Bridging quantum semantics and interactive narrative

Storyworlds as Sparse Autoencoders

A storyworld is a formally structured interactive environment: a set of characters, variables, encounters presenting multiple options, formulas for unlocking options and determining the reaction to them, and subsequent effects on variables and encounter transitions, which constrain how narratives can unfold. Unlike linear stories or unstructured text, storyworlds explicitly encode goals, affordances, conflicts, and consequences.

When used for training and as a tool of inter-AI communication, these environments function as compressed, human-legible representations of complex systems. Agents must reason about state, intent, causality, and trade-offs within a bounded semantic space—precisely the dimensions we want to make visible inside learned models. Skills and a Verifiers env have been developed to assist in AI generation and iteration of storyworlds based on user prompts or ranges of historical/literary adaptation, as well as use in the context of diplomacy strategy games. An experiment harness called Storyforge facilitates the study of this output using MLFlow.

This research framework treats storyworlds as a form of narrative sparse autoencoder: structured inputs that encourage models to learn disentangled, interpretable internal features. By aligning model training with how humans naturally reason about agents and worlds, we aim to bridge the gap between mechanistic interpretability and human-understandable concepts.