Novalist directs five agents from premise to canon.
Gravitas is the first world built with Novalist. Architect, Loremaster, Scribe, Critic, and Editor — five specialized roles — challenge one another toward a coherent, canon-consistent manuscript, then hand it downstream to Nexus Kinetis, which grows the living canon. Explore the growing world, or read the pages they wrote.
5
Creative Roles
collaborating in the loop
3
Chapters in Manuscript
across the current acts
13
Canon Entities
in the current knowledge graph
"Chapters in Manuscript" and "Canon Entities" reflect this demo world's current content, not a live count.
The Agent Pipeline
Five conceptual expert perspectives — Architect, Loremaster, Scribe, Critic, Editor — inform three real execution stages: Plan → Write → Polish. A Human Director oversees the run and is the only one who can approve a manuscript version; nothing reaches canon just because a draft was produced. Retrieval and context-gathering (pulling in relevant existing canon, which may involve an embedding call) is separate from drafting itself — it informs the agents, it isn't one of the three drafting calls. Downstream, Nexus Kinetis — this world's reusable canon engine, configured here with a creative-fiction domain package defining its ontology and expert roles — turns Director-approved pages into proposed knowledge; a separate human review of that proposed knowledge (not the manuscript approval itself) is what actually makes it canon in the world. Hover or select any node to trace how they connect.
1 · Plan
Architect (structure) · Loremaster (canon context)
Breaks the premise into acts and beats, and pulls in relevant existing canon before any prose is written.
2 · Write
Scribe (prose)
Drafts full prose scenes from the plan.
3 · Polish
Critic (review) · Editor (continuity & final pass)
Reviews the draft for continuity and tone, then produces the version the Director sees.
A Human Director approves each version before anything reaches canon.
Evaluation & Experiments
Local tooling used to test and prototype this system, and — the first card only — how runtime data is actually validated in the deployed app.
Pydantic + Zod
Two different scopes, not one system
Pydantic structures evaluation graph state, local to the Python harness. Zod validates application API contracts and runtime data throughout the deployed app.
Langfuse
Optional trace export
Exports evaluation runs for inspection when configured; cloud delivery is unverified.
LangGraph
Experimental eval executor
Runs the evaluation dataset through a local graph — a prototype, not the discovery path.
Temporal
Experimental workflow test
Explores durable-workflow retry and cancellation locally, outside the deployed app.