GPT Review — P10A-2B v0.2 Corrupted Prompt

Date: 2026-04-29

Verdict

BLOCKED — do not dispatch.

P10A-2B v0.2 as pasted is corrupted/duplicated and not safe for execution. The five requested guardrails are conceptually correct, but the final prompt text is malformed.

Blocking defects

1. The prompt body is duplicated midway through Phase 1.
2. The UUID regex line is truncated/corrupted: ^[0-9a-f]{8}-...-[0-9a-f]{12}# P10A-2B...
3. Render command is split/corrupted: -F appears incomplete, and \t appears after the document footer.
4. Section numbering is duplicated/inconsistent: 3.2, 3.3, then later another 3.2.
5. Footer repeats P10A-2B v0.1 after a v0.2 prompt.
6. PASS criteria conflict: one table says Preflight 4/4, another says 5/5.
7. The duplicated block creates ambiguity about which instruction is final.

Law / constitutional check

• Zero Trust: blocked. A malformed execution prompt cannot be considered certainly correct.
• Điều 38 / Điều 33: goal remains aligned, but execution must not proceed with corrupted instructions.
• Gate separation: still valid if corrected; no P10A-2B dispatch until fixed.

Directive to Opus 4.6

Produce a clean P10A-2B v0.3 prompt. Do not add new scope.

Required fixes:

1. Remove all duplicated text.
2. Repair UUID verification:
   • parse pub_id from insert output;
   • verify with full UUID regex;
   • confirm exactly 1 row in tac_publication.
3. Repair render command:
   • save raw TSV to /tmp/p10a-2b-render-raw.tsv;
   • use a valid separator such as -F $'\t' or equivalent shell-safe quoting.
4. Use one coherent Phase 3 numbering.
5. Footer must say P10A-2B v0.3.
6. PASS criteria must match the actual preflight count: PF-1 to PF-4 if there are four PFs, or rename to PF-5 if SHA is separate.
7. Keep the five prior guardrails:
   • backup/safety evidence;
   • stronger schema/triggers/collision check;
   • robust pub_id extraction;
   • no automatic rollback after committed mismatch;
   • explicit raw render to markdown conversion.

After v0.3 is clean, return it for GPT final review. No Agent dispatch until GPT authorizes.

Current state

• P10A-2A package: PASS accepted.
• P10A-2B v0.2: blocked due to corrupted prompt text.
• P10A-2B execution: not authorized.