A Roman-period birth declaration appears in Apollo Restore’s research paper with proposed words marked in red. The September preprint describes a model that fills gaps from the surviving Greek around them, without requiring a count of missing characters.[1]
On papyrus, handwriting can compress several letters into a small space. A broken edge may remove the boundary needed to estimate the gap. Apollo can accept an exact count, a range or no length hint.[1]
Why the shape of the question matters
The earlier Ithaca system, published in Nature in 2022, illustrates the established approach. Its restoration test deliberately hid between one and ten characters of surviving inscriptions. Researchers supplied the missing positions, then assessed which letters the model put back. Ithaca offered 20 ranked suggestions and visualised which surrounding text influenced its prediction.[2]
Its examples also show why a plausible completion matters historically. In an Athenian decree, the alternatives included terms for an alliance, an assembly and a relationship between a state and a foreigner. Each could sound appropriate in a political document while implying something different about its contents. The researcher needs to compare the candidate with the particular inscription and its context.[2]
Imagine a torn modern form with “received from” before the tear and “for rent” after it. A surname, a full name and a name followed by an address create different possible completions. Fixing the number of missing characters at the start silently rules some of them out. This is an illustration of the problem, not a translation of an ancient document.
The model can work when the count is absent
On held-out documentary papyri, Apollo matched deliberately hidden text with its first suggestion 33.2% of the time without a size hint, versus 43.2% with the correct count. Those figures give equal weight to gap lengths from one to 20 characters. The scoring ignores spaces, punctuation and numerals.[1]
For a historical editor, that is a useful question to put to the software: what readings become available if the presumed length changes? A shorter suggestion might join two surviving fragments neatly; a longer one might imply that more material has disappeared. Comparing those possibilities makes the assumption about the gap visible.
The Austrian Academy of Sciences has made Apollo available through a free public interface.[3]
A restored name, office or relationship can turn an isolated scrap into a more informative record. Keeping the surviving text and proposed additions distinguishable lets later readers revisit that interpretation as further evidence appears.
Sources & further reading
- Apollo Restore: A Foundation LLM for Historical Greek Optimized for Fill-in-the-Middle Restoration of Ancient Greek Texts, version 4 McGovern and colleagues / arXiv · 7 Oct 2026 · Accessed 19 Oct 2026
- Restoring and attributing ancient texts using deep neural networks Assael and colleagues / Nature · 9 Mar 2022 · Accessed 19 Oct 2026
- AI Speaks Ancient Greek: Apollo Restores 2,000-Year-Old Texts Austrian Academy of Sciences · 23 Sep 2026 · Accessed 19 Oct 2026