The Eye Jumped to the Wrong Line

A scribe copying by hand looks at the exemplar, looks down at the copy, writes, and looks back up.

The moment of looking back up is where the text gets damaged, and it does not get damaged randomly.

If two nearby lines end with the same word, the eye returning to the page can land on the second occurrence instead of the first. Everything between them disappears, silently, with the copy reading perfectly well. Textual critics call the trigger homoeoteleuton — similar ending — and the general act of losing your place parablepsis, looking aside. Homoeoarchy is the mirror case, where the two lines begin alike.

Both are omissions. Both are also, and this is the useful part, self-documenting: the missing span is exactly the distance between the two similar strings. An editor who spots the repeated ending can reconstruct not just that something was dropped, but approximately how much, and why the eye went where it did.

The error carries a description of its own cause.

A taxonomy, not a category #

Textual criticism does not have a concept called “copying error.” It has a list, and each entry names a trigger.

Dittography is writing twice what appeared once. Haplography is writing once what appeared twice — the same mechanism failing in opposite directions, both triggered by repeated material sitting close together. Iotacism is substitution driven by sound: Greek vowels that had merged in pronunciation get swapped by a copyist working from dictation or subvocalising, producing an error impossible for a scribe who was purely copying shapes.

That last one is the tell for the whole approach. Iotacism cannot occur in a script whose sounds have not merged. It is not available. The error requires the language to be in a particular state, and where the language is not in that state, no amount of careless copying produces it.

So the corpus of possible errors is set by the medium, and the corpus of likely errors is set by the medium’s specific properties. Textual criticism records error profile as script- and language-specific for exactly this reason.

What the encoding decides #

Writing systems differ along axes that map directly onto which errors are available.

An abjad writes consonants and leaves vowels to the reader, so a class of vowel errors simply cannot be recorded — and a different class, where the reader supplies the wrong vowel, becomes the dominant failure. A logographic system with thousands of signs has a large inventory and high information per sign, so a single-sign error destroys more and is easier to spot. An alphabet with a small inventory produces errors that are individually cheap and individually inconspicuous.

Inventory size against per-sign density is not a curiosity of script typology. It is a statement about how much a single corruption costs and how likely it is to be noticed.

Which means, given a script, you can predict the error profile before seeing any errors. And given an error profile that does not match the script, something other than ordinary copying has happened.

The same reasoning, applied to formats #

Every data format has an available corruption set, and it is determined the same way.

Fixed-width truncates, silently, at the field boundary, and the loss is invisible because the record still parses. Delimited formats break on embedded delimiters, so the corruption occurs specifically in the records containing the character the format reserved — which is to say, disproportionately in the unusual records rather than at random. Nested formats lose subtrees, and a dropped subtree leaves a document that is still valid.

Each of these has a signature. Fixed-width truncation produces values of exactly the field length. Delimiter breakage produces field counts that are wrong by the number of embedded delimiters. A dropped nested element leaves a parent whose child count disagrees with a sibling’s.

Almost nobody records which one happened.

Data quality reporting counts errors, rates completeness, and produces a percentage. That is a measure of volume, and volume is the least informative property available, because it says nothing about mechanism and therefore nothing about what to fix or what else is wrong.

The equivalent in textual criticism would be an editor noting that a manuscript contains 340 errors and stopping. No editor does that, because the distribution of error types is what tells you where the manuscript was copied, from what, by whom, and under what conditions — and because a type appearing that the script should not permit is the single most informative thing a tradition can show you.

The check #

For any loss, ask what the two boundaries have in common.

If the span that vanished begins and ends at similar-looking content, the cause is structural and will recur at every other place with that property. If losses cluster in records containing one particular character, the format chose which records to damage. If the distribution has no positional structure at all, then and only then is it random, which is the rarest case and the one usually assumed.

A scribe who drops a line between two identical endings has not been careless in a general way. He has been defeated by a specific configuration, which is present elsewhere in the same text, and which will defeat the next copyist too.

That is why the list has names. Naming the trigger is what makes the next instance findable before it costs anything.