Large language models have made AI-driven games and interactive narratives newly plausible, but most current implementations remain structurally fragile. They rely on prompts, hidden context, and model-specific behavior rather than on durable, inspectable representations of world knowledge. The result is familiar: inconsistent continuity, weak portability across models and engines, limited auditability, and costly rewrites when execution layers change. This paper argues that AI-native worlds require a canonical substrate: a schema-governed, versioned, machine-readable knowledge layer that preserves the durable meaning of a world independently of any single model, runtime, or engine. Drawing an analogy to chemistry’s InChI identifiers and Pixar’s Universal Scene Description (USD), such a substrate provides a stable representational layer through which characters, quests, mechanics, locations, and relationships can be authored once and interpreted across multiple tools and execution contexts. We extend this thesis into a practical production architecture organized around three interacting domains: Brain, Runtime, and Body. Brain stores canonical world knowledge and design intent as validated documents with stable identifiers, explicit references, and governed schemas. Runtime interprets that knowledge into executable logic, state transitions, and engine-facing actions, while remaining distinct from any individual runtime instance. Body embodies that execution within a playable product through interfaces, levels, and scene implementation. This separation allows AI systems to operate as interpreters of structured world knowledge rather than as sole owners of canon, improving modularity, validation, reuse, auditability, and bounded reproducibility across execution contexts. We further examine implications for interoperability with existing structured content ecosystems and for production workflows in which bottom-up inputs, such as level design or embodied prototyping, are normalized back into canonical form. Our central claim is that for AI-native worlds to become durable, testable, and shippable products rather than brittle prompt-driven demos, game knowledge must be treated as production infrastructure.
Alex Haviland (Wed,) studied this question.