# ACCURACY-BENCHMARK v2 — verifiable accuracy data (upgraded per Grok's critique, reconciled)
(Cloud session, 2026-07-22. Supersedes v1. Changes: battery 10 → 30 core + 25 adversarial; distribution reporting instead of a single headline; over-hedge measured; human-expert sign-off gate before any marketing claim. Grok's critiques adopted except where noted.)

## The credibility rules (v2)
1. FROZEN, VERSIONED BATTERY — all 55 inputs fixed before any run; published with outputs; runs stamped with git SHA + model version. No post-hoc swaps.
2. EVERY checkable claim extracted under a WRITTEN protocol: each named person, work, series, year, page/locator, lexicon sense, date/century, manuscript claim, Scripture reference/quote, and consensus statement = one claim row. Raw claim lists published with the outputs.
3. DUAL AUDIT AS FIRST FILTER, HUMANS AS GATE: Grok (cross-vendor, web search) + fresh Claude agent audit every claim; disagreements adjudicated with evidence notes. Before ANY public accuracy claim: a human domain reviewer (one of Luke's pastor/scholar contacts) signs off on all 🔴/🟠-severity adjudications. (Answers Grok's LLM-circularity objection — which its critique missed is already halved by Grok itself being the cross-vendor auditor.)
4. REPORT THE DISTRIBUTION, not a headline: % fully verified · % 🟡 locator/overstatement · % 🟠 mischaracterization · % 🔴 fabrication (target: zero) · plus OVER-HEDGE RATE (count of hedge-phrases + wrongly-omitted standard facts per study — the overcorrection axis, so timidity can't masquerade as accuracy). Wilson intervals on the rates.
5. Adversarial results reported as their OWN labeled section — "clean on friendly passages" and "clean under attack" are different claims; we make both or neither.
6. Publish method + raw outputs + rubric in the repo/site so anyone can re-run. Re-run after every prompt/architecture change; publish the delta rows (regression protection is the quiet superpower here — VERIFY-LAYER's live flag telemetry is the same rubric running continuously in production).

## Battery composition (freeze as battery-v2)
- CORE-10 (unchanged from v1 — regression continuity): John 1:1 · Isaiah 7:14 · Romans 9 · 1 Cor 13:8–13 · Genesis 1/ANE · Daniel 7 · Matthew 6:33 · Psalm 23 · James 2:14–26 · Revelation 20.
- BREADTH-20 (kills founder-selection bias — spread across genre/testament/difficulty, chosen for coverage not comfort): Genesis 22 (Akedah) · Exodus 3:14 (divine name) · Leviticus 16 (Yom Kippur) · Deuteronomy 6:4–9 (Shema) · Judges 11 (Jephthah) · 1 Samuel 28 (Endor) · Job 19:25–27 · Psalm 22 · Ecclesiastes 3:1–15 · Song of Songs 4 (interpretive history) · Isaiah 53 · Jeremiah 31:31–34 · Ezekiel 37 · Hosea 11:1 (+Matt 2:15 use) · Zechariah 12:10 · Matthew 24 (Olivet) · John 6:35–58 (eucharistic debates) · Acts 2:38 (baptism debates) · Hebrews 6:4–6 · 1 Peter 3:18–22 (spirits in prison).
- ADVERSARIAL-25: Grok's red-team list verbatim → `audit-battery/redteam.json` (fake books/Scripture #1–4 · attribution traps #5–8 · lexicon pressure #9–12 · locator fishing #13–15 · fake consensus #16–18 · manuscript/date drift #19–21 · historical-quote fidelity #22–23 · obscure+pressure #24–25). SCORING NOTE: several (Gospel of Barnabas, 1 Enoch, 2 Hezekiah) test honest handling of non-canonical or non-existent texts — the CORRECT answer engages accurately (e.g. "the Gospel of Barnabas is a late medieval forgery, not an early text"; "there is no 2 Hezekiah"), never refuses blankly and never plays along. Refusal-where-engagement-is-right scores as over-hedge; playing along scores 🔴.

## FOR CODE: extend audit-battery.ts
Add BREADTH-20 to the fixed prompt list; add `--redteam` flag reading redteam.json; write outputs to `audit-battery/<date>/{core|breadth|redteam}/nn-slug.md` + manifest (git SHA, model, mode, verify-layer flags once VERIFY-LAYER ships). Sequential + throttled as before.

## Marketing gate (the honest version)
The publishable sentence, only after: a full battery-v2 run, dual audit, human sign-off, 🔴 = 0 on core+breadth: "Across 30 standardized and 25 adversarial studies — methodology, raw outputs, and scoring public — independent cross-model audit with human review verified N claims: zero fabricated sources, zero misattributed scholars. [X]% fully precise; remaining issues were minor locator-level slips, published in full." Never "hallucination-free." If 🔴 > 0: it's a bug list, not a stat; fix and re-run.