer allowed to sit undifferentiated. Every open question in canon must be answered, ruled on with a decision, or explicitly deferred with a named trigger, something specific that will force the question to be resolved when it happens. A deferral with no trigger counts as a defect in the canon file rather than a legitimate state.
Why the undifferentiated state got closed off
There’s a specific reason for that. Earlier in the project, a mystery was introduced on the page without anyone having actually decided what its answer was. Nothing was wrong with introducing it unresolved, that’s normal craft, but the project let it sit as a bare open question with no plan for resolving it, and chapters kept accumulating evidence around it. By the time anyone checked, there was a real risk that no answer would fit everything already written. The rule that followed forbids that specific state: an open question now has to be either resolved, ruled as a working decision, or deferred with a stated trigger, no exceptions.
A related part of the system, kept separate from ordinary canon, is a directory of sealed answers: for mysteries the story depends on, the actual answer is written down before or at the moment the mystery appears on the page, but it’s written only by a subagent and never shown to David unless he explicitly asks. This exists so that a later check can verify every clue is coherent with a real answer, without spoiling the answer for the person who’s going to read the finished book like everyone else. 2026-08-27, checking a sealed answer exists is one such session: it asks, deliberately without wanting the content, whether a coherent explanation for the mystery actually existed.
Where facts come from and how they grow
Canon isn’t written speculatively ahead of the story. A skill called consistency-backfill runs after a chapter is drafted and checked. It reads what the chapter implies, not just what it states outright, and updates canon files to match: a stated fact becomes Established, an implied one becomes Working. Where canon was previously silent on something the chapter now depends on, it decides the smallest binding commitment and records that as Working too, preferring to under-specify rather than over-commit. Where a chapter conflicts with existing canon, the fix is proposed but never applied silently: it goes back to David as a flagged decision, quoted, with the reason and a note of which is cheaper, changing the chapter or changing canon.
The counterpart skill, consistency-check, runs the other direction. Given a draft, it checks every claim against the canon files in priority order, core facts and continuity first, with entries about theme weighted more softly. It reports contradictions, or the weaker category the project calls drift, where a claim is technically compatible with canon but strains it. Neither skill is allowed to weaken or delete an existing canon entry to make a draft fit. If canon and an earlier accepted chapter turn out to disagree, canon wins and the earlier chapter is flagged for revision, rather than the model quietly editing canon to fit what was already written.