NovelKit and Hermes: A Controlled Studio for Long-Running Fiction
NovelKit doesn't hand a whole series to one model and wait for magic. Hermes splits the work into artifacts, agents, and gates so each chapter can be checked before it becomes canon.
Long-running fiction doesn't need an all-powerful "AI novelist". It needs a system that knows who is doing what, which data is the truth, which chapters are approved, and which errors must be blocked before they spread to later parts. NovelKit is built around that need on the Hermes architecture: file-first canon, a task DAG, multi-layer memory, and a group of agents with separated roles.
The point worth noting isn't the number of models called. The point worth noting is that a model isn't allowed to turn a draft into canon on its own.
Project DNA and Story Bible: Locking the Rules Before Writing
A project begins with Project DNA. This file answers the questions a chatbot usually leaves vague: who the story is for, what the genre promise is, how fast the leveling goes, what level of violence is allowed, whether the ending is HE or BE, whether the narration is close or distant. DNA isn't a marketing brief. It's a contract the agent must read before creating any artifact.
After DNA comes the Story Bible. Characters have want, need, lie, and behavioral limits. The world has operating rules. The timeline records events and durations. Plot threads carry a status of open, in progress, or resolved. For cultivation fiction, the bible also holds realms and breakthrough fuel. For system fiction, it holds stats, quests, and the cost of a cheat.
If a new chapter conflicts with an old chat but matches the approved canon file, the file wins. This principle sounds dry, but it saves editors from arguments like "I think back in chapter 12 we said something different".
Five Memory Layers, Five Different Jobs
NovelKit doesn't call everything memory and dump it all into one store. Hermes separates five layers:
- Canon Files: Project DNA, database, outline, chapters, and approved reviews. This is the authoritative source.
- RAG SQLite: chunks the content to retrieve relevant passages faster.
- Vector DB: finds semantic similarity when the keywords don't match exactly.
- Episodic Memory: records what happens chapter by chapter after sync.
- Curated Memory.md: a short, curated working set used for the next writing turn.
The retrieval layers can be stale or incomplete. That's why they can't override canon. If a vector search pulls back a rejected draft, the Quality Auditor must reconcile it against the owner file. Good memory isn't remembering every word. It's remembering the right thing, with the right authority, at the right time.
The Genre Squad Under Lãng Khách's Orchestration
The orchestrator Lãng Khách doesn't write everything directly. It reads task state and assigns work in a hub-and-spoke model. The specialists don't call each other on their own, so the pipeline knows which artifact came from which role.
| Role | Primary responsibility | Not allowed to do on its own |
|---|---|---|
| Character Architect | Profiles, motivations, arcs, and relationships | Write a complete chapter in its place |
| World Builder | World rules, locations, factions | Change a locked premise |
| Plot Weaver | Arcs, foreshadowing, chapter beats | Sync canon |
| Prose Writer | Write the draft following the outline and style profile | Raise a realm or edit the bible on its own |
| Quality Auditor | Check continuity, OOC, timeline, style, and genre rules | Publish in the Publisher's place |
Separating roles doesn't make the writing less natural. It makes responsibility clearer. If the power scaling is wrong, you can trace it back to the outline, the prose, or the canon. If everything comes from a single chat thread, the cause usually dissolves into the conversation history.
Pipeline DAG and Sync Gate: A Draft Has No Right to Create Consequences Yet
Each chapter goes through a chain of dependencies: outline, write, review, revise if needed, then sync. The task DAG blocks chapter N+1 while chapter N is incomplete. The Quality Auditor scores against a rubric. In the current configuration, only a PASS at 85 or above is synced; 70-84 is a SOFT-FAIL that needs fixing; a blocking error or a low score creates a HARD-FAIL.
The score of 85 is an internal operating threshold, not a literary truth. It's useful because it turns a vague decision into an auditable gate. A Publisher can still require 100% human review, a higher threshold, or a custom rubric for premium titles.
The most important thing is the order in which state is mutated. The write stage creates a draft but doesn't officially update memory. The review stage catches errors. Only the sync stage, after passing, updates episodic memory, reindexes, and lets later chapters use the result. A bad draft can exist in the sandbox; it can't contaminate the story's history.
What Should "Copyright-Clean" Mean in 2026?
From 9 April 2026, Decree 134/2026/NĐ-CP adds provisions on copyright in cases where AI is used. The Copyright Office summarizes the principle: protection applies only to creations with a personal imprint, creative choices, and direct human control. This is why a promise like "AI writes 100%, copyright by default" is both careless and out of step with the current legal direction.
In NovelKit, the clean-sourcing commitment should be broken into checkable tasks:
- not importing pirated fiction libraries or unlicensed translations into the project's canon;
- storing the prompt brief, DNA, outline choices, review notes, and the owner's edit history;
- checking similarity and reference sources according to project policy;
- having a human approve the creative decisions and the final release;
- recording the model, agent, timestamp, and related artifacts in the provenance.
This process produces better evidence for the Publisher. It doesn't replace a lawyer, doesn't guarantee every dispute will end favorably, and doesn't turn a quality-gate score into a copyright certificate. That transparency is less flashy than the phrase "100% automated". In return, it holds up under real due diligence.
From a Digital Manuscript to a Package Complete Enough to Keep Producing
A minimum handoff package includes the synced manuscript, Project DNA, Story Bible, timeline, plot thread map, glossary, and review notes. For a white-label partnership, the Publisher also needs the provenance, the quality-gate configuration, the production schedule, and appropriate access rights.
To print a physical book, the manuscript still needs publishing edits, typesetting, permits, and the publisher's legal process. NovelKit prepares structured material so that team can work faster; the system doesn't grant publishing licenses on its own.
When Is This Model Unnecessary?
A short story of a few chapters, a voice-test project, or a manuscript used only internally may not need the whole stack. A chatbot and a good editor are sometimes enough. Hermes proves its value when the title is long, many people are involved, it needs to run in parallel, or rights and QA records must be handed off. In other words, the process should scale with production risk. Building a factory for a short story is wasteful; using a single chat box for a commercial catalog is too.
That's the line NovelKit chooses: automate the repetitive parts, keep the accountable decisions with humans. Post 5 moves from architecture to the business problem: how to use this production line to test a catalog of 200 chapters a week without turning throughput into spam.
References
- Copyright Office: Decree 134/2026/NĐ-CP and works using AI
- NovelKit: Canon file-first
- NovelKit: Quality Gate and the PASS 85 threshold
- Does NovelKit use pirated content as canon?
- The project policy is not to import pirated corpora or unlicensed translations into canon. Partners should write this into contract terms and check the handoff provenance.
- Can you publish a physical book right after PASS 85?
- No. PASS 85 is an internal production gate. Publishing a physical book still requires editing, procedures, and a publisher as required by regulation.