Rebuilding Thayer's Greek Lexicon, entry by entry, against the 1889 page scans

Thayer's Greek–English Lexicon has been public domain for over a century, and no first-class digital text of it exists in the open. We know because we swept every natural habitat before starting — the Bible-app ecosystem, commercial Bible software, the open-data community on GitHub, continental scholarship, Perseus, the research libraries, Wikisource — and recorded why each candidate failed.

The failures share a fingerprint. We test every candidate at one entry, βάπτισμα, against the printed page, which reads …Acts i. 22; x. 37; xviii. 25; [xix. 3]… — and the freely circulating digitisations all render it Acts 18:25; (); — the bracketed reference dropped, its punctuation left orphaned. One ancestral digitisation made that mistake decades ago, and everything downstream inherited it. Other datasets are not Thayer at all but gloss-sets and outlines wearing his name. In machine-readable form, the only complete Thayer that existed was a scan of ink on paper: the Harper & Brothers 1889 Corrected Edition, 716 body pages of dense, two-column Greek and English.

So we rebuilt it from the page scans. All of it. Here is how, because the how is the reason you can trust the result — and it is the same discipline behind everything Darash serves.

The model kept remembering the wrong Thayer

Transcription started the obvious way: show a vision model a page, get the text back. It failed in an unexpected and telling way — the model's memorisation guard kept firing. The widely-replicated abridged Thayer is all over every model's training data, so when the model saw a real 1889 page, it recognised something close enough to a memorised text and refused to transcribe, to avoid reciting.

Think about what that means: the corrupted version of this book is so pervasive that an AI cannot look at the original without being haunted by the copy. That is the problem in one image.

The escape was tiling: cut each column into as many as sixteen pieces, because a shorter span matches a memorised one less often. Four transcription routes in escalating order — whole page, then column, then tiles, then a second, independent model family entirely — each page taking the cheapest route that worked. Pages no engine could deliver were transcribed by hand, each hand file carrying the exact crop command that produced the view, so a reviewer can regenerate precisely what the transcriber saw.

Two witnesses, and judges paid to break it

Every page was then transcribed twice, by unrelated model families, and the two witnesses collated: 98.91% agreement across 856,793 tokens. Agreement is not truth, but disagreement is a finding — the collation cut the review queue from 716 pages to 16 pages plus a ranked list of individual spots.

Then came the part most digitisation projects skip. We ran adversarial judges against the work — independent reviewers instructed to break the claims, not confirm them. One refused to accept our own derived data as evidence that no dictionary entries had been lost, on the correct grounds that you cannot audit a thing with an artifact built from it. It went instead to three signals the transcription never touched — the printed running heads, an alphabetical-gap test whose own recall it measured first, and a third model as an independent reader — and put the no-lost-entries claim at odds of about 10²³ to one.

The judges found real defects. The best one: a page where the scanner's ink read ἐνπ- for ἐντ-, which caused a matching cursor to jump — and seven entries silently vanished into their neighbour, one short article swelling to 8,295 characters. A summary line had reported this as out of order: 33, which reads like a sorting nit. It was data loss. A count is not a check — that sentence is now written on the project.

The governing rule

One rule governed every repair: never write a character that is not on the page. A fix was allowed only where the mapping had no judgement in it, or where the correct reading was taken off the page image by eye and recorded as such. Everything else stayed flagged. There is one digit in a Josephus reference that no eye could resolve on the scan — it is preserved as "?", because an honest question mark outranks a confident guess.

Finally, every one of the 5,866 entries was read in full against its page image — six readers, the great entry for ἄν made whole at 7,312 characters, block scrambles on two pages relocated to where the printer actually put them — until the invariants held: zero duplicated entries, zero empty ones, zero silent drops, text conservation proven end to end.

Why a Bible company does this

Not for the love of digital humanities. We had to do this simply to be able to provide Thayer to people at all — there was no first-class text to license, buy, or borrow. When the open ecosystem offers only a defective lineage and gloss-sets wearing the name, the only way to hand readers the real book is to make it exist.

And because trust in a research engine is not a marketing claim — it is a property of the pipeline that filled it. When your AI assistant reads the Greek with you through Darash and cites Thayer, the citation has to be Thayer — the 1889 text, not a gloss-set wearing his name. As far as our documented source sweep found, this is the only first-class digital Thayer in the open. It is entering Darash now, held back until every entry had survived verification, because a Thayer that isn't Thayer helps nobody.

The same standard runs through the rest of the estate — the lexicons, the morphology, the 4,200+ published editions — and it is the standard we bring when we build for churches, too. Rigor is not a feature we added. It is the company.