It is often iterated that if one only knows how to wield a hammer, most problems start to look like a nail. This effect is known as Maslow’s Hammer or more colloquially as Golden Hammer syndrome (Maslow 1966, 15-16). Maslow’s Hammer encapsulates the essence of a cognitive bias: if one has a limited repertoire of methods or techniques, we might use them in many situations where they are actually inappropriate, and we may overlook or reject alternative approaches that may be more suitable or effective. Usually the value of recognizing this bias is understood simply as “it is useful to consider multiple techniques or problem solving strategies in any situation”. However, there is a less trivial, subtler, and epistemologically more interesting cognitive effect that we want to consider here: apparently knowing a specific technique shapes, to a certain extent, our understanding of any given problem. This is classic Kuhn. Our scientific experience and paradigm condition how we perceive a problem (Khun 2012[1962]). But what if we know two techniques, and both appear equally effective for solving our problem. How do we choose? From a pragmatic perspective we may be tempted to base our methodological choice on context-sensitive aspects such as tractability and resource demands. But where is the epistemological proposition in that? Epistemologists and philosophers of science would rather appeal to theoretical virtues, such as “accuracy, consistency, scope, simplicity, and fruitfulness” (Kuhn 1977:357). The question we thus raise here is whether some methodological Ockham’s razor-like approach is also justified when choosing techniques in digital humanities or whether other epistemological considerations should prevail.

Figure 1, The “Sphere of the Animals” in München, BSB Cgm 312 f. 122r.

The case we discuss in our short paper contribution concerns the transcription of a number of “sortes texts”, also referred to as lot books or books of fortune. These texts represent a popular medieval and early modern divination practice with roots in Late Antiquity (Heiles et al. 2021). Rather than being designed for linear reading, these texts required randomising tools to initiate a series of hints that would guide readers from a specific question, typically about the future, via several steps through the various parts of the manuscript to an answer (Heiles 2018). A particular feature of a subselection of the surviving manuscripts is the use of numerous radial diagrams (see Figure 1 for an example). Earlier editions of these books of fortune have often neglected the intricate layout designs found in the manuscripts, favouring some linearized print representation for pragmatic and economic reasons. Digital editions of these manuscripts may represent a much fuller experience of their materiality and design aspects (Cugliana et al. 2024). It would highly facilitate the production of these digital representations if HTR (Handwritten Text Recognition) technology could support the transcription of these diagrams. However current HTR tools, such as Transkribus (Kahle et al. 2024) or Loghi (Van Koert et al. 2017), do not deal with circular or radial text layouts.

The problem to solve in this case is to linearize the circular diagrams, quite literally: to turn the radial diagrams into horizontal representations. In this case two techniques were tested. The first is a geometrical projection (Mercator projection) that produces the result given in figure 2. Color histogram analysis and Hough transforms (Ballard 1981) using OpenCV (Bradski 2000) are subsequently applied to distinguish the “cells” across which the text is distributed. The second technique involves applying a multi modal large language model (MLLM), Google Gemini 2.5 Flash (Werner and Cugliana 2026). In this case in-image hinting (i.e. coloring one of the radial ‘wedges’ of a circular diagram in combination with a descriptive prompt) led to successful isolation of the individual text components of the diagram (figure 3). Both methods produce results that are equally susceptible to further processing by HTR or MLL models.

[Figure 2, Mercator projection of a radial diagram]

[Figure 3, Radial text wedges from München, BSB Cgm 312, f. 135v. ]

The epistemological issue that we raise in our paper is this: given that both techniques produce equal results, should we prefer one? We suppose we could evaluate both methods using Kuhn’s aspects of epistemological virtue: accuracy, consistency, scope, simplicity, and fruitfulness. As far as we can see, there are no differences between the first three of these virtues for either method. As to simplicity and fruitfulness, the case seems rather ambivalent. It seems unjustified to call MLLMs “simple”. However the mathematics underlying Hough transforms are far from straightforward either. A pragmatic argument could be that the first method, based for the most part on basic trigonometry and well understood Hough transforms, is much less computational intensive and can be deployed on an average capable laptop computer. This democratizes both epistemology and practicality (e.g. in contexts with less or no sophisticated infrastructure and personnel support, where only “minimal computing” is an option). In contrast, using MLLMs requires serious (i.e. third party) compute capacity. This could inspire the epistemological observation that one method might be more open to scrutiny and reuse, as it requires less advanced technologies. In this sense, the first technique might be more “fruitfull” to use Kuhn’s expression. However, we could also argue that the second technique will actually be more fruitful. As it relies on natural language prompting, rather than on mathematical programming, it utilizes concepts directly related to the content domain – that is terminology about the visual structure and design of the actual texts.

We do have two hammers. They are definitely of different shape and type. What our paper demonstrates, is that both make the problem indeed look different. One casts it as a problem of mathematically coding projections and shape analysis, the other as a problem of creating visual examples and verbal descriptions of text layout. Still it is hard to say what these differences entail on an epistemological level exactly, but we can see that “both these approaches contribute something different and potentially useful” (Gillespie et al. 2024). Considering the increasing tendency to delegate tasks to (M)LLMs, we aim to shift the focus to the cognitive value of operationalisation. In the case of applying mathematical reasoning, as in our first approach, the actual process, that is, the execution of the operation, is followed step by step by the scholar. In the second approach presented, by contrast, the effort lies in formulating instructions through language, within an iterative process (which, in a sense, proceeds by trial and error, albeit “informed” trial and error). This effort, however, eventually reaches a limit: the actual realisation of the operation remains, to a certain extent, inaccessible, and the reproducibility of the operation is not always guaranteed. If operationalisation, understood as modelling and as “thinking in practice,” can contribute to the generation of knowledge (Ciula et al. 2023), it is therefore necessary to be aware of the different contributions made by different approaches to operationalisation itself and to carefully assess which “hammer” to choose for one’s own nail. Not only is the nail perceived differently depending on the hammer at hand, but the very act of hammering also takes different forms, each with its own advantages and disadvantages, as we demonstrate in our paper.