One Pipeline, Every Field: The Structural Inevitability of Disciplinary Convergence Under AI Methodology

(Part 2 of a 3-part series.) (Revised version) Disciplines are not natural kinds. They are methodological accidents—territories defined by the tools available to study them, not by any inherent difference in what they study. This is no longer a philosophical provocation. It is an observable fact. By 2023, the bibliometric crossover had already occurred: in chemistry, biology, and physics, AI-methodology publications surpassed domain-specific publications. The methodological lattice that held disciplines apart has dissolved. What remains is only the difference of questions. This paper does not argue that convergence will happen. It documents that it has. We introduce the Learned Interaction Pipeline (LIP)—a four-stage computational structure (tokenize, interact, aggregate, decode)—as the common form that AI-based scientific computation takes across domains. We show that this structure is not specific to any domain but inherent to all AI-based data analysis: wherever phenomena can be indexed, their interactions learned from data, and outputs decoded, the computation converges on LIP. This paper demonstrates this inevitability in three ways. First, structurally: the four models map onto the same pipeline across tokenization, interaction, aggregation, and decoding stages. Second, empirically: when a chemistry model (MACE-MP-0) and a biology model (ESM-2) process the same molecules, their internal representations converge at interaction layers and diverge at output layers—a pattern reproduced across 27 proteins (p = .0001, permutation test; p = 2.3 × 10⁻¹², paired t-test). Third, functionally: the chemistry model's representations predict protein residue contacts above chance, confirming that functional information is transferable across disciplinary boundaries via the shared pipeline. A critical control establishes the boundary: a general-purpose language model (GPT-2) shows substantially lower representational convergence (CKA = 0.207, p = 0.026), confirming that pipeline sharing alone is insufficient—convergence requires training on data from the natural world. Domain-specific training does not erase this convergence but modulates it: the effect is absent below 23 residues and consistently present above 28. Large-scale validation (N = 88 proteins, 10–640 residues) confirms gradient persistence. (v3 to v4) Version 4 — What's new The thesis has shifted from a predictive claim ("convergence is structurally inevitable") to a descriptive one ("convergence has already occurred"). Bibliometric crossover evidence is now foregrounded as the primary empirical anchor. The framing of disciplines as "methodological accidents rather than natural kinds" is new to this version.