1QIsa-a, the Great Isaiah Scroll, discovered in 1947 in Cave 1 at Qumran, is one of the most important biblical manuscripts found in the Dead Sea Scrolls collection. It is notable for its length (734 cm) and preservation, containing the full ancient square script Hebrew text of the Book of Isaiah from ca. 125 BCE, making it one of the oldest surviving biblical texts (additional information and images can be accessed at http://dss.collections.imj.org.il/isaiah) [8]. The study of ancient manuscripts, such as the 1QIsa-a, combines historical research with various methodologies to uncover both the textual content and the scribal practices behind these texts. While the primary focus often lies in extracting historical facts from the content itself, the application of contextualization, sourcing, and corroboration is essential in the scholarly approach to understanding these manuscripts. As Van Drie et al. describe, corroboration involves comparing documents to address historical questions or confirm claims, with the identification of comparable documents—such as those written by the same scribe— providing deeper insights into their creation and context [9]. For ancient manuscripts, this requires identifying comparable documents, often determined by whether they were written by the same scribe. Documents attributed to the same scribe can provide additional comparative material for deeper contextual insights [5].

Paleography, the study of ancient handwriting, is a vital tool in this process. This methodology, as defined by Wakelin, involves analyzing features such as handwriting size, letterforms, and corrections to reveal individual scribal practices and styles [10]. Tov’s work emphasized that examining these characteristics in Hebrew manuscripts can help identify distinct scribes, allowing scholars to trace the evolution and transmission of texts [11]. The consensus surrounding the authorship of the 1QIsa-a Scroll is not uniform. Traditionally, paleographic methods suggested that the entire manuscript was copied by a single scribe, with subtle variations in handwriting attributed to personal idiosyncrasies or occasional inconsistencies. However, some scholars also proposed that the manuscript was actually the work of two distinct scribes—one responsible for Columns I–XXVII and another for Columns XXVIII–LIV. Recent advancements, particularly through AI techniques that detect micro-level handwriting variations, have revealed the involvement of at least two scribes in its creation [12]. This breakthrough demonstrates the power of combining traditional paleographic methods with modern technology to enhance our understanding of ancient texts, as well as the scribes who worked on them.

The study of writer verification has grown significantly in recent decades, driven by advances in machine learning and its applications to handwriting analysis. At least three key components differentiate computational approaches: the token level of textual input (e.g., character, word, sentence), the feature extraction methods, and the classification techniques (see Table I).

Across diverse studies, research on Latin and non-Latin scripts often employs similar feature extraction techniques and classifiers, demonstrating the adaptability of machine learning frameworks across scripts. However, the choice of input level significantly influences the system’s focus and accuracy. For example, character-level analysis can capture fine-grained handwriting traits, while word- or sentence-level inputs provide broader context. Despite these advances, integrating decision scores across token levels in a systematic manner remains a major challenge, as highlighted in Table I.

For historical manuscripts like the 1QIsa-a Scrolls, it is essential to select features and classifiers that align with paleographic practices. In this study, we employ edge-directional and hinge features, as these methods are intuitive to scholars studying historical manuscripts. These features have been successfully applied in handwriting analysis from their introduction to recent studies in capturing stroke directionality and curvature patterns (Refs. [31] and [32]). Additionally, we use an SVM classifier, which has demonstrated robust performance in tasks involving historical and modern handwriting (Refs. [30] and [33]).

Despite significant advancements, computational writer verification faces critical limitations when applied to historical manuscripts:

These gaps are particularly relevant for the 1QIsa-a Scrolls, where evidence suggests multi-scribe authorship. The inability to address such challenges limits the scope and accuracy of existing verification methods. Furthermore, due to the varying consensus on the authorship of the 1QIsa-a Scrolls, no standardized protocol exists for verifying proposed authorship claims through a writer verification system. Current systems lack the capability to assess and validate the theories surrounding the manuscript’s scribal origins.

This study introduces a building-block approach to address the gaps identified. At the character level, this framework-based approach assigns probability scores to individual characters based on their likelihood of having been written by the same scribe. These scores are then aggregated hierarchically at word and sentence levels, allowing for a comprehensive verification process. This modular approach is well-suited for manuscripts like the 1QIsa-a Scrolls, where the complexity of multi-scribe authorship requires a flexible and scalable framework. By integrating paleographic expertise with computational methods, the proposed system offers an intuitive yet powerful solution for historical manuscript studies. Furthermore, this framework provides scholars with a deeper understanding of the methods employed, bridging the gap between disciplines and enhancing collaboration.

The writer recognition framework mentioned in Section 2.2 mainly adopts an absolute decision strategy in identifying, verifying, or authenticating a specific token. For example, in a page-level writer recognition system, the recognition rate is usually computed using global features, and the decision process stops at this level. This method typically assesses whether the entire page can be attributed to a particular writer, relying on features that represent the overall style of the text. However, in the context of historical manuscripts, a more granular analysis is often required. In such cases, it is important to consider the recognition rate at smaller units, such as characters, words, or sentences. This is because, in historical manuscript studies, the goal is not only to determine whether an entire page was written by the same hand but also to investigate whether different hands may have written paragraphs within the same text, sentences within a paragraph, or even individual words within a sentence.

Unlike modern datasets, where the identity of the writer is typically labeled (often in controlled lab environments), historical manuscripts lack such clear labeling and may involve multiple hands contributing to the same document. In these cases, a detailed analysis at finer levels is necessary to assess the possibility of multiple authors contributing to a single page, paragraph, or even a specific word. This approach provides a more nuanced understanding of the document’s authorship, which is especially important when studying manuscripts where the attribution of authorship is uncertain or disputed. Thus, in this study, we integrated the building-block approach in the decision strategy stage with a machine learning- based writer verification system, as illustrated in Figure 1.

The building-block approach uses a layered decision-making strategy built on the outcomes of a machine learning model. At the lowest level, raw probability scores result from testing and evaluation within the machine learning framework. These scores serve as foundational inputs, which are further used to compute the intermediate-level and highest-level scores, allowing for a hierarchical assessment of writer verification. This decision strategy consists of three hierarchical layers. At the highest level (Layer 2), S2 represents the sentence- level probability score, indicating the likelihood that the entire sentence is written by a specific scribe. The intermediate level (Layer 1), where S1, i represents the word-level probability score for the ith word in the sentence, indicating the likelihood that the word is associated with a specific writer. At the lowest level (Layer 0), p0, i, j represents the raw probability scores of each individual character in word i, with j indexing the character within the word. This layered structure permits a detailed analysis, ranging from characters to words and sentences, facilitating writer verification at multiple levels.

To implement the proposed framework, we built an adaptation system to investigate the hands that wrote an ancient manuscript written in square script Hebrew, known as the Great Isaiah Scrolls. Our baseline used edge-directional and hinge feature extraction methods proposed by Bulacu et al. [30]. Edge-directional and hinge methods extract angular information differently—Edge-directional features compute the angle between handwriting strokes and a horizontal reference line, while hinge features measure angles formed between pairs of adjacent strokes. In both methods, these angles are categorized into structured bins based on their frequency of occurrence, with each bin assigned an empirical probability. These binned values represent the feature vectors, which are then used as input for training a machine learning model. Training-wise, we used an RBF kernel Support Vector Machine (SVM) to train and predict data [33]. SVM is known as a robust classifier that is suitable for datasets with high-dimensional and non-linearly separable features, which aligns with the characteristics of the data used in this study. To assess the contribution of the building-block approach, we leveraged SVM’s probabilistic outputs by opting for class probability scores as the output of the testing stage. These scores were then processed for soft voting and majority voting systems.

The Great Isaiah scroll consists of 54 pages, commonly called columns, of which it was assumed (by humanities [20]) that Scribe A wrote Col. I–XXVII and Scribe B Col. XXVIII–LIV. Research done by Popovic et al. [29] proved that assumption by implementing unsupervised learning for all columns. Based on these findings, the adaptation system aims to build a dataset and verify the data based on the claimed columns belonging to each scribe. We present an illustrated version of the letter samples, traced from the original images (see Figure 2). To consider all tokens that are available in large quantities, we chose tokens to be investigated from the lowest level, i.e., characters, words, and one sentence. We constructed the character-specific model based on the non-final form of the 22 letters of ancient square script Hebrew. We excluded the five final form letters since their frequency of occurrence was so limited that the machine learning-based system was unlikely to benefit from the small dataset.

For data collection, we expanded our custom dataset using the same approach described in our earlier study [34]. A total of 283 samples representing each letter (except for the letter tet, with only 145 samples, and the letter samekh with 232 samples, due to limited availability) were taken from each scribe’s corresponding columns. The training-validation set referring to Scribe A comprised single characters extracted from Col. I–XXVI, while the set referring to Scribe B included single characters extracted from Col. XXVIII–LIII. The test set consisted of one sentence from Col. XXVII (referring to Scribe A’s handwriting) and one from Col. LIV (referring to Scribe B’s handwriting).

This separation of data into Scribe A and Scribe B’s columns follows the presumed authorship as well as AI-based writer identification methods discussed in Section 2. The goal of this work was to verify these assumptions using the proposed system, leveraging the computational approach to confirm the authorship of the individual columns attributed to each scribe.

To validate the proposed framework, we utilized edge-directional and hinge feature extraction methods and an SVM-based classification technique. A major advantage of implementing handcrafted features, specifically edge-directional and hinge features, is that unique information based on the slant angle and curvatures of handwriting projection can be extracted. The collaborative nature of SVM and handcrafted features also brings an advantage for robust implementation, especially with our relatively small dataset, which could perform poorly during model training with a deep learning network.

Table II presents the best results of parameter tuning obtained from the training and validation section of the proposed writer verification system. To determine which letters are more representative or less representative of their respective features (edge- directional and hinge), we can analyze the training, validation, and gap scores provided for each letter. Representative letters typically have higher scores (indicating better alignment with the feature set), while less representative letters present with lower scores, suggesting that the features struggle to capture the stylistic traits of those letters. Letters like dalet, tet, and qof are the most representative, as evidenced by their consistently high scores in both training and validation datasets. Their distinct stylistic traits make it easier to classify them with edge-directional and hinge features. Letters like zayin, vav, and samekh are less representative due to lower training and validation scores. This indicates that these letters either lack strong distinguishing features or that the current feature extraction methods struggle to model their traits effectively.

To assess the effectiveness of the hinge and edge-directional features, we analyzed the mean differences in training scores, validation scores, and the gaps between these scores. Figure 3 intuitively depicts the trends in training and validation scores for each feature extraction method, complementing the results summarized in Table III.

The mean training score for hinge method was 78.9%, while edge-directional achieved a mean score of 67.2%, resulting in a mean difference of 11.7%. This indicates that hinge consistently outperformed edge-directional in capturing patterns within the training data. The higher training scores for hinge suggest that it is more effective at modeling the stylistic differences in the handwriting captured during training, which is crucial for distinguishing between writers. For validation scores, hinge achieved a mean of 61.3%, whereas edge- directional attained a mean score of 58.4%, resulting in a smaller mean difference of 2.9%. While hinge maintains an advantage on unseen data, the reduced gap between the validation scores suggests that edge-directional performs closer to hinge when generalizing to new data. This indicates that edge-directional might generalize more consistently but is overall less accurate. The gap between training and validation scores for hinge was 17.6%, while edge- directional exhibited a smaller gap of 8.8%, with a mean difference of 8.8%. The larger gap for hinge highlights a potential overfitting issue, as its performance on training data was significantly higher than on validation data. In contrast, edge-directional had a smaller gap, suggesting better generalization to new data despite its lower overall accuracy. While hinge demonstrated higher accuracy on both training and validation data, its larger gap between training and validation scores suggests it may be more prone to overfitting. Edge-directional, with its smaller gap, appears to generalize better but sacrifices some level of accuracy. Based on these results, hinge is recommended if achieving the highest possible accuracy is the primary goal, and overfitting can be mitigated through regularization techniques or additional data augmentation. However, if generalization is more critical, edge-directional may be a more robust choice.

Following the interpretation of character-specific training and validation scores, it is imperative to analyze the results of the decision strategy on the building-block approach for determining the probabilities at word and sentence levels. This approach provides a practical perspective on how the trained features translate into the testing phase and helps identify characteristics associated with each scribe’s writing. The results are detailed in Table IV (showing probabilities for Scribe A, Col. XXVII) and Table V (showing probabilities for Scribe B, Col. LIV).

As shown in Table IV, for Scribe A, edge-directional outperformed hinge at sentence- level probabilities under both soft voting and majority voting strategies. Under soft voting, edge-directional achieved a sentence-level probability of 54.4%, which is slightly higher than hinge’s 53.8%. Word-level probabilities for edge-directional under this strategy ranged from 33.2% to 73.6%, indicating a slightly wider spread but with consistently fewer extremely low values compared to hinge. Under majority voting, edge-directional also demonstrated stronger performance with a sentence-level probability of 53.8%, significantly higher than hinge’s 23.1%. This marked difference indicates that edge-directional, despite variability in word-level probabilities, aggregates better in majority voting.

It can be inferred from Table V that Scribe B presented a different trend. Under soft voting, edge-directional again slightly outperformed hinge, with a sentence-level probability of 49.0% compared to 48.6%. However, the margin of difference was narrower for Scribe B than for Scribe A. Word-level probabilities for edge- directional ranged from 39.0% to 60.7%, showing less spread compared to Scribe A’s data and a more consistent performance. For majority voting, hinge surprisingly outperformed edge- directional, with a sentence-level probability of 44.4% compared to 33.3%. This indicates that hinge may be more robust when handling shorter sentence structures, as Scribe B’s dataset consists of only 9 words per sentence compared to Scribe A’s 13 words.

Analyzing the results further, several key observations emerge when comparing the impact of sentence length and voting strategies on edge-directional and hinge features:

Sentence Length Impact: Scribe B’s shorter sentences (9 words) amplify the impact of low word-level probabilities on the aggregated sentence-level result. This effect is particularly evident in majority voting, where the edge-directional feature’s performance drops significantly for Scribe B (33.3%) compared to Scribe A (53.8%).

Edge-directional’s Stability: Edge-directional consistently outperformed hinge under soft voting for both scribes, indicating that edge-directional is more stable when aggregating probabilities using this strategy. However, its performance under majority voting varies more, particularly for shorter sentences (Scribe B).

Hinge’s Robustness in Majority Voting: While hinge performance was lower than edge-directional, it showed better sentence-level probabilities for Scribe B under majority voting. This suggests that hinge may handle extreme variability at the word level better than edge- directional in certain conditions.

The results emphasize the importance of selecting an appropriate voting strategy and feature representation. Soft voting appears more stable across scribes and features, making it a preferred approach when building sentence-level probabilities. Edge-directional emerges as the stronger candidate overall due to its higher sentence-level probabilities and more consistent word-level predictions under soft voting. However, the lower performance of both features for Scribe B highlights challenges in handling shorter sentences. Shorter units of analysis amplify inconsistencies in word-level predictions, especially under majority voting, where extreme values (e.g., 0%) can disproportionately affect the aggregated result.

The relatively low probability scores observed in the analysis can be attributed to several underlying factors related to the data, feature extraction, and the nature of the scribes’ handwriting. Two key possibilities contributing to these results are discussed below.