The Performance of Religious Belief in Medieval Inquisition Records: A Structural Analysis using LLM Data Extraction and Multilevel Regression

How can computational methods engage with something as elusive and mediated as historical religious belief? This chapter approaches belief as a spoken and recorded performance and tests this approach on medieval inquisition records, in which belief was legally central yet unevenly inscribed in text. Our corpus comprises 801 depositions from two registers produced around the turn of the fourteenth century: Toulouse (1273–1282) and Bologna (1291–1307), the latter including a distinctive subset of documents known as the Liber securitatum. Using a large language model (Claude Sonnet 4.5), we transformed the depositions into a clause-level representation of 27,850 clauses annotated for content topics, belief subtypes, narrative and immediate agency, and question–response structure; expert validation yielded an overall accuracy of 87%. We then applied four progressively elaborated sets of mixed-effects logistic regression models. Belief content proves to be neither ubiquitous nor uniformly elicited, but emerges in rare, concentrated bursts. Register and gender effects (clauses in female depositions being more likely to express belief) are largely mediated by thematic context: narration of material support, encounters, and other tangible actions strongly suppresses belief expression. Belief is more likely to surface where questioning is intensive and the question–response structure is preserved by the notary. Once introduced, belief "sticks": each additional belief clause in the surrounding window roughly doubles the odds of further belief content, and modelling this clustering reduces unexplained deposition-level variance by about 94%. Beyond these substantive findings on how inquisitors, deponents, and notaries shaped the performance and recording of belief, the chapter offers a transferable framework for combining LLM-assisted data extraction with statistical modelling to study mediated evidence of religious belief. This is a preprint version of a chapter submitted to De Gruyter's "Introductions to Digital Humanities". An accompanying repository with extended information about data and methods is accessible at https://github.com/DISSINET/belief-acting-annotations.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-10-20
DOI
https://doi.org/10.5281/zenodo.23493919
Primary Topic
Digital Humanities and Scholarship
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
preprint

The Performance of Religious Belief in Medieval Inquisition Records: A Structural Analysis using LLM Data Extraction and Multilevel Regression

Robert L. J. Shaw, Tomáš Hampejs, David Zbíral
Zenodo (CERN European Organization for Nuclear Research)
Digital Humanities and Scholarship
preprint

The Performance of Religious Belief in Medieval Inquisition Records: A Structural Analysis using LLM Data Extraction and Multilevel Regression

Robert L. J. Shaw, Tomáš Hampejs, David Zbíral
preprint en

Abstract

How can computational methods engage with something as elusive and mediated as historical religious belief? This chapter approaches belief as a spoken and recorded performance and tests this approach on medieval inquisition records, in which belief was legally central yet unevenly inscribed in text. Our corpus comprises 801 depositions from two registers produced around the turn of the fourteenth century: Toulouse (1273–1282) and Bologna (1291–1307), the latter including a distinctive subset of documents known as the Liber securitatum. Using a large language model (Claude Sonnet 4.5), we transformed the depositions into a clause-level representation of 27,850 clauses annotated for content topics, belief subtypes, narrative and immediate agency, and question–response structure; expert validation yielded an overall accuracy of 87%. We then applied four progressively elaborated sets of mixed-effects logistic regression models. Belief content proves to be neither ubiquitous nor uniformly elicited, but emerges in rare, concentrated bursts. Register and gender effects (clauses in female depositions being more likely to express belief) are largely mediated by thematic context: narration of material support, encounters, and other tangible actions strongly suppresses belief expression. Belief is more likely to surface where questioning is intensive and the question–response structure is preserved by the notary. Once introduced, belief "sticks": each additional belief clause in the surrounding window roughly doubles the odds of further belief content, and modelling this clustering reduces unexplained deposition-level variance by about 94%. Beyond these substantive findings on how inquisitors, deponents, and notaries shaped the performance and recording of belief, the chapter offers a transferable framework for combining LLM-assisted data extraction with statistical modelling to study mediated evidence of religious belief. This is a preprint version of a chapter submitted to De Gruyter's "Introductions to Digital Humanities". An accompanying repository with extended information about data and methods is accessible at https://github.com/DISSINET/belief-acting-annotations.

Zenodo (CERN European Organization for Nuclear Research)
Masaryk University (CZ)
Digital Humanities and Scholarship
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.