GPT Review — P10B-2B D28 Package Generation Prompt v0.1
GPT Review — P10B-2B D28 Package Generation Prompt v0.1
Date: 2026-04-29
Verdict
DO NOT DISPATCH YET — patch required.
The prompt is directionally correct and has no conceptual conflict with the Constitution / laws after patch. However, it contains one critical input-certainty flaw and several precision gaps.
Legal / constitutional check
Relevant constraints from KB / context-pack red zones:
- No case-dispatch per section.
- No built-in renderer fallback.
- No ALTER TABLE / DDL.
- No manual mutation outside approved gates.
- Zero Trust: if input is not certainly right, it is wrong.
P10B-2B is read-only package generation, so the scope is acceptable if the prompt is patched.
Blockers / required patches
B1 — Critical: candidate-units-r2.json location is not guaranteed
The prompt assumes:
/tmp/p10b-2a-r2/candidate-units-r2.json
But the Agent report for R2 said it ran from local Mac and only /tmp/p10b-2a-r2/* ephemeral files were created. P10B-2B targets VPS/DB and cannot assume that local /tmp file exists on the same host.
Patch:
- Do not treat
/tmp/p10b-2a-r2/candidate-units-r2.jsonas canonical. - Canonical input must be the KB report:
knowledge/dev/laws/dieu38-trien-khai/reports/p10b-2a-d28-hash-r2-evidence-2026-04-29.md. - Agent must reconstruct/export
candidate-units-r2.jsonfrom the KB report only if the report includes full body for all 27 units; otherwise STOP and request the JSON artifact upload. - Better: require a separate KB artifact path for
candidate-units-r2.json, or embed full JSON in the report. No package generation from missing/tmpfile.
B2 — Require full-body availability check before SQL generation
P10B-2B must not generate SQL from hash/excerpt only.
Patch:
- Verify candidate JSON has full
bodyfor all 27 units. - Verify every body SHA equals R2 per-unit hash.
- Verify JSON SHA equals
e47775e33cc752656468edb287cca7b58539804678443b6c1b1dd03b165de8ad. - If full bodies are not available, STOP. Do not re-segment source and do not infer body boundaries.
B3 — Strengthen SQL literal/body safety
Đ28 contains SQL fences, markdown tables, pipes, quotes, and special characters.
Patch:
- Generator must use a safe SQL literal escaping function for all text fields.
- Report the escaping method.
- Include a body preservation check: count code fences and markdown table rows in candidate JSON vs generated SQL/report, or at least sample S3-P2/S3-P3.
- No manual edits after generation.
B4 — Clarify generated SQL transaction terminator pattern
D32 PF found COMMIT/ROLLBACK were both commented. For Đ28, package must explicitly report the ending pattern.
Patch:
- Generated insert SQL must end with both decision lines commented, same as D32, or clearly documented.
- Report exact tail lines.
- P10B-2C-PF will inspect before execute.
B5 — Full hardcode audit must inspect helper scripts, not only grep narrow patterns
Patch:
- Inspect all
.py/.shhelper scripts. - Allowed: JSON data contains
D38-DIEU28-*. - Not allowed: arrays/case-dispatch in logic such as Roman numeral lists,
if section ==, or manual unit arrays. - Report audit method and findings.
Direction to Opus
Patch P10B-2B prompt v0.1 → v0.2 with B1–B5.
Most important: solve the canonical JSON input problem before dispatch.
After patch, if the prompt requires Agent to fetch candidate JSON from KB/report and verify full body + SHA before generation, it may be marked GPT AUTHORIZED.
No SQL execution. No INSERT. No mutation.