Daejeon, July 27–31
Religious belief, as an interior state, is never directly observable; it only appears as enacted through speech and practice. Anthropologists have questioned whether the category is universal or a distinctly Judeo-Christian mode of confessing inner convictions whose recognition depends on the interrogating framework (Needham, 1972; Asad, 1986). Recent scholarship reconceptualizes medieval religious belief as performative action enacted in situated speech acts (Arnold, 2023); inquisitorial registers, often preserving their question–answer structure, render that performance as a series of textual narrative acts. Yet accessing it systematically at corpus scale requires computational tools that preserve clause-level semantic granularity.
This paper addresses that challenge through LLM-assisted clause-level annotation combined with mixed-effects regression modelling. Prior computational approaches have focused on topic modeling (Brookes / McEnery, 2018) or manual clause-level frameworks such as Quantitative Narrative Analysis (Franzosi, 2004) and Computer-Assisted Semantic Text Modelling (Zbíral et al., 2022; Shaw et al., 2026). We automate and extend such annotation via a single LLM prompt focusing on explicit belief acts: verbal clauses in which deponents or inquisitors assert, attribute, or deny religious conviction; then ask what structural conditions govern when and how they surface. The central question is what kind of belief occurs, and also when and why it surfaces: (1) how do broad trial context (register) and deponent sex shape belief frequency and subtype; (2) how does surrounding thematic content condition whether belief can emerge; and (3) how does belief behave as a sequential phenomenon once it enters the record.
Our dataset comprises 801 depositions (27,850 clauses) from Toulouse (1273–82; Reg. 0, n=192; Biller et al., 2011) and Bologna (1291–1310; Paolini / Orioli, 1982). The Bologna material divides into non-LS proceedings (Reg. 1, n=207) and a distinct civic unrest campaign, the Liber securitatum (Reg. 2, n=402), whose depositions are much shorter (median 9.5 clauses vs. 45 in Toulouse and 41 in Bologna non-LS). The structural difference matters: in a nine-clause deposition, a single belief utterance constitutes 11% of the text; in a hundred-clause deposition, it is 1%.
Table 1. Corpus overview by register.
| Register | n dep. | Total clauses | Median clauses | % belief clauses | Dominant subtype |
| Toulouse (Reg. 0) | 192 | 12,126 | 45 | 5.7% | socio-theological |
| Bologna non-LS (Reg. 1) | 207 | 10,767 | 41 | 9.3% | socio-theological |
| Bologna LS (Reg. 2) | 402 | 4,957 | 9.5 | 10.6% | socio-moral |
Two contrasting depositions illustrate the range. In Toulouse, Guillelmus de Moleriis narrated how "Bernardus proclaimed, that there was no salvation in any state except the heretical state" — socio-theological belief in an open interrogation. In the Bologna LS, Bertholomea confessed only that "the said friars would be spoken ill of because they had seized the said Bompetrus" — socio-moral belief, minimal context, no recorded questioning. Both are beliefs with very different conditions of their emergence — shaped by register, surrounding thematic content, and interrogative context.
All 801 medieval Latin depositions were processed via Anthropic Claude Sonnet 4.5 API (temperature=0) with cached prompts. The prompt decomposes depositions into verbal-complex-based clauses, applies multi-dimensional classification, and expands elided actions. All clauses are classified by content type — four belief subtypes (theological, socio-theological, socio-moral, generic) and non-belief content types — enabling identification of explicit belief acts within their full discursive context, not by keyword matching.
Validation proceeded in two phases. Two historians independently reviewed 300 randomly selected clauses: overall accuracy 87%, with highest performance for question–response tracking (97%), temporality (91%), and discourse agency (89%). A blind-coding phase tested 100 belief-classified clauses; historian agreement averaged 78%, with disagreements at category boundaries reflecting interpretive ambiguity. Derived data, prompt and code will be made publicly available.
Four progressively elaborated mixed-effects logistic regressions (GLMM, binomial, logit link; Gries, 2015; pymer4, Jolly, 2018) predict belief presence at the clause level — both aggregate and per subtype — with random intercepts per deposition: M1: register and sex; M2: deposition-level thematic context (proportional competing content: material support, social networks, spatio-temporal, encounters/interactions, religious action); M3: agency configurations and interrogative intensity; M4: local sequential clustering via ±3-clause burst measures.
Table 2. Model progression for agg. Belief as outcome: predictors, headline results, deposition-level random effect SD, and AIC.
| Model | Predictors added | Key result | Dep. SD | AIC |
| M1 | Register, sex | LS OR=2.18***; male OR=0.50*** | 1.38 | 13,080 |
| M2 | + Thematic content | LS reverses OR=0.56***; competing content OR≈0.00*** | 1.00 | 12,831 |
| M3 | + Agency, questioning | deponent+general OR=16.04***; LS×intensity OR=0.24*** | 1.09 | 10,534 |
| M4 | + Burst clustering | burst OR different by subtypes 1.95–2.55***; dep. variance −94% | 0.33 | 9,714 |
Explicit belief acts are rarer than the records' purpose implies. Across the corpus, 75.1% of all clauses occur in windows with no belief content at all; 95.6% have no theological belief in their immediate ±3-clause environment. Explicit belief is not the ambient content of these trials, it is episodic and concentrated.
M1 (register and sex). The Bologna LS shows higher aggregate belief likelihood (OR=2.18, p<.001), but this is driven almost entirely by socio-moral belief (subtype OR=6.88, p<.001). Male deponents show significantly lower belief likelihood (OR=0.50, p<.001). Deposition-level random variance is enormous (SD=1.38 aggregate; SD=5.88 theological): register and sex explain almost none of the variation between depositions.
M2 (thematic competition). The surrounding thematic landscape is the larger driver of variation. All competing content types suppress belief — material support and exchange (OR≈0.00, p<.001) and encounters/interactions (OR≈0.00, p<.001) reduce belief probability to near zero. When depositions reconstruct movements, meetings, and exchanges, belief collapses. Surprisingly, the Bologna LS register effect reverses once thematic context is controlled (OR=0.56, p<.001). The register that looked most belief-dense in M1 is not institutionally predisposed to belief, it simply has less competing content. Its brief depositions left no room for social-network narration, so socio-moral belief filled the available space.
M3 (interactional structure). Agency configuration is the primary organiser within depositions. Clauses with generalised immediate agency (God, the devil, abstract institutions) in either deponent or inquisitor narration show the largest effects (OR=16.04 for deponent+general, p<.001): theological subjects predict theological belief. Notarial discourse suppresses belief (OR=0.13, p<.001); answers to explicit questions predict it (OR=1.76, p<.001). Question intensity is a positive predictor in Toulouse (OR=1.66, p<.01) but reverses in the LS (interaction OR=0.24, p<.001), where socio-moral utterances required little interrogative scaffolding.
M4 (stickiness and clustering — the central finding). Each additional belief clause within a ±3-clause window roughly doubles the odds that the target clause also contains belief (ORs: 1.95–2.55 across subtypes, all p<.001). Generic belief is the structural adhesive—the strongest burst effect (OR=2.55)—the moment when metadiscourse about faith itself sustains a self-reinforcing belief episode. Register-level fixed effects lose statistical significance entirely in M4: institutional differences operate through how belief is distributed within depositions, not through direct clause-level effects. And deposition-level random variance collapses by 94% (SD: 1.38 → 0.33). What appeared as deposition-level variation almost entirely reflects local clustering. Model fit confirms this progression: AIC 13,080 → 12,831 → 10,534 → 9,714 across the four models.
Fig. 1. Register (broad trial context) effects (OR, aggregate belief) across four progressive models. Bologna LS (Reg. 2) and Bologna non-LS (Reg. 1) both deflate M1→M2 when thematic content is controlled, but only LS reverses below OR=1. M3–M4: both effects collapse to non-significance. Filled red circles = Bologna LS, p<.001; open red circles = Bologna LS, NS; grey squares (dashed) = Bologna non-LS.
Belief in inquisitorial depositions is a threshold phenomenon, not a background one. It rarely saturates testimony. Instead, it emerges when competing thematic content recedes, when the inquisitor and deponent are no longer reconstructing social networks and material transactions; and once it enters the record, it cascades. The same interactional logic governs Toulouse and Bologna: register creates the conditions, but within each deposition, belief follows the same clustering dynamic.
For interpreting inquisitorial records, the LS case is the clearest test. LS socio-moral belief remains elevated after thematic control (M2: subtype OR=2.66): the community did express strong views about the executions. What the model adds is a claim about mechanism. In Toulouse and non-LS Bologna, belief is a threshold event within an open interrogation: interrogative pressure pushes it over the threshold, competing content displaces it, question intensity positively predicts it. In the LS, question intensity is negatively associated with belief (M3 interaction). Belief was not elicited through escalating questioning, it was the primary expected output of the situational frame itself; deponents were summoned to confess their reaction to the executions. These are two structurally different genres of belief performance: in the first, interrogative escalation is the felicity-producing mechanism; in the second, the situational frame pre-establishes the conditions for successful belief acts.
Female deponents showed consistently higher belief probability across all models (OR=0.76 for male in final model, p<.001), consistent with recent parallel findings on inquisitorial sources (Biller, 2005; Shaw et al., 2026). Whether this reflects different inquisitorial handling, gendered self-presentation, or simply less competing material-social content in female depositions remains an open question, but it is structurally robust.
By transforming 801 depositions into 27,850 annotated clauses and modelling them through four progressive steps, we show that religious belief as recorded in medieval inquisitorial depositions is rare, episodic, and structurally conditioned — not a direct function of institutional register. The strongest predictors are suspension of material-social narration, intense recorded questioning, generalised grammatical agency, and above all the local clustering dynamics of M4. What appears as individual deposition variation almost entirely reflects local clustering dynamics — whether belief entered at all and then cascaded.
Methodologically, LLM-assisted clause-level annotation can extend close-reading frameworks (CASTEMO, QNA) at corpus scale with validation accuracy sufficient for exploratory analysis (cf. Karjus, 2025). Sequence analysis would better recover whether belief episodes escalate internally toward direct avowal. The LS case also illustrates a general problem: naive semantic comparison across variable-length texts is confounded by deposition length. Religious belief in its explicit rendering is not the ambient content of inquisition trials, it is a threshold, interactionally produced, and locally self-sustaining. That is a finding both about medieval inquisition and about what computational methods can recover from it.