# Codex H Subreview Synthesis

## Intent

Adversarially review whether the Codex H evidence dossier missed major/critical textual variants, overclaimed the no-critical conclusion, mishandled supplied text as manuscript evidence, or confused SBLGNT with NA28/UBS5.

## Reviewers

- Claude: completed, 9 findings (`0` critical, `6` warnings, `3` nits). Receipt: `outputs/audits/2026-06-03-codex-h/subreview/claude.md`; full plan: `/Users/borker/.claude/plans/reviewer-prompt-template-harmonic-flask.md`.
- Gemini: failed before review because this installed Gemini CLI rejects the wrapper's `--skip-trust` / `skipTrust` argument. Receipt: `outputs/audits/2026-06-03-codex-h/subreview/gemini.md`.

## Act On

- Escalate Heb 4:15 from "fragmentary/non-evidentiary" to "major unresolved candidate." The TEI has Heb 4:15 ending on folio `57v` after "our weaknesses" and Heb 4:16 resuming on folio `58r` at "grace," with no explicit 4:15 gap marker in the extracted `<ab>`. Because the missing clause includes "without sin," the final report must not quietly downgrade it.
- Add scribal-hand provenance for app/correction readings where relevant. Firsthand corrections count differently from later corrector hands.
- Make visible/supplied discipline auditable in prose. The current extraction has verse-level supplied counts but not per-token supplied flags in each diff op; the final report should say this is a limitation and avoid claiming supplied NA28 text as manuscript evidence.
- Add row-level NA28/SBLGNT caveats. SBLGNT is only the mechanical comparator; rows such as Heb 1:3, Heb 2:9, Col 2:2, 1Cor 10:28, and 1Tim 1:17 need "NA28/UBS5 unchecked or known divergence" language.

## Consider

- Re-emit `codex_h_comparison.json` with token-level metadata: `visible`, `supplied`, `source`, `unclear`, `gap`, `hand`, and `folio`.
- Add a coarse `vs_NA28` field to every major row once licensed apparatus access is available: `agrees`, `differs`, `mixed`, or `unchecked`.
- Pull folio images for Heb 4:15/4:16 before any scholarly/public claim that Codex H omits the whole final clause.

## Noted

- Heb 10:34 should mention the `δεσμοῖς`/`δεσμίοις` issue, even if downgraded as spelling/itacism plus supplied text.
- The coverage section should explicitly name missing Pauline material: Romans, Ephesians, Philippians, 2 Thessalonians, and large gaps within preserved letters.
- The funnel from 257 automated variant candidates plus 7 near matches to a small set of material variants should be explained as a filtered triage, not as a complete apparatus.

## Dismissed

- The core conclusion is not overturned. Claude explicitly found the "major variants, no critical findings" conclusion supported on the evidence as classified.

## Agreement Map

- Agreement: The dossier's headline conclusion is supported; problems are auditability and a small number of row-level classifications.
- Disagreement/uncertainty: Heb 4:15 classification needs stronger treatment; SBLGNT-vs-NA28 caveats must be row-specific.
- Missing reviewer: Gemini did not review due CLI argument incompatibility.