Method

How a plus becomes a verdict. Every step leaves a record in the published data.

1. The difference layer

The Septuagint and the Masoretic text were aligned from the CATSS parallel files, which the project uses as an index only; the Greek published here is Swete's. Each Addition was cut into units of one clause or one motif, and given a draft English translation and a one-line summary of what it contributes.

2. The bench

27 works, 2,122 paragraphs, pulled from Sefaria and Hebrew Wikisource at pinned versions, each paragraph carrying its source, licence and date. Composite works are dated per stratum.

3. Search

First a lexical pass: two to four Hebrew retroversions per unit, matched against the bench with weighted word-sequence overlap. Then three reading passes by language models under materially different briefs, each reading every bench paragraph against every unit and keeping anything plausible. Together they produced 1,249 candidate pairs.

4. Verification and rulings

A separate verification pass read each candidate against its unit and ruled it verbatim, motif or unrelated under written rejection criteria: material the Masoretic text already supplies, stock biblical idiom, long paragraphs matched on vocabulary alone, a shared word with a different referent, anything explicable from a different biblical book, a passage that contradicts the plus, and a copy of another witness. The editor spot-checked forty of its confident calls, ruled how each work transmitted what it carries, and bracketed disputed paragraphs as "flagged for further study." An adversarial reviewer then attacked every published claim; its findings were adopted in full.

Verbatim means the witness reproduces two or more of the plus's clauses, in the plus's order, with the same content and at least one rendering of a specific word of the plus that is neither the Masoretic text's nor stock. Language does not matter. Motif means a specific shared detail the Masoretic text does not supply.

Final count: 65 verbatim, 284 motif, 225 flagged, 675 unrelated.

5. How much to trust the motif tier

Not much, pair by pair. In a seeded sample of forty motif pairs the adversarial reviewer rejected 58% (95% interval about 42–72%). The negative controls showed the same leak. The motif pairs are therefore labelled candidates everywhere they appear.

6. Controls

Pairs fixed from the scholarly literature before any search, to test recall (positive), precision (negative), and the machinery itself.

7. Conduits

Each kept parallel carries a judgment about how the material reached the witness: a medieval retranslation from Latin, an Aramaic targum, oral tradition, a lost written intermediary, or a common Hebrew source. No pair on this bench carries the last.