When small signs raise big questions

Historical manuscripts often appear at first glance as texts that only need to be read. A page is digitized, an AI system or HTR model analyzes it, and at the end a transcription is produced.

It is not that simple. Medieval and early modern manuscripts do not consist only of letters in the modern sense. They contain abbreviation marks, ligatures, superscript forms, special signs, damaged glyphs, uncertain minim clusters, marginal signs, corrections and scribal habits that cannot simply be translated into modern characters.

This is where one of the major challenges of digital manuscript analysis begins.

The problem is not only reading

Many digital systems are designed to produce text from an image as quickly as possible. For many document types that is useful. For historical manuscripts, however, it can become problematic.

Automatic transcription often answers only one question:

What text might stand here?

For source-bound research that is not enough. The more important question is:

Which visible findings lead to this reading?

Between image and text lies an entire analytical path. A line may be part of a letter, an abbreviation sign, a correction, a stain, the remnant of a damaged stroke, or a form that only becomes intelligible through comparison with other positions.

If a system skips this intermediate layer, an uncertain visual finding can quickly become an apparently secure word.

Small signs disappear first

Very small sign forms are especially critical: abbreviation strokes, dots, small hooks, superscript signs, tildes, nasal bars, thin connecting strokes, small correction signs and hard-to-separate minim sequences.

Such details are often only a few pixels wide. When an image is reduced, compressed or processed as a working copy, precisely this information may be weakened, smoothed or lost completely.

For a broad page description this is often unproblematic. Layout, columns, larger text blocks and margins usually remain visible in reduced representations. But for deciding whether a tiny form above a line is a relevant abbreviation mark, a reduced image version is often insufficient.

A sign is not automatically a character

Another problem lies in digital encoding. What is visible in the manuscript is first a concrete image form. That image form is not automatically identical with a modern Unicode character.

For historical texts, additional reference systems are relevant, including MUFI, the Medieval Unicode Font Initiative. MUFI provides recommendations and character references for medieval special signs, especially where standard Unicode coverage is insufficient or where signs are located in the Private Use Area.

This is important for digital editions. But the rule remains:

A MUFI or Unicode assignment does not replace analysis of the source.

A sign may be encoded correctly and still be displayed incorrectly if the necessary font is missing. Another sign may look visually similar but serve a different function in the source. Sometimes the visible finding is simply not sufficient for a secure decision.

visible manuscript form ≠ modern Unicode character ≠ MUFI codepoint ≠ font rendering ≠ secure reading

Why AI must be treated carefully here

AI models can work impressively well with images and text. They can describe manuscript pages, recognize structures, propose transcriptions and place historical texts into broader contexts.

They also have a structural weakness: they tend to transform uncertainty into fluent language.

A model can turn a difficult sign into a plausible reading. That reading may be linguistically convincing even though the visual basis is not secure enough.

For historical manuscripts, the decisive question is not only whether a reading sounds probable. The decisive question is whether it can be checked against the image evidence.

A scholarly usable analysis must therefore do more than output text. It must show where a reading is secure, where it remains uncertain and which visible places require special attention.

From image to auditable reading

Digital manuscript analysis therefore requires a different view of the process. Not:

Image → automatic transcription

but:

Image → visible finding → striking sign forms → uncertainty → possible reading

The decisive step takes place before transcription. First, it has to be checked which areas of a source are visually critical.

These may be places where small signs stand above the line. Or areas where several minims lie so close together that modern systems easily add a word without really justifying the individual signs. Or special forms that later have to be encoded cleanly for a digital edition.

Such places should not be silently turned into smooth readings. They should be marked, described and kept visible as uncertain or in need of review.

Why this matters for non-specialists

Manuscript analysis is not only a topic for specialists. More and more historical sources are digitally accessible, and many users work with manuscript images without being trained palaeographers.

A good digital tool should not only say: here is the text. It should also say: this place is visually striking; the reading is not fully secured; an alternative interpretation or special-character encoding may be possible.

This does not overwhelm users. It gives orientation. They do not need to be experts in medieval abbreviations, Unicode or MUFI to see where caution is required.

Documentation instead of false certainty

The greatest danger of automatic analysis is not the error itself. Errors can be checked and corrected. More dangerous is an output that leaves no trace of the path by which it was produced.

If the analyzed image version, the critical place, the certainty of the sign and the relation between visual finding and reading remain invisible, false certainty is created.

For scholarly work this is problematic. For digital editions it is even more problematic. An edition is not only readable text; it is a documented decision about a source.

Digital manuscript analysis must therefore make its own limits visible.

HistoriaMP and source-bound analysis

HistoriaMP follows this approach. The focus is not rapid text generation, but the traceable path from image to reading.

The project understands historical manuscripts not as a simple OCR task, but as a multi-step analysis:

Source → image finding → layout → segment → glyph → uncertainty → reading → quality control

HistoriaMP does not generate historical truth. It makes visible what a reading rests on.

This is especially important where small signs have large consequences: abbreviations, special signs, minim clusters and uncertain glyph forms.

Conclusion

Historical manuscripts pose a special challenge for digital systems. It is not enough to extract text from an image as quickly as possible.

The decisive issue is whether the path from visible finding to reading remains traceable.

Small signs, abbreviations and special forms show why digital palaeography needs more than automatic text recognition. It needs image fidelity, clean documentation, uncertainty marking and a strict distinction between visible source, technical encoding and interpreted reading.

The most important question is therefore not only:

What does the system read?

But before that:

What does the system really see - and what does the reading rest on?

This is exactly where source-bound manuscript analysis begins.