📄 Ugaritic Spelling Errors

Summary

The paper asks how error-prone the scribes of ancient Ugarit were, and whether their slips look like the corruptions that textual critics assume when they emend the Hebrew Bible. No duplicate Ugaritic tablets exist, and the only known ancient translation is loose, so Richardson uses a different check. He compares passages that repeat inside a single tablet written by one scribe. His main case is the Krt legend, where El's instructions in a vision are later retold almost word for word as narration. He also examines repeated passages in the Baal cycle. Each variant is sorted by its level (letter-stroke, letter, word, phrase) and by its effect. A simple error leaves the meaning intact. A complex error produces another valid word that also fits the context.

Errors that cause real misreading turn out to be almost absent. Most slips fall at the lowest levels, and differences at word or phrase level are usually deliberate stylistic choices. This suggests critics should be cautious about conjectural emendation and should give more weight to oral tradition.

Key points

• Ilimilku, a 13th-century BC scribe known for his neat hand, wrote a Baal tablet of about 310 lines. Of these, 180 survive, and they contain at least twenty spelling errors, roughly 3–4% of words.
• For comparison, F. Delitzsch (1920) catalogued over three thousand supposed errors in the Old Testament. The tablets are cited from Herdner's CTA (1963).
• The Krt tablet is the main test case:
  • The vision occupies lines 52–153 and the narration lines 154–300, covering five shared episodes plus two added ones.
  • A seven-day motif appears four times.
  • A bribe motif recurs six times.
• Errors become rarer as the unit grows: letter-constituent slips outnumber letter slips, which outnumber word slips. Rare words were spelled carefully, while mistakes cluster in common words.
• Several apparent "errors" are better read as free variation, dialect-driven phonetic spelling, or poetic restructuring. Examples are a dropped word divider, an omitted copula, and a variable locative /h/. The author warns against emending the parallel passages about Anat and Mot.

Interesting ideas

• I1. When no external copy exists, a text's own repeated passages can act as independent witnesses that confirm or reject a suspected corruption.
• I2. A slip only truly matters if the wrong form is itself a real word that fits the context; otherwise it is harmless noise.
• I3. Variants are classified as surface-only or deep changes; dropping a copula or word divider leaves the underlying structure untouched.
• I4. Which error type dominates reveals how the text was produced. Duplications point to slow, careful copying, while omissions point to hasty writing by someone who understood the text.
• I5. Real scribal errors usually hide a common word, so "correcting" a common form into a rare one is suspect.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Repeated or paraphrased sentences within one document could serve as an extra agreement check for silver parses, flagging cases where parallel clauses receive inconsistent trees.
• I2 (tokens): The spelling-repair step could rank candidate typos by whether the misspelled form is a valid in-context word, and spend careful disambiguation only on those.
• I3 (back): When scoring the text → graph → text compressor, count only differences that alter the meaning graph, and treat such surface variation as equivalent.
• I4 (learn): Profiling whether a parser mostly drops or duplicates attachments could diagnose its failure mode and guide which self-induced repair rules to try.
• I5 (tokens): The normalizer should prefer frequent lemmas as the likely originals behind odd spellings, rather than reaching for rare vocabulary.