Extended Controlled Deformation Protocol_ENG

This note presents a formal description of the E-CDP (Extended Controlled Deformation Protocol), an experimental protocol designed to address challenges in the digital humanities. The protocol employs large language models (LLMs) as a tool for the quantitative analysis of text corpora, rather than as an object of study in itself. The core idea is as follows: using a target corpus, several versions of a training dataset are created — one unchanged (control) and three deformed versions for each semantic domain under investigation (semantic removal, introduction of logical contradiction, and stylistic paraphrase as a control). A base LLM is independently fine-tuned on each version using identical hyperparameters, yielding a family of equivalent models. The changes induced by these deformations are assessed using a system of binary questions with expert-predefined answer keys. The quantitative metric — the Structural Significance Coefficient (SSC) — captures how critical a domain is for preserving the integrity of the discursive worldview embedded in the corpus. This note sets out the methodological architecture of the protocol; all implementation details, examples, program code, and current experimental results are published sequentially in the online appendix.