Toward Ludic-Aware Narrative Generation: A Neuro-Symbolic Framework for Evaluating Playability in LLM-Generated Backstories

Abstract

Large Language Models produce fluent prose, but for interactive media they tend to generate narratively sterile backstories whose conflicts are all resolved before play begins: the Closure Paradox. Conventional metrics, from n-gram overlap to retrieval-augmented faithfulness, do not reward openness. We propose a dual-axis framework separating Temporal Consistency from Ludic Potential, anchored by two metrics computed deterministically over a symbolic Fact Ledger: the Ludic Potential Index (LPI), a weighted measure of a backstory’s openness to future play, and the Potential Conflict Value (PCV), a per-event measure of conflict fuel. A four-stage pipeline extracts the Ledger, scores it deterministically, and has a separate agent stress-test it by building a first-session encounter. A pre-registered ablation gates generation on LPI and PCV and measures Adaptation Effort, the share of session entities a synthetic Game Master must invent. Gating helps or harms depending on the generator’s capability, which size predicts only loosely: it lowers a commodity 14B generator’s load, leaves a frontier model unchanged, and raises it for some smaller generators. An expert study finds LPI tracks judgements of openness and overall playability modestly (r=0.24), with single-rater agreement below the pre-registered threshold. PCV adds no prediction beyond LPI.