NovZon is an independent, founder-led software startup established in January 2026 in Ankara, Türkiye, building infrastructure for digital reading, creator workflows, and stateful interactive narrative. The ecosystem unites:
Machine-readable documentation for AI search agents and LLMs is published at https://novzon.net/llms.txt and https://novzon.net/llms-full.txt.
NovZon Multiverse is an advanced interactive storytelling engine developed by NovZon. It pairs a deterministic stateful narrative kernel with model-assisted generation to deliver non-linear, branching stories where reader agency creates real narrative consequences. Key capabilities include:
NovZon Multiverse is an experimental stateful narrative system where player intent can reshape future events while a deterministic kernel preserves canon, world rules and continuity.
Multiverse explores a controlled middle ground: authored structure defines what the world is; model-backed reasoning helps interpret what can happen next.
An architectural boundary separating immutable, observed history from bounded emergent futures. World state remains deterministic; prose rendering remains responsive.
Free-form probabilistic generation directly answering player prompts.
Every player action transitions through explicit structural checks before prose rendering and persistence.
Player selects an authored choice or provides free-form narrative input.
Intent is parsed into typed, inspectable parameter tokens.
Kernel checks invariants, character location, inventory & rules.
Legal state diff commits. Illegal moves are safely blocked.
Model receives masked, bounded context to render literary prose.
Automated validator checks prose against invariants before delivery.
Turn commits to branch tree; checkpoint created for rollback.
Key system concepts separating deterministic narrative infrastructure from probabilistic generation.
World state, inventories, and legal transitions remain explicit code execution rather than being delegated to free-form prompt completion.
Observed history is never silently rewritten when a player explores an alternate branch. Established events remain immutable foundation.
Free-form player actions are compiled into bounded, inspectable action tokens before they ever touch canonical state variables.
Alternate scenarios fork cleanly from verified checkpoints without mutating or polluting the author's primary canonical timeline.
Language models receive strictly filtered state projections. Secret author notes and future outcomes remain masked from player prompts.
Rendered prose is validated against state diffs and continuity requirements before the player ever sees it. Faulty outputs are rejected.
A working architectural framework for controlled narrative possibility.
Select an illustrative branch decision below to observe how player agency alters trajectory without violating past canon or breaking convergence milestones.
Player leverages trust with Archivist. Avoids guard confrontation; reveals hidden archive catalog.
Player challenges security threshold. Triggers lockdown state and raises faction alert level.
Player bypasses main entrance. Discovers sealed cipher without alerting archive personnel.
Current implementation status verified within the engineering codebase.
Typed schema validation, StoryIR loader and structural linting.
Immutable version history and config release staging.
Discrete turn tracking, legal state transitions, invariants.
Relational attribute maps, character state and inventory.
Point-of-no-return markers and branch rollback support.
User-scoped counterfactual scenario branch creation.
Secrets masking preventing model leak to player perspective.
Authenticated turn-based player execution environment.
Offline harness, invariant audit and consistency metrics.
Compiling free-text into legal kernel action tokens.
Benchmarking frontier reasoning and prose fidelity models.
Visual branch designer and canon constraint editor.
Multiverse is model-assisted rather than model-owned. We are evaluating frontier models for post-kernel roles such as free-form intent interpretation, narrative rendering, long-context character/world consistency and output evaluation.
Translating unconstrained player text into candidate action tokens that the deterministic kernel can validate or reject.
Synthesizing evocative, literary prose strictly bounded by canonical facts, visible player perspective, and authorial style.
Operating as an automated reviewer verifying generated narrative against world constraints and character voice before display.
Rigorous engineering benchmarks prioritizing reliability and consistency over unverified marketing claims.
Does generated narrative violate established world rules or secret boundaries? Zero silent canon drift remains the target invariant.
How do context strategies, bounded masking and provider selection affect latency, throughput and runtime cost?
Can the system preserve literary coherence and authentic character voice while granting the player meaningful agency?
“Every choice changes what comes next.
Not what has already happened.”
We welcome dialogue with researchers, narrative systems designers and engineers working on stateful AI architectures.
Rozet Acildi
