# ScriptureLeaf V2 Reader Guide

**Study:** five-translation V2 research release
**Translations:** BSB, ESV, NASB 2020, CSB 2020, NIV 2011
**Corpus:** 160 passages: 120 difficulty-enriched challenge passages and 40 controls

## How to read the study

V2 reports separate translation dimensions rather than one overall score. OpenAI and Claude reviewed the frozen instrument independently; their results remain separate. Agreement is a robustness signal, not truth, and disagreement is not a penalty against a translation.

The corpus is deliberately enriched for difficult translation decisions. Results describe this study corpus and must not be read as whole-Bible prevalence estimates.

## What each construct means

- **Textual Basis** — Which recognised source-text reading a rendering appears to follow; descriptive, not a quality score.
- **Semantic Representation** — How source meaning is represented, including preserved/functional-equivalent, narrowing, broadening and material loss/change.
- **Morphosyntax** — Meaning-bearing grammatical structure; functional restructuring can preserve meaning and is not penalised for different word order.
- **Interpretive Explicitness** — How much the English makes implicit or ambiguous source relations explicit; more explicit is not automatically better.
- **Ambiguity Handling** — Whether a live source ambiguity is retained, favoured but open, resolved, displaced or replaced by a new ambiguity.
- **Lexical Continuity** — How consistently the same prospectively frozen source sense is rendered across linked passages, allowing justified contextual variation.
- **Intertextual Visibility** — How visible an OT→NT source relationship remains without demanding maximal verbal identity.
- **Reader Disclosure** — Whether exact-edition official notes disclose textual or interpretive alternatives; companion evidence only.
- **English Clarity** — Source-blind ratings of syntactic clarity, referential clarity and idiomatic naturalness, reported separately.
- **Reviewer Sensitivity** — Differences between OpenAI and Claude judgements, treated as methodological uncertainty rather than translation performance.

## Why OpenAI and Claude are shown separately

The two AI reviewer systems used the same frozen measurement contract independently. V2 does not average them into a consensus result. Where their distributions differ, the public presentation should make that reviewer sensitivity visible.

## Challenge passages and controls

The 120 challenge passages belong to eight prospectively frozen strata. The 40 controls are reported separately and are not treated as a ninth challenge stratum. The pooled 160-passage view is descriptive of the study corpus only.

## How to inspect a case

Every worked case identifies the canonical Bible reference, translation or translation pair, reviewer system, construct/opportunity, frozen outcome, and source-evidence references. The case package deliberately excludes restricted translation wording. A BibleHub comparison link is supplied so readers can inspect current published wording externally.

## How translation philosophy is used

Publisher philosophy and intended-use statements come only from first-party sources frozen before scoring. They are added after the translation results are frozen and are descriptive context only. V2 does not assign a formal-to-dynamic number, philosophy-fit score, or causal effect.

## What the study cannot tell you

V2 does not identify a best translation, create an overall accuracy percentage, combine clarity with fidelity, infer whole-Bible prevalence, or replace human biblical-language scholarship. It is an independent AI-assisted, reproducible research project whose evidence and limitations are intended to be inspectable.