This paper presents a computational approach to redaction criticism in the Book of Isaiah, utilising a dual-model Transformer ensemble to detect stylistic discontinuities at the verse level. Rather than classifying text into predefined authorial strata, the method identifies locations where linguistic expectations shift abruptly. A causal language model (GPT-Neo) measures contextual disruption via perplexity, whilst a masked language model (DictaBERT) evaluates internal morphosyntactic coherence via pseudo-log-likelihood. These signals are normalised and synthesised into a Combined Fracture Index (CFI) across Isaiah’s 1,291 verses (Westminster Leningrad Codex).

The empirical results align with established biblical scholarship. The macro-structural transition between chapters 39 and 40 exhibits a severe statistical shift, with Isa 40:1 registering one of the strongest spikes in the corpus. The Deutero/Trito-Isaiah boundary is similarly detected, marked by a steep CFI drop from 55:13 to 56:1. Furthermore, verses within the Fourth Servant Song (notably 53:8 and 53:11) rank among the most anomalous overall, supporting its status as a distinct compositional unit.

Crucially, the method isolates unexpected micro-fractures. Chapter 28 emerges as the most internally spliced chapter, containing two of the highest global CFI scores (28:12, 28:17). Chapters 32-33 form a statistically unstable 'hot bridge'. In Deutero-Isaiah, dramatic spikes at 40:19 and 46:5 provide independent quantitative evidence that these anti-idol polemics are stylistically alien interpolations. Finally, Isa 63:13 appears as a severe, unexpected outlier within Trito-Isaiah.

Having validated that this methodology reliably detects physical editorial seams in Isaiah, the fully implemented computational pipeline is ready for deployment across the wider Hebrew Bible. This scalable infrastructure offers a transparent, replicable framework for generating new quantitative data to complement traditional philological inquiry