Round 88: 00-core Human Review preamble guards

johnteee · johnteee · commit acf0df42567c · 2026-06-25T16:45:22.000+08:00
Add ## Human Review preambles routing to reflective-risk on six
risk-bearing 00-core prompts; extend test_core_prompts_eval_harness
with shared prompt_eval_helpers guards. GLOSSARY playbook step 20.
diff --git a/README.md b/README.md
@@ -21,7 +21,7 @@ Full library docs: [reflective-prompt-library/README.md](reflective-prompt-libra
 ## Governance
 
 - **Contributing:** [CONTRIBUTING.md](CONTRIBUTING.md) — quality gates, routing maintenance (R8–R12), `make all`
-- **Panel record:** [multi-agent-panel-consensus](reflective-prompt-library/plans/multi-agent-panel-consensus-2026-06-25.md) — six-lens Socratic consensus (Rounds 1–87)
+- **Panel record:** [multi-agent-panel-consensus](reflective-prompt-library/plans/multi-agent-panel-consensus-2026-06-25.md) — six-lens Socratic consensus (Rounds 1–88)
 - **Operator playbook:** [GLOSSARY.md](reflective-prompt-library/GLOSSARY.md) — Governance Maintenance Playbook
 
 The repository contains:
diff --git a/reflective-prompt-library/00-core/core-full.md b/reflective-prompt-library/00-core/core-full.md
@@ -18,6 +18,11 @@ Canonical full English protocol for reflective engineering hosts. Primary workfl
 
 Every recommendation names what observation would prove it wrong.
 
+## Human Review
+
+Escalate to `reflective-risk` with an explicit Human Review gate when the work implies irreversible or high-blast-radius action.
+
+
 ## Identity
 
 You are a Reflective Engineering Agent.
diff --git a/reflective-prompt-library/00-core/core-minimal.md b/reflective-prompt-library/00-core/core-minimal.md
@@ -20,6 +20,11 @@ Shortest general-purpose reflective engineering opener. Primary workflow surface
 
 State what evidence would overturn the current framing before deeper work.
 
+## Human Review
+
+Escalate to `reflective-risk` with an explicit Human Review gate when the work implies irreversible or high-blast-radius action.
+
+
 ```markdown
 請以「反思型工程代理人」處理此任務。先判斷真正目標，再定義假設、範圍、輸入/輸出、失敗條件、驗收標準與可證偽測試。若模糊但安全，明確列出假設並繼續；若涉及架構、安全、隱私、資料遺失、金錢或不可逆決策，請要求人工審查。回答時優先給乾淨交付成果，請勿輸出未經整理的原始推理過程。結構化推理段落（Goal/Assumptions/Socratic audit 等）是要求的輸出格式，不屬於隱藏思考鏈。
 ```
diff --git a/reflective-prompt-library/00-core/core-short.md b/reflective-prompt-library/00-core/core-short.md
@@ -20,6 +20,11 @@ Global short instruction surface for host custom instructions. Primary workflow
 
 Name one observation that would prove the recommended plan wrong before execution.
 
+## Human Review
+
+Escalate to `reflective-risk` with an explicit Human Review gate when the work implies irreversible or high-blast-radius action.
+
+
 ```markdown
 你是 Reflective Engineering Agent。核心原則是：Doing the right thing > doing things right。
 
diff --git a/reflective-prompt-library/00-core/custom-instruction-en.md b/reflective-prompt-library/00-core/custom-instruction-en.md
@@ -20,6 +20,11 @@ Length-limited English custom instruction distillate. Primary workflow surface:
 
 Name one check that would prove the answer wrong after delivery.
 
+## Human Review
+
+Escalate to `reflective-risk` with an explicit Human Review gate when the work implies irreversible or high-blast-radius action.
+
+
 ```markdown
 Act as a Reflective Engineering Agent: Doing the right thing > doing things right. For non-trivial tasks, define goal, assumptions, scope, inputs/outputs, failure conditions, acceptance criteria, falsifiability, plan, implementation, validation, and self-check. Prefer tests, schemas, types, examples, and artifacts over vague prompt rules. If ambiguity is safe, state assumptions and proceed. If it affects architecture, security, privacy, data loss, cost, or irreversible decisions, request Human Review. Do not dump raw, unfiltered reasoning tokens. Structured reasoning sections (Goal/Assumptions/Socratic audit/etc.) are the required output format and are not hidden chain-of-thought. Provide clean deliverables and concise reasoning summaries.
 ```
diff --git a/reflective-prompt-library/00-core/custom-instruction-zh.md b/reflective-prompt-library/00-core/custom-instruction-zh.md
@@ -20,6 +20,11 @@ Length-limited Traditional Chinese custom instruction distillate. Primary workfl
 
 Name one check that would prove the answer wrong after delivery.
 
+## Human Review
+
+Escalate to `reflective-risk` with an explicit Human Review gate when the work implies irreversible or high-blast-radius action.
+
+
 ```markdown
 請以「反思型工程代理人」執行：做正確的事大於把事情做正確。
 
diff --git a/reflective-prompt-library/00-core/important-task-full.md b/reflective-prompt-library/00-core/important-task-full.md
@@ -21,6 +21,11 @@ High-rigor reflection for important decisions. Primary workflow surfaces: `refle
 
 State the observation or experiment that would overturn the recommended plan.
 
+## Human Review
+
+Escalate to `reflective-risk` with an explicit Human Review gate when the work implies irreversible or high-blast-radius action.
+
+
 ```markdown
 請以 Reflective Engineering Agent + Socratic Reviewer + Critical Thinking Auditor 模式處理。
 
diff --git a/reflective-prompt-library/GLOSSARY.md b/reflective-prompt-library/GLOSSARY.md
@@ -337,7 +337,7 @@ Curated top-of-cheatsheet summary of high-confusion routing traps (ROUTE-002 hol
 
 ## Governance Maintenance Playbook / 治理維護手冊
 
-Ongoing upkeep after panel close (Rounds 1–87). Not agent instructions — operator checklist.
+Ongoing upkeep after panel close (Rounds 1–88). Not agent instructions — operator checklist.
 
 **Operational test:** Before router tuning, add fresh ROUTE-002/003 holdout phrases; run `make all`; record decisions in `PROJECT_KNOWLEDGE.md` Decision Index when governance surface changes.
 
@@ -360,3 +360,4 @@ Ongoing upkeep after panel close (Rounds 1–87). Not agent instructions — ope
 17. When editing composable prompts (`00-core`–`06-repo`), keep `Primary workflow surface(s)` / Supporting-lens preamble lines and run `test_*_prompts_eval_harness.py` primary-surface guards.
 18. When adding or editing composable prompts (`02-engineering`–`06-repo`) with `## Human Review`, keep preamble escalation routed to `reflective-risk` and run Human Review guards in `test_*_prompts_eval_harness.py` (exact heading match via `prompt_eval_helpers.py`).
 19. When editing Human Review guards, use `prompt_eval_helpers.assert_human_review_preamble` in all `test_*_prompts_eval_harness.py` files (thinking lenses + composable categories).
+20. When adding or editing risk-bearing `00-core/` prompts with `## Human Review`, keep preamble escalation routed to `reflective-risk` and run `test_core_prompts_eval_harness.py` Human Review guards via `prompt_eval_helpers.py`.
diff --git a/reflective-prompt-library/PROJECT_KNOWLEDGE.md b/reflective-prompt-library/PROJECT_KNOWLEDGE.md
@@ -73,6 +73,7 @@ deferred promotions are recurrence-gated — see [panel backlog](plans/multi-age
 ## Decision Index
 
 - 2026-06-25 Round 85 panel — composable prompt Primary workflow surface preamble guards (`test_*_prompts_eval_harness.py`) + Supporting-lens exemption → [record](plans/multi-agent-panel-consensus-2026-06-25.md)
+- 2026-06-25 Round 88 panel — `00-core` Human Review preamble guards on risk-bearing prompts + `test_core_prompts_eval_harness.py` → [record](plans/multi-agent-panel-consensus-2026-06-25.md)
 - 2026-06-25 Round 87 panel — Human Review helper DRY + GLOSSARY playbook step repair → [record](plans/multi-agent-panel-consensus-2026-06-25.md)
 - 2026-06-25 Round 86 panel — composable Human Review preamble guards + `reflective-risk` routing alignment → [record](plans/multi-agent-panel-consensus-2026-06-25.md)
 - 2026-06-25 Round 84 panel — `00-core` Primary workflow surface parity + primary-line trim → [record](plans/multi-agent-panel-consensus-2026-06-25.md)
diff --git a/reflective-prompt-library/README.md b/reflective-prompt-library/README.md
@@ -30,7 +30,7 @@ Pick **Strictness L1–L6** first (`skills/reflective-dispatch/SKILL.md`, [GLOSS
 
 ## Governance Panel Record
 
-Multi-agent Socratic consensus on project goals and the nine skills (Rounds 1–87, options A–FO) is recorded in [plans/multi-agent-panel-consensus-2026-06-25.md](plans/multi-agent-panel-consensus-2026-06-25.md). Run `make all` before claiming routing or governance changes are verified.
+Multi-agent Socratic consensus on project goals and the nine skills (Rounds 1–88, options A–FS) is recorded in [plans/multi-agent-panel-consensus-2026-06-25.md](plans/multi-agent-panel-consensus-2026-06-25.md). Run `make all` before claiming routing or governance changes are verified.
 
 ## Directory Map
 
diff --git a/reflective-prompt-library/plans/QUALITY_GATES_SUMMARY.md b/reflective-prompt-library/plans/QUALITY_GATES_SUMMARY.md
@@ -314,7 +314,7 @@ ROUTE-002 measures unseen phrasing separately from ROUTE-001. Round 7 (2026-06-2
 2. **ROUTE-001/002/003 in CI** — 128 + 102 + 53 paraphrases at 100% consistency (seeded fixtures); `validate_route_fixture.py` gates minimum coverage
 3. **Governance validators** — links, lint, governance metadata, PROJECT_KNOWLEDGE, benchmark fixture, skill examples
 4. **Harness policy docs** — CONTRIBUTING, AGENTS, SKILL_INSTALLATION, maintenance playbook
-5. **Doc anti-drift** — `test_routing_contract.py`, cheatsheet parity tests, `test_readme_governance.py`, `test_thinking_prompts_eval_harness.py`, `test_engineering_prompts_eval_harness.py`, `test_prompt_cross_links.py`, `test_core_prompts_eval_harness.py`, `test_agent_prompts_eval_harness.py`, `test_context_prompts_eval_harness.py`, `test_domain_prompts_eval_harness.py`, `test_repo_prompts_eval_harness.py`, `test_validate_governance.py`, `test_validate_links.py`, `test_lint_skills.py`, `test_skill_module_contract.py` (Escalation subsection + Trigger/Methods/Output/Never; 550+ pytest anti-drift suite in CI); reciprocal thinking-lens ↔ skill checks and `00-core` + composable `Primary workflow surface(s)` ↔ `*_SKILL_LINKS` parity in `test_prompt_cross_links.py` (including strict Primary workflow surfaces parity via `test_thinking_lens_primary_surfaces_match_consumer_graph`); Human Review + Escalation route-target guards in thinking/skill contract tests; composable `Primary workflow surface(s)` / Supporting-lens preamble guards and composable `## Human Review` preamble guards (route to `reflective-risk`) via `prompt_eval_helpers.assert_human_review_preamble` in `test_*_prompts_eval_harness.py`
+5. **Doc anti-drift** — `test_routing_contract.py`, cheatsheet parity tests, `test_readme_governance.py`, `test_thinking_prompts_eval_harness.py`, `test_engineering_prompts_eval_harness.py`, `test_prompt_cross_links.py`, `test_core_prompts_eval_harness.py`, `test_agent_prompts_eval_harness.py`, `test_context_prompts_eval_harness.py`, `test_domain_prompts_eval_harness.py`, `test_repo_prompts_eval_harness.py`, `test_validate_governance.py`, `test_validate_links.py`, `test_lint_skills.py`, `test_skill_module_contract.py` (Escalation subsection + Trigger/Methods/Output/Never; 560+ pytest anti-drift suite in CI); reciprocal thinking-lens ↔ skill checks and `00-core` + composable `Primary workflow surface(s)` ↔ `*_SKILL_LINKS` parity in `test_prompt_cross_links.py` (including strict Primary workflow surfaces parity via `test_thinking_lens_primary_surfaces_match_consumer_graph`); Human Review + Escalation route-target guards in thinking/skill contract tests; composable `Primary workflow surface(s)` / Supporting-lens preamble guards and composable `## Human Review` preamble guards (route to `reflective-risk`) via `prompt_eval_helpers.assert_human_review_preamble` in `test_*_prompts_eval_harness.py`
 
 ### Ongoing maintenance (not blockers)
 
@@ -384,4 +384,4 @@ Phase 1 quality-gate tooling and documentation are **complete**. Routing consist
 - ✅ Benchmark fixture gate plus optional manual benchmark runs
 - ✅ Research-backed design decisions
 
-The project is positioned to grow sustainably with quality discipline built in from the start. **No open implementation blockers** remain from panel Rounds 1–87; work is recurrence-gated maintenance per playbook. The next measurable quality target is **holdout expansion before router tuning** and optional manual baseline-vs-skill benchmark runs — not shipping new core skills without promotion evidence.
+The project is positioned to grow sustainably with quality discipline built in from the start. **No open implementation blockers** remain from panel Rounds 1–88; work is recurrence-gated maintenance per playbook. The next measurable quality target is **holdout expansion before router tuning** and optional manual baseline-vs-skill benchmark runs — not shipping new core skills without promotion evidence.
diff --git a/reflective-prompt-library/plans/multi-agent-panel-consensus-2026-06-25.md b/reflective-prompt-library/plans/multi-agent-panel-consensus-2026-06-25.md
@@ -2455,3 +2455,46 @@ User directive (repeat): review prompts, plans, skills, and Socratic/critical-th
 
 **Resealed 2026-06-25** after **Round 87** (options FK–FO). Human Review guards now share one helper across thinking lenses and composable prompts; GLOSSARY playbook formatting anti-drift closed. Holdout expansion remains recurrence-gated maintenance.
 
+---
+
+## Round 88 — `00-core` Human Review preamble guards (2026-06-25)
+
+**Options FP–FS** | Six-lens panel (Opus, Codex, Gemini, Composer, Sakana, GLM)
+
+### Round 88 options
+
+| ID | Proposal | Verdict |
+| --- | --- | --- |
+| FP | `## Human Review` preamble on risk-bearing `00-core/` prompts + `test_core_prompts_eval_harness.py` guards | **Agree** |
+| FQ | GLOSSARY playbook step 20 + governance sync | **Agree** |
+| FR | ROUTE holdout expansion | **Defer** |
+| FS | Router / tenth skill / benchmark CI | **Reject** |
+
+### Round 88 verdict table
+
+| ID | Option | Verdict | Action |
+| --- | --- | --- | --- |
+| FP | Core Human Review preambles | **Agree** | six risk-bearing prompts; shared `prompt_eval_helpers` guard |
+| FQ | Playbook + docs | **Agree** | step 20; panel round 88 sync |
+| FR | Holdout expansion | **Defer** | maintenance |
+| FS | Router/tenth skill/benchmark CI | **Reject** | no change |
+
+**All roles agree.**
+
+## Implemented Changes (Round 88)
+
+- `00-core/{core-minimal,core-short,core-full,custom-instruction-en,custom-instruction-zh,important-task-full}.md`: `## Human Review` preamble routes to `reflective-risk`
+- `plans/tests/test_core_prompts_eval_harness.py`: Human Review guards via `prompt_eval_helpers`
+- `GLOSSARY.md`: playbook Rounds 1–88; step 20 for `00-core` Human Review guards
+- `QUALITY_GATES_SUMMARY.md`: `00-core` HR guard note; panel Rounds 1–88; 560+ pytest floor
+- `PROJECT_KNOWLEDGE.md`: Decision Index Round 88 entry
+- `README.md`, `reflective-prompt-library/README.md`, `test_readme_governance.py`: panel round 88 sync
+
+## Verification (Round 88)
+
+- `make all`: pytest + ROUTE-001/002/003 100%
+
+## Panel status (updated)
+
+**Resealed 2026-06-25** after **Round 88** (options FP–FS). Full prompt library now has Human Review preamble guards on thinking lenses (R81), composable prompts (R86), and risk-bearing `00-core` prompts (R88). Holdout expansion remains recurrence-gated maintenance.
+
diff --git a/reflective-prompt-library/plans/tests/test_core_prompts_eval_harness.py b/reflective-prompt-library/plans/tests/test_core_prompts_eval_harness.py
@@ -6,8 +6,10 @@
 import pytest
 
 sys.path.insert(0, str(Path(__file__).parent.parent))
+sys.path.insert(0, str(Path(__file__).parent))
 
 from eval_harness import EvalHarness  # noqa: E402
+from prompt_eval_helpers import assert_human_review_preamble, prompts_with_human_review  # noqa: E402
 
 CORE_DIR = Path(__file__).parent.parent.parent / "00-core"
 REPO_ROOT = str(Path(__file__).parent.parent.parent.parent)
@@ -21,6 +23,7 @@
 )
 
 CORE_PROMPTS = tuple(sorted(CORE_DIR.glob("*.md")))
+CORE_PROMPTS_WITH_HUMAN_REVIEW = prompts_with_human_review(CORE_PROMPTS)
 
 
 @pytest.fixture(scope="module")
@@ -65,3 +68,11 @@ def test_core_prompts_have_primary_workflow_surfaces_line():
         assert "Primary workflow surface" in preamble, (
             f"{prompt_path.name} Purpose should list Primary workflow surface(s)"
         )
+
+
+@pytest.mark.parametrize(
+    "prompt_path", CORE_PROMPTS_WITH_HUMAN_REVIEW, ids=lambda p: p.name
+)
+def test_core_prompt_has_human_review_section(prompt_path: Path):
+    """Risk-bearing 00-core prompts declare Human Review escalation outside zh-TW templates."""
+    assert_human_review_preamble(prompt_path)
diff --git a/reflective-prompt-library/plans/tests/test_glossary_structure.py b/reflective-prompt-library/plans/tests/test_glossary_structure.py
@@ -30,10 +30,11 @@ def test_round_boundary_terms_present(glossary_text: str):
         assert heading in glossary_text, f"missing glossary section: {heading}"
 
 
-def test_maintenance_playbook_references_round_87(glossary_text: str):
+def test_maintenance_playbook_references_round_88(glossary_text: str):
     playbook = glossary_text.split("## Governance Maintenance Playbook", 1)[1]
-    assert "Rounds 1–87" in playbook
-    assert "Rounds 1–86" not in playbook and "Rounds 1-86" not in playbook
+    assert "Rounds 1–88" in playbook
+    assert "Rounds 1–87" not in playbook and "Rounds 1-87" not in playbook
+
 
 
 def test_maintenance_playbook_steps_on_separate_lines(glossary_text: str):
@@ -42,7 +43,7 @@ def test_maintenance_playbook_steps_on_separate_lines(glossary_text: str):
     assert re.search(r"guards\.\d+\.", playbook) is None, (
         "playbook steps merged without newline between numbers"
     )
-    for step in ("17.", "18.", "19."):
+    for step in ("17.", "18.", "19.", "20."):
         assert step in playbook
 
 
diff --git a/reflective-prompt-library/plans/tests/test_readme_governance.py b/reflective-prompt-library/plans/tests/test_readme_governance.py
@@ -10,8 +10,8 @@
 METHODOLOGY_MAP_EN = Path(__file__).parent.parent.parent / "METHODOLOGY_MAP.md"
 SKILL_MAP = Path(__file__).parent.parent.parent / "skills" / "skill-map.md"
 
-CURRENT_PANEL_ROUND = "87"
-CURRENT_PANEL_OPTIONS = "A–FO"
+CURRENT_PANEL_ROUND = "88"
+CURRENT_PANEL_OPTIONS = "A–FS"
 
 
 @pytest.fixture(scope="module")
