Large Language Models (LLMs) are taking on increasingly agentic roles in narrative creation. As they engage in core stages of creative writing, including ideation, character development, and plot structuring (Yuan et al., 2022; Wang et al., 2024; Venkatraman et al., 2025; Park et al., 2026), their capacity for coherent narrative world-building becomes as important as output quality. In particular, how LLMs decide which narrative elements fit together within a single story deserves closer attention—whether, for example, an era of early ritual can coexist with digital rooms shaped by code. This is difficult to examine through generated texts alone, since both the elements involved and the criteria for judging their compatibility remain implicit in the final output. To address this, we draw on a narrative selection task and observe how LLMs select compatible sets of narrative elements.

In probing LLMs' narrative compatibility judgment, a reasonable concern arises: what appears to be assembling a compatible set may, for LLMs, amount to little more than combining semantically similar constraints. Recent work on LLMs supports this concern. Li et al. (2026), for example, find that LLMs follow salient surface cues at the expense of implicit feasibility conditions—the implicit preconditions that must be satisfied but are not explicitly stated in the prompt—and that this pattern generalizes across heuristics such as semantic similarity, cost, and efficiency. Specifically in the narrative domain, De Langis et al. (2026) demonstrate that LLMs' overt responses fail to fully distinguish coherent from incoherent narratives, raising the question of what underlies these judgments. We therefore ask whether a model's selection of compatible elements reduces to semantic similarity, or whether other factors beyond semantic similarity shape their selections.

To investigate this, we use the Narrative Constraints Library (Jung et al., 2026), which comprises 200 constraints across four narrative elements (Event, Style, Character, Setting). We prompt three LLMs (GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro) to perform two complementary selection tasks. In the free-selection task (Task 1), each model selects 20 constraints from the full pool of 200 without an anchor, treating each set as constituting a single coherent narrative (300 runs). In the anchor-conditioned task (Task 2), each model receives one constraint as an anchor and selects 19 additional constraints that best fit it, under explicit instructions emphasizing compatibility (200 anchors × 10 runs = 2,000 runs). Task 1 yields each constraint's baseline selection frequency, used as a popularity control; Task 2 isolates compatibility judgments when compatibility is made the explicit criterion.

In our analysis, we examine one focal variable (semantic similarity), one control (popularity), one absorption mechanism (anchor fixed effects), and one structural variable (candidate element identity). Semantic similarity is the pairwise cosine similarity of all 200 constraint embeddings (jina-embeddings-v5-text-small). Popularity is the log-transformed and z-standardized marginal selection frequency from Task 1. Anchor fixed effects absorb anchor-level heterogeneity in baseline selection rates. Candidate element identity captures any systematic tendency for compatibility judgments to favor particular element categories; because each anchor's element category is fixed, anchor-element effects are absorbed by the anchor fixed effects. Using these components, we fit binomial GLMs predicting per-anchor selection probability and decompose the McFadden pseudo-R² across nested specifications, quantifying both the incremental and unique contributions of each component relative to semantic similarity. We additionally examine element-level residuals from the full model to identify systematic over- or under-selection associated with specific element categories.

Our analyses show that the LLMs' narrative compatibility judgments are not reducible to semantic similarity. For GPT-5.2, for instance, similarity is the strongest single predictor of selection probability, yet accounts for only a modest fraction of variance in selection in absolute terms (McFadden R² = 0.118); even the full model reaches an R² of only 0.264. Within the explained variance, similarity accounts for 45%, with the other three components contributing the remaining 55%. Element-level residuals reveal systematic asymmetries: Style constraints are universally over-selected and Character universally under-selected—patterns independent of similarity, suggesting element-specific structural heuristics whereby LLMs treat Style as a universal complement and Character as a scarce, slot-limited resource.

These findings have two implications for how we understand and use LLMs in creative writing. First, semantic similarity remains a significant factor in models' compatibility judgments, but it does not fully account for them: other systematic patterns operate alongside similarity. Second, the mechanisms producing these patterns remain underexplored, leaving open how closely they correspond to human considerations in narrative composition. Writers and tool designers deploying LLMs as creative collaborators should be aware that what an LLM judges "compatible" may follow an internal logic of its own.