The Proof is in the Pudding: Structural Experimentation as Close Reading Strategy in the Age of LLMs

The rapid emergence of Large Language Models presents an emergency for humanistic education. LLMs excel at summarizing, ranking, and generating plausible text — but these very capacities threaten to make close reading obsolete. When students and scholars rely on AI-generated summaries rather than direct engagement with texts, critical thinking erodes. The ability to form independent questions, gather evidence, and construct interpretive arguments is displaced by algorithmic selection that chooses for the reader rather than supporting her. Attempts to move beyond the close/distant binary have diversified what computation does with literature — Piper's alternative reading modes (2018), Underwood's conciliatory "new scale of description" (2019) — and a smaller tradition has applied computational methods to individual texts (Eve 2019; Sá Pereira 2019). Yet even these single-text studies use computation to observe formal features. None uses computation to construct a new text as a way of experimentally testing how literary form produces meaning. Our approach, structural experimentation, extends the practice of "deformance" — the deliberate deformation of a text to release its latent formal logic — theorized by McGann and Samuels (2001) and advocated by Ramsay (2011) — by extracting the structures operating at the semantic and topical levels and applying them to new content, testing their meaning-producing operations in an alien domain. Novels have proven to be an excellent benchmark for assessing AI's contextual processing of extended texts (Wang 2024; Bonomo 2025). Current LLMs still face significant limits when processing long narratives: models with context windows larger than 200K tokens struggle when provided context exceeds 100K tokens, and their performance on meaning and relevance tests fails to reach 30% (Wang 2024). Operating on computationally decomposed formal templates extracted from discrete textual units, our approach sidesteps the context-window problem entirely while enabling a genuinely new mode of AI-assisted criticism. We demonstrate this through a technique of "intertextual mapping" applied to Alessandro Manzoni's The Betrothed (~420K tokens). The novel's opening chapter founds a shared, future-oriented historic space through specific formal operations (Raimondi 1975, 1987). We computationally decompose its formal architecture — syntactic hierarchy, perspectival sequence, and grammar of autonomous agency — into a template specifying the preservable structural properties of Manzoni's prose. This template is then applied to a different foundational scenario: the transformation of a contemporary Italian urban neighborhood. The AI performs formal parsing, constraint monitoring, and alternative generation; the human scholar provides philological judgment and evaluates the results. By detaching Manzoni's formal architecture from its original content and applying it to politically charged contemporary material, the experiment is designed to make visible formal operations — such as the naturalization of process, perspectival authority, and the grammatical exclusion of human agents — that remain invisible when form and content are fused in the original. We propose that the current emergency contains within it an emergence: AI tools that support the relationship between formal attention and critical thought. Realizing this requires design choices that prioritize structural analysis over summary, experimentation over observation, and the scholar's interpretive agency over algorithmic selection.