chore(governance): add ROUTE-003 minimality trap and expand adversarial holdout

Close the reflective-minimality gap in ROUTE-003 with a dedicated
minimality_not_implement_trap group (8 phrases), thicken thin adversarial
groups (+8 phrases), sync fixture minimums to 15 groups / 53 phrases, and
add cheatsheet parity + anti-drift probe coverage. Router unchanged.

Author: johnteee

reflective-prompt-library/PROJECT_KNOWLEDGE.md

@@ -72,6 +72,7 @@ deferred promotions are recurrence-gated — see [panel backlog](plans/multi-age
 
 ## Decision Index
 
+- 2026-06-25 Post-Round 68 holdout expansion — ROUTE-003 minimality trap + thin adversarial group refresh (15 groups / 53 phrases) → [QUALITY_GATES_SUMMARY.md](plans/QUALITY_GATES_SUMMARY.md)
 > Pointers to the causal trail — plans, reflections, tests, commits. Detail is
 > not duplicated here; this is a map, not an archive.
 

reflective-prompt-library/plans/QUALITY_GATES_SUMMARY.md

@@ -208,7 +208,7 @@ python3 reflective-prompt-library/plans/validate_skill_examples.py
 - Post-panel maintenance: ROUTING_CONTRACT **R11** approved-spec delivery (`implement_not_plan_trap`) at 100%
 
 **Results:**
-- Tested 14 adversarial groups with 37 paraphrases
+- Tested 15 adversarial groups with 53 paraphrases
 - Overall consistency: 100.0% (passes Phase-1 threshold >=80% and aspirational target >=90%)
 - Low-confidence route trace coverage: 100.0%
 - Output: `plans/route-003-results.json`
@@ -224,7 +224,7 @@ python3 reflective-prompt-library/plans/route_paraphrase_eval.py reflective-prom
 
 **What it does:**
 - Enforces minimum ROUTE-001/002/003 group and phrase counts before paraphrase eval runs
-- Current minimums: ROUTE-001 (12 intent + 4 adversarial groups, 128 phrases); ROUTE-002 (36 holdout groups, 102 phrases); ROUTE-003 (14 adversarial groups, 37 phrases)
+- Current minimums: ROUTE-001 (12 intent + 4 adversarial groups, 128 phrases); ROUTE-002 (36 holdout groups, 102 phrases); ROUTE-003 (15 adversarial groups, 53 phrases)
 - Round 22 panel compromise: deterministic hygiene without YAML dependency explosion
 - Integrated in `make validate` after skill examples gate; mirrored by pytest in `test_validate_route_fixture.py`
 - Cheatsheet parity anti-drift: `test_cheatsheet_boundary_quick_cues.py` (R12 quick-cue summary), `test_cheatsheet_*_parity.py` (holdout probe cues in EN/zh-TW cheatsheets)
@@ -280,7 +280,7 @@ The implementation aligns with research findings:
 | Benchmark tasks | 24 | ✅ Ready |
 | Routing consistency | 100.0% | ✅ Passes ROUTE-001 expanded boundary eval |
 | Holdout routing consistency | 100.0% | ✅ ROUTE-002 (36 groups, 102 paraphrases) |
-| Adversarial routing consistency | 100.0% | ✅ ROUTE-003 (14 groups, 37 paraphrases) |
+| Adversarial routing consistency | 100.0% | ✅ ROUTE-003 (15 groups, 53 paraphrases) |
 | Skill example coverage | 9/9 | ✅ validate_skill_examples.py |
 | Linting errors | 0 | ✅ Clean |
 
@@ -304,14 +304,14 @@ The latest router improvement uses concept-level boundary rules for these cases
 
 ### Holdout Tracking
 
-ROUTE-002 measures unseen phrasing separately from ROUTE-001. Round 7 (2026-06-25) added Traditional Chinese holdout groups with matching router intent keywords — fairness test without full `SKILL.md` translation. Round 65 expanded to 36 holdout groups; post-Round 68 maintenance added ROUTING_CONTRACT **R11** boundaries (102 ROUTE-002 phrases, 14 ROUTE-003 adversarial groups / 37 phrases including `implement_not_plan_trap`, `approved_spec_plan_not_implement_trap`, and `dispatch_meta_skill_trap`) at 100%. Treat this as a seeded holdout, not proof of broad semantic routing; add fresh cases before further router tuning.
+ROUTE-002 measures unseen phrasing separately from ROUTE-001. Round 7 (2026-06-25) added Traditional Chinese holdout groups with matching router intent keywords — fairness test without full `SKILL.md` translation. Round 65 expanded to 36 holdout groups; post-Round 68 maintenance added ROUTING_CONTRACT **R11** boundaries (102 ROUTE-002 phrases, 15 ROUTE-003 adversarial groups / 53 phrases including `implement_not_plan_trap`, `approved_spec_plan_not_implement_trap`, `dispatch_meta_skill_trap`, and `minimality_not_implement_trap`) at 100%. Treat this as a seeded holdout, not proof of broad semantic routing; add fresh cases before further router tuning.
 
 ## Phase 2 Status (post-Round 68 maintenance)
 
 ### Done ✅
 
 1. **CI/CD** — `.github/workflows/python-tools.yml` runs `make all` on push/PR
-2. **ROUTE-001/002/003 in CI** — 128 + 102 + 37 paraphrases at 100% consistency (seeded fixtures); `validate_route_fixture.py` gates minimum coverage
+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` (190+ pytest anti-drift suite in CI)

reflective-prompt-library/plans/route-003-adversarial-eval.yaml

@@ -46,6 +46,8 @@ adversarial_sets:
     phrases:
       - verify production auth change will not expose secrets
       - check rollback before touching production billing data
+      - verify production deploy will not leak secrets
+      - production auth change rollback plan before merge
 
   - name: plain_review_not_risk
     expected_workflow: reflective-review
@@ -55,6 +57,8 @@ adversarial_sets:
     phrases:
       - inspect this patch for regressions before merge
       - review the behavior changes carefully in the diff
+      - check this patch for readability not security
+      - inspect patch for regressions not production deploy risk
 
   - name: multi_voice_not_dispatch
     expected_workflow: reflective-research
@@ -64,6 +68,8 @@ adversarial_sets:
     phrases:
       - six-lens debate on whether to merge these skills
       - multi-perspective strategic rethink with explicit persona lenses
+      - six-lens panel debate with explicit persona lenses
+      - source-backed library comparison before choosing
 
   - name: mixed_language_implement
     expected_workflow: reflective-implement
@@ -99,6 +105,8 @@ adversarial_sets:
     phrases:
       - lessons learned retro after this sprint
       - sprint retrospective with lessons learned
+      - capture lessons learned and promotion candidates from this sprint
+      - session handoff summary with reusable rules for next agent
 
   - name: implement_not_plan_trap
     expected_workflow: reflective-implement
@@ -146,6 +154,21 @@ adversarial_sets:
       - which reflective workflow skill fits a routing-only mixed intent
       - which reflective workflow skill should handle routing-only mixed intent
 
+  - name: minimality_not_implement_trap
+    expected_workflow: reflective-minimality
+    expected_rigor: medium
+    expected_enhancements:
+      - complexity_cut_list
+    phrases:
+      - challenge whether this abstraction layer is needed
+      - run a minimality gate before adding another helper
+      - can we delete this wrapper and use stdlib instead
+      - avoid overengineering this small formatting helper
+      - which dependencies can we remove without breaking tests
+      - cut unnecessary abstraction before shipping the feature
+      - dependency removal review for this module
+      - 少寫程式 surface area in this module
+
   - name: brief_not_plan_trap
     expected_workflow: reflective-brief
     expected_rigor: low

reflective-prompt-library/plans/tests/test_quality_gates_summary.py

@@ -98,6 +98,7 @@ def test_holdout_tracking_paragraph_matches_fixture_minimums(summary_text: str):
     ) in section
     assert "approved_spec_plan_not_implement_trap" in section
     assert "dispatch_meta_skill_trap" in section
+    assert "minimality_not_implement_trap" in section
     assert "100 phrases" not in section
 
 

reflective-prompt-library/plans/tests/test_validate_route_fixture.py

@@ -108,6 +108,12 @@ def test_round_51_boundary_probes():
     "which reflective workflow skill should handle routing-only mixed intent",
 )
 
+MINIMALITY_NOT_IMPLEMENT_TRAP_PROBES = (
+    "dependency removal review for this module",
+    "which dependencies can we remove without breaking tests",
+    "cut unnecessary abstraction before shipping the feature",
+)
+
 ROUTE_003_ADVERSARIAL_BOUNDARY_PROBES = (
     ("design a handoff workflow specification without runtime code", "reflective-spec-plan"),
     ("patch the trivial null check in code", "reflective-implement"),
@@ -247,3 +253,16 @@ def test_dispatch_meta_skill_trap_covers_probes():
     missing = [p for p in DISPATCH_META_SKILL_TRAP_PROBES if p not in fixture_phrases]
     assert not missing, f"dispatch_meta_skill_trap missing probes: {missing}"
 
+def _minimality_not_implement_trap_phrases() -> set[str]:
+    config = load_route_eval_config(PLANS / "route-003-adversarial-eval.yaml")
+    for group in config["adversarial_sets"]:
+        if group["name"] == "minimality_not_implement_trap":
+            return set(group.get("phrases", []))
+    raise AssertionError("missing ROUTE-003 minimality_not_implement_trap group")
+
+
+def test_minimality_not_implement_trap_covers_probes():
+    """Anti-drift: ROUTE-003 minimality boundary probes live in dedicated trap group."""
+    fixture_phrases = _minimality_not_implement_trap_phrases()
+    missing = [p for p in MINIMALITY_NOT_IMPLEMENT_TRAP_PROBES if p not in fixture_phrases]
+    assert not missing, f"minimality_not_implement_trap missing probes: {missing}"

reflective-prompt-library/plans/validate_route_fixture.py

@@ -31,8 +31,8 @@
 ROUTE_002_MIN_HOLDOUT_GROUPS = 36
 ROUTE_002_MIN_PHRASES = 102
 
-ROUTE_003_MIN_ADVERSARIAL_GROUPS = 14
-ROUTE_003_MIN_PHRASES = 37
+ROUTE_003_MIN_ADVERSARIAL_GROUPS = 15
+ROUTE_003_MIN_PHRASES = 53
 
 REQUIRED_TRACE_FIELDS = {
     "canonical_intent",

reflective-prompt-library/skills/SKILL_TRIGGER_CHEATSHEET.md

@@ -137,6 +137,7 @@ Trigger cues:
 - "Review this diff only for complexity we can cut."
 
 - "What dependencies can we remove from this module?"
+- "Dependency removal review for this module."
 
 Do not use when:
 

reflective-prompt-library/skills/SKILL_TRIGGER_CHEATSHEET.zh-TW.md

@@ -135,6 +135,7 @@ Minimality 訊號掃描（skill 內建）：
 - 只想從 diff 裡找可刪的複雜度
 
 - 「What dependencies can we remove from this module？」
+- 「Dependency removal review for this module。」
 
 不要用在：