Reconstructing a Fragmentary Text from Priors: Language-Model Agents and Learned Alignment for P. Fouad 266b (848)

Reconstructing a fragmentary manuscript of a known text rests on prior knowledge that editors apply implicitly and never evaluate. We formalise this knowledge as four priors (the reference text, the scribe's orthography and letter metrics, the physical geometry of the roll, and a rule for reading ambiguous ink) and combine them with large language model (LLM) agents and a constrained alignment model in a three-stage pipeline: reading, placement and layout. We evaluate it on the Greek Deuteronomy roll P. Fouad 266b (Rahlfs 848), which survives in some 140 fragments. An LLM agent, tuned by reflective prompt evolution, transcribes the plates and restores them against the reference text, missing 1 of 943 present letters on held-out plates. Fragments are placed by a banded, information-weighted edit distance with scribal substitution costs, chained across rows by a line-advance model and committed only above a confidence threshold. On a two-fold, plate-level split the placement model places 88% of held-out units (110 of 125) with no false placement (95% upper bound on the error rate about 3%). A deterministic constraint solver then rebuilds 85 columns that keep all 6,736 inked letters at their text positions. Unplaced units are mostly too short to place; the four that no setting places depart in wording from both reference texts, so the model's failures locate textual variants rather than errors. Transferred unchanged to a second roll, P. Fouad 266c (847), it places one unidentified fragment.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24408888
Primary Topic
Digital Humanities and Scholarship
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Reconstructing a Fragmentary Text from Priors: Language-Model Agents and Learned Alignment for P. Fouad 266b (848)

Samiya Ahsan, Tayyaba Saad Ahsan
1 links
Zenodo (CERN European Organization for Nuclear Research)
Digital Humanities and Scholarship
preprint

Reconstructing a Fragmentary Text from Priors: Language-Model Agents and Learned Alignment for P. Fouad 266b (848)

Samiya Ahsan, Tayyaba Saad Ahsan
preprint en

Abstract

Reconstructing a fragmentary manuscript of a known text rests on prior knowledge that editors apply implicitly and never evaluate. We formalise this knowledge as four priors (the reference text, the scribe's orthography and letter metrics, the physical geometry of the roll, and a rule for reading ambiguous ink) and combine them with large language model (LLM) agents and a constrained alignment model in a three-stage pipeline: reading, placement and layout. We evaluate it on the Greek Deuteronomy roll P. Fouad 266b (Rahlfs 848), which survives in some 140 fragments. An LLM agent, tuned by reflective prompt evolution, transcribes the plates and restores them against the reference text, missing 1 of 943 present letters on held-out plates. Fragments are placed by a banded, information-weighted edit distance with scribal substitution costs, chained across rows by a line-advance model and committed only above a confidence threshold. On a two-fold, plate-level split the placement model places 88% of held-out units (110 of 125) with no false placement (95% upper bound on the error rate about 3%). A deterministic constraint solver then rebuilds 85 columns that keep all 6,736 inked letters at their text positions. Unplaced units are mostly too short to place; the four that no setting places depart in wording from both reference texts, so the model's failures locate textual variants rather than errors. Transferred unchanged to a second roll, P. Fouad 266c (847), it places one unidentified fragment.

Zenodo (CERN European Organization for Nuclear Research)
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.

Laika access coming soon