GPT Review — P10B-2A-HASH-R2 Prompt

Date: 2026-04-29

Verdict

APPROVE WITH MINOR PATCHES before dispatch.

The prompt correctly fixes the prior mistake: evidence must match the corrected 27-unit candidate before package generation. Agent should run a read-only HASH-R2, not package SQL.

Required patches

P1 — Clarify anti-hardcode scope

The prompt says “NO hardcoded unit key lists” but also includes a full 27-unit table. This is acceptable only if the table is treated as candidate data, not logic.

Patch wording:

The 27-unit table is expected candidate data. Helper scripts must not contain manual arrays/case-dispatch of units. If a script needs the candidate, store it as candidate-units-r2.json and iterate over that JSON.

P2 — Produce candidate-units-r2.json as evidence

Add output requirement:

• Create /tmp/p10b-2a-r2/candidate-units-r2.json with all 27 units, including full body, source span, body_sha256, word_count, section_type, parent, sort_order.
• Compute SHA256 of this JSON and include it in the report.
• Next step P10B-2B must use this JSON as source-of-truth input.

P3 — Vocab must be re-queried if DB is accessible; no silent skip

Patch T4:

Re-query production vocab if DB/VPS access is available. If DB is unavailable, report DB unavailable and cite R1 as fallback evidence, but mark this as warning. Do not claim fresh vocab verification if only citing R1.

P4 — Add hardcode audit for helper scripts

Add final audit:

• Inspect helper scripts for manually enumerated unit arrays or Roman numeral case-dispatch.
• Allowed: candidate-units-r2.json contains unit keys.
• Not allowed: Python/bash logic with ['I','II','III'], if section == 'VII', or arrays of D38-DIEU28-*.

Report audit method and result.

Direction

After P1-P4 patch, dispatch Agent to run P10B-2A-HASH-R2.

No SQL package. No INSERT. No mutation.