Does Next Token Prediction Explain Output Structure? Behavioral Evidence for Implicit Blueprints in Creative Text Generation

It has been hypothesized, from probing analyses of hidden states, that large language models (LLMs) possess an implicit blueprint—a pre-determined output structure—before generation begins (Li et al., 2024). But this hypothesis depends on access to model internals; no behavioral evidence existed.This paper provides behavioral evidence for the implicit blueprint without access to model internals. In a creative text generation task where an output attractor (Oyama, 2026) yields 100% convergence, forcing the opening token to an unrelated word (the sword name "Claymore"), or even reversing an attribute ("fire dragon"), leaves the blueprint vocabulary almost fully intact. Barnaby escape (Oyama, 2026), by contrast, breaks the blueprint—and once broken, forcing the original name back does not restore it. A blind evaluation by Claude Opus 4.8 (40 judgments) confirms these vocabulary-level results at the structural level.The opening token neither determines output structure nor restores it. Output structure is decided before the opening token—behavioral evidence consistent with Li et al.'s implicit-blueprint hypothesis, requiring no model internals. We verify this across 2 model families, 4 models (9B–35B), and 3,800 generations, all reproducible on a single consumer GPU. Finally, we ask whether what we recognize as intelligence in these outputs is inseparable from the structure that shaped them.