Shadow-Loom: Causal Reasoning over Graphical World Model of Narratives

Shadow-Loom's core move is treating narrative comprehension as a causal inference problem rather than a sequence modeling one. By encoding reader-state dynamics (mystery, suspense, surprise) as measurable properties of versioned world models, the work positions LLMs as tools for extraction and rendering rather than reasoning engines. This inverts the typical pipeline.

This connects directly to the semantic role labeling modernization from May 4th, which argued that explicit structured tasks remain valuable precisely because they offer interpretability and efficiency that implicit LLM representations lack. Shadow-Loom extends that principle: rather than asking an LLM to reason about narrative causality end-to-end, it builds a formal graph that LLMs can populate and query. The same logic appears in the SCISENSE-LM work from May 1st, which showed that constraining LLM reasoning through explicit cognitive scaffolding actually improves both fidelity and novelty. Shadow-Loom is another instance of the field moving away from end-to-end generation toward human-aligned structured workflows.

If Shadow-Loom's framework successfully predicts reader confusion or surprise on held-out narratives (measured against human annotations) at accuracy rates above 75%, that validates the causal graph encoding. If it fails to generalize beyond the training domain or requires heavy manual annotation of world states, the approach remains a research artifact rather than a practical tool for narrative analysis at scale.