Transparency
Revisions
Every verse goes through an independent review pass — a second model rereads the draft against the original source and either signs off or proposes changes. This page reports both: the review coverage (how many verses have been read by a second pair of eyes) and the applied edits that came out of those reviews. Click any edit to read it in context.
How a revision happens
1. Source text
SBLGNT (Greek NT) or WLC (Hebrew OT)
2. GPT-5.4 drafts
Per-verse English with full lexical reasoning
3. Gemini 3.1 Pro reviews
Independent second opinion via Vertex AI, anchored in the source text (earlier passes used Gemini 2.5 Pro)
4. Adjudicator applies
Claude Opus 4.7 on majors; mechanical auto-apply on safe stylistic polish
5. Public commit
Every revision is a git commit with rationale
What counts as a revision?
A revision is a targeted change to a committed verse — an English polish, a restored rhetorical force, a preserved wordplay, or a project-wide consistency fix (like the Χριστός → Messiah normalization). Revisions don't overwrite the original draft; they layer on top, and the previous rendering is kept as a footnote on every revised verse. Detailed policy in REVISION_METHODOLOGY.md.
What do the tiers mean?
Tier 1 is mechanical stylistic polish — awkward English cleaned up where the suggested rewrite shares most of its words with the draft. These are auto-applied with a word-overlap safety check.
Tier 2 is Claude Opus 4.7 adjudicating Gemini's flagged majors — mistranslations, lexical fixes, and grammar losses whose rationale cites specific Hebrew / Greek evidence. A Tier-2 rewrite is only applied if the rationale grounds the change in BDAG, HALOT, BDB, or in source-language grammar.
Anything requiring policy judgment (theological weight, cross-verse consistency) is escalated to a human rather than auto-applied.