Methodology

How a translation is produced

When you translate a short text (up to 500 characters, in either direction), your browser sends it to our server. The server validates the request, applies rate limits, and forwards the text to a third-party AI model instructed to answer in a strict structured format: a translation, brief notes, a list of uncertain choices, and any modern concepts that have no Old English equivalent. What comes back is shown to you as a draft — a starting point, never a certified translation. Where possible, words in the draft are linked to entries in our dictionary so you can check them yourself.

The three evidence tiers

Every piece of Old English text on this site carries exactly one label:

• Attested example

  Quoted from a manuscript or a scholarly edition, with the work and location named (for example, “Beowulf, line 1”). The label applies to the quotation exactly as printed.

• Adapted learning example

  Edited or constructed to illustrate usage. Useful for learning, but not a historical quotation, and never presented as one.

• AI-generated draft

  Produced by the AI model a moment ago in response to your input. It can contain errors, including subtle ones, and no label changes that.

Labels do not travel. A label describes the text exactly as it was delivered. If you edit the text, or regenerate a result, the new wording does not inherit the old label — an edited attested quotation is no longer an attested quotation.

What the dictionary is — and its limits

The dictionary is a hand-curated set of entries. Each entry cites Bosworth–Toller’s Anglo-Saxon Dictionary and links the corresponding Wiktionary word page, so you can verify every claim against a published source. Draft results link only to this curated index — the AI model cannot invent dictionary citations.

Coverage is deliberately small: a working set of common words, not a complete Old English lexicon. If a word is missing, that is a gap in our coverage, not evidence that the word does not exist.

What we have not done yet

A formal linguistic evaluation of draft quality is still pending. That is why you will find no accuracy percentages on this site: we do not have measured results, and we will not publish numbers we have not earned. Treat every draft as unverified until you have checked it against the cited sources.

How errors get corrected

Report a problem via the contact page: include the word or phrase id (dictionary entries carry an anchor such as #hw-…), what you expected, and the source you checked. A human reviews each report against the cited source before an entry or label changes — corrections are edited, not crowd-guessed.