diff --git a/README.md b/README.md
index b0fd1ee..73172a8 100644
--- a/README.md
+++ b/README.md
@@ -99,8 +99,10 @@ ATS API를 probe해 치과 키워드가 실제로 있는 보드만 등록한다.
 검증 전에 **role gate**(제목 규칙)가 보조/비임상 직무를 걸러낸다 — Dental Assistant·Hygienist·
 Coordinator·Payable 등은 수집·검증 단계에서 FILTERED. ortho 신호는 페이지 전체가 아니라
 **제목 + 공고 설명 영역**에서만 확인하므로, 치과 고용주 회사소개 문구("... & Orthodontics")로 인한
-오탐이 없다. 경계 타이틀은 오프라인에서 `role-audit`(gemma4)이 검토해 용어를 제안한다 — 런타임은
-여전히 non-AI이며, 제안 용어는 `sites/role_terms.auto.yaml`로 병합된다.
+오탐이 없다. 경계 타이틀은 런 후 오프라인으로 `role-audit`(gemma4)이 자동 검토한다 —
+support/non_clinical 제안은 `sites/role_terms.auto.yaml`에 자동 반영되고, target 제안은 게이트
+완화 위험이 있어 리포트(`workspace/manifests/role-audit-*.md`)에만 기록된다. 런타임 수집·검증은
+여전히 non-AI이며, Ollama가 없으면 audit만 조용히 건너뛴다.
 
 ```bash
 uv run gimme-job proactive run                    # 1회 실행 (발견+검증+저장+보고)
diff --git a/gimme_job/cli.py b/gimme_job/cli.py
index f21797a..cbee965 100644
--- a/gimme_job/cli.py
+++ b/gimme_job/cli.py
@@ -493,6 +493,9 @@ def proactive_run(
     if result.report_path:
         uri = Path(result.report_path).resolve().as_uri()
         console.print(f"Report: [link={uri}]{result.report_path}[/link]")
+    if result.role_audit_path:
+        uri = Path(result.role_audit_path).resolve().as_uri()
+        console.print(f"Role audit: [link={uri}]{result.role_audit_path}[/link]")
 
 
 @proactive_app.command("reverify")
@@ -691,7 +694,9 @@ def proactive_role_audit(
     date_: Optional[str] = typer.Option(
         None, "--date", help="Audit date YYYY-MM-DD (default: today)"
     ),
-    model: str = typer.Option("gemma4:26b-mlx", "--model", help="Ollama model"),
+    model: Optional[str] = typer.Option(
+        None, "--model", help="Ollama model (default: plan role_audit.model)"
+    ),
     limit: int = typer.Option(60, "--limit", help="Max distinct titles to review"),
     titles: Optional[str] = typer.Option(
         None, "--titles", help="Comma-separated titles (skip the daily candidate log)"
@@ -707,8 +712,11 @@ def proactive_role_audit(
     from datetime import date as _date
     from datetime import datetime as _dt
 
+    from gimme_job.proactive.plan import ProactivePlan
     from gimme_job.proactive.role_audit import append_auto_terms, run_audit
 
+    plan = ProactivePlan.load()
+    model = model or plan.role_audit.model
     run_date = _dt.strptime(date_, "%Y-%m-%d").date() if date_ else _date.today()
     title_list = (
         [t.strip() for t in titles.split(",") if t.strip()] if titles else None
diff --git a/gimme_job/proactive/engine.py b/gimme_job/proactive/engine.py
index 16b6cd0..b182972 100644
--- a/gimme_job/proactive/engine.py
+++ b/gimme_job/proactive/engine.py
@@ -79,6 +79,7 @@ class ProactiveRunResult:
     engine_blocks: int = 0
     errors: list[str] = field(default_factory=list)
     report_path: Optional[str] = None
+    role_audit_path: Optional[str] = None
 
     def finish(self) -> None:
         self.ended_at = datetime.utcnow()
@@ -143,6 +144,7 @@ class ProactiveEngine:
                 self._persist(verified, result)
                 self._record_run(result)
                 self._deliver_report(result, plan)
+                self._run_role_audit(plan, result)
         finally:
             bm.close()
 
@@ -691,4 +693,49 @@ class ProactiveEngine:
                 except Exception as e:
                     logger.debug(f"[proactive] Report path update failed: {e}")
         except Exception as e:
-            logger.error(f"[proactive] Report delivery failed: {e}")
\ No newline at end of file
+            logger.error(f"[proactive] Report delivery failed: {e}")
+
+    def _run_role_audit(self, plan: ProactivePlan, result: ProactiveRunResult) -> None:
+        """Daily offline LLM pass over borderline titles (never blocks the run).
+
+        gemma4 reviews unknown/support/non_clinical candidates logged during
+        discovery. Support/non-clinical proposals are auto-applied to the
+        machine-owned role_terms.auto.yaml; target proposals stay in the audit
+        report for human review (they would loosen the gate).
+        """
+        cfg = plan.role_audit
+        if not cfg.enabled:
+            return
+        try:
+            from gimme_job.proactive.html_report import terminal_link
+            from gimme_job.proactive.role_audit import append_auto_terms, run_audit
+
+            path, parsed, entries = run_audit(
+                result.run_date,
+                model=cfg.model,
+                base_url=cfg.base_url,
+                max_titles=cfg.max_titles,
+            )
+            result.role_audit_path = str(path)
+            logger.info(
+                f"[proactive] Role audit: {len(entries)} candidates → "
+                f"{terminal_link(path, path)}"
+            )
+
+            proposals = parsed.get("proposals") or []
+            applicable = [p for p in proposals if p.get("role") in cfg.apply_roles]
+            if cfg.write and applicable:
+                auto_path, added = append_auto_terms(applicable)
+                logger.info(
+                    f"[proactive] Role audit: {added} term(s) applied → {auto_path}"
+                )
+            skipped = len(proposals) - len(applicable)
+            if skipped:
+                logger.info(
+                    f"[proactive] Role audit: {skipped} proposal(s) report-only "
+                    "(target/unknown — review the report)"
+                )
+        except ValueError as e:
+            logger.debug(f"[proactive] Role audit skipped: {e}")
+        except Exception as e:
+            logger.warning(f"[proactive] Role audit failed (non-fatal): {e}")
\ No newline at end of file
diff --git a/gimme_job/proactive/plan.py b/gimme_job/proactive/plan.py
index 7c3e819..89613b4 100644
--- a/gimme_job/proactive/plan.py
+++ b/gimme_job/proactive/plan.py
@@ -177,6 +177,19 @@ class ReportConfig(BaseModel):
     retention_days: int = 30
 
 
+class RoleAuditConfig(BaseModel):
+    """Daily offline LLM pass over borderline titles (never in the verify path)."""
+
+    enabled: bool = True
+    model: str = "gemma4:26b-mlx"
+    base_url: str = "http://127.0.0.1:11434"
+    max_titles: int = 60
+    write: bool = True
+    # Proposals for these roles are auto-applied; target proposals stay
+    # report-only to avoid loosening the gate without review.
+    apply_roles: list[str] = Field(default_factory=lambda: ["support", "non_clinical"])
+
+
 class ProactivePlan(BaseModel):
     schedule: ScheduleConfig = Field(default_factory=ScheduleConfig)
     engines: EngineSettings = Field(default_factory=EngineSettings)
@@ -191,6 +204,7 @@ class ProactivePlan(BaseModel):
     boards: BoardsConfig = Field(default_factory=BoardsConfig)
     outreach: OutreachConfig = Field(default_factory=OutreachConfig)
     report: ReportConfig = Field(default_factory=ReportConfig)
+    role_audit: RoleAuditConfig = Field(default_factory=RoleAuditConfig)
 
     @classmethod
     def load(cls, path: Optional[Path] = None) -> "ProactivePlan":
diff --git a/gimme_job/proactive/role_audit.py b/gimme_job/proactive/role_audit.py
index 487944f..bbc928b 100644
--- a/gimme_job/proactive/role_audit.py
+++ b/gimme_job/proactive/role_audit.py
@@ -28,6 +28,14 @@ DEFAULT_MODEL = "gemma4:26b-mlx"
 DEFAULT_BASE_URL = "http://127.0.0.1:11434"
 _VALID_ROLES = {TARGET, SUPPORT, NON_CLINICAL, UNKNOWN}
 
+# Terms too generic to ever apply automatically (would misfire on every title)
+_BLOCKED_TERMS = {
+    "a", "an", "the", "and", "or", "of", "to", "in", "at", "on", "for", "with",
+    "by", "is", "are", "be", "no", "not", "all", "any", "new", "other",
+    "job", "jobs", "career", "careers", "work", "role", "position", "positions",
+    "staff", "team", "member", "full", "part", "time",
+}
+
 
 def build_audit_prompt(entries: list[dict], max_titles: int = 60) -> str:
     lines: list[str] = []
@@ -145,6 +153,9 @@ def append_auto_terms(
         role, term = proposal["role"], proposal["term"]
         if term in DEFAULT_ROLE_TERMS.get(role, []):
             continue  # already covered by code defaults
+        if term in _BLOCKED_TERMS or len(term) < 2:
+            logger.debug(f"[role-audit] blocked generic term: {term!r}")
+            continue
         current = terms.setdefault(role, [])
         if term not in current:
             current.append(term)
diff --git a/gimme_job/proactive/scheduler.py b/gimme_job/proactive/scheduler.py
index 442407e..ea69cab 100644
--- a/gimme_job/proactive/scheduler.py
+++ b/gimme_job/proactive/scheduler.py
@@ -89,6 +89,13 @@ def run_forever(
                     logger.info(
                         f"Report: {terminal_link(result.report_path, result.report_path)}"
                     )
+                if result.role_audit_path:
+                    from gimme_job.proactive.html_report import terminal_link
+
+                    logger.info(
+                        f"Role audit: "
+                        f"{terminal_link(result.role_audit_path, result.role_audit_path)}"
+                    )
             except Exception as e:
                 logger.error(f"Run #{run_count} failed: {e}")
     except KeyboardInterrupt:
diff --git a/sites/proactive.yaml b/sites/proactive.yaml
index b5c2127..4574cfd 100644
--- a/sites/proactive.yaml
+++ b/sites/proactive.yaml
@@ -61,6 +61,15 @@ report:
   html: true
   retention_days: 30           # 한 달 지난 리포트(.html/.md) 자동 삭제
 
+# 매일 런 후 오프라인 LLM(gemma4)으로 경계 타이틀 검토 → 용어 자동 보강
+# (런타임 수집/검증 경로에는 관여하지 않음. Ollama 미실행 시 조용히 skip)
+role_audit:
+  enabled: true
+  model: gemma4:26b-mlx
+  max_titles: 60
+  write: true                  # 제안 용어를 sites/role_terms.auto.yaml에 반영
+  apply_roles: [support, non_clinical]   # target 제안은 리포트에만 (게이트 완화 방지)
+
 states:
   - Alabama
   - Alaska
diff --git a/sites/role_terms.auto.yaml b/sites/role_terms.auto.yaml
new file mode 100644
index 0000000..03b28e3
--- /dev/null
+++ b/sites/role_terms.auto.yaml
@@ -0,0 +1,7 @@
+# 자동 제안 role 용어 — `gimme-job proactive role-audit --write`가 생성/갱신.
+# 수동 편집 금지: 직접 추가하려면 sites/proactive.yaml의 verification.role_terms.
+
+role_terms:
+  non_clinical:
+  - it
+  - business
diff --git a/tests/proactive/test_role_audit.py b/tests/proactive/test_role_audit.py
index 924b3fe..63fb2e4 100644
--- a/tests/proactive/test_role_audit.py
+++ b/tests/proactive/test_role_audit.py
@@ -47,9 +47,10 @@ def test_append_auto_terms_dedupes(tmp_path):
         {"role": "support", "term": "surgical tech", "reason": "r"},
         {"role": "non_clinical", "term": "revenue cycle", "reason": ""},
         {"role": "target", "term": "dentist", "reason": "already a code default"},
+        {"role": "non_clinical", "term": "the", "reason": "generic stopword"},
     ]
     _, added = role_audit.append_auto_terms(proposals, path=path)
-    assert added == 2  # duplicate + code-default terms skipped
+    assert added == 2  # duplicate + code-default + stopword terms skipped
     _, added2 = role_audit.append_auto_terms(proposals, path=path)
     assert added2 == 0
 
@@ -73,3 +74,66 @@ def test_run_audit_writes_report(monkeypatch, tmp_path):
     assert "Role audit" in path.read_text(encoding="utf-8")
     assert entries[0]["title"] == "Orthodontic Clinician I"
     assert parsed["proposals"] == []
+
+
+# ── Engine daily hook ────────────────────────────────────────────────────────
+
+
+def test_engine_role_audit_applies_only_safe_roles(monkeypatch, tmp_path):
+    from gimme_job.proactive.engine import ProactiveEngine, ProactiveRunResult
+    from gimme_job.proactive.plan import ProactivePlan
+
+    report = tmp_path / "role-audit-2026-09-13.md"
+    parsed = {
+        "items": [],
+        "proposals": [
+            {"role": "support", "term": "surgical tech", "reason": ""},
+            {"role": "target", "term": "clinician", "reason": ""},
+        ],
+    }
+    monkeypatch.setattr(
+        role_audit, "run_audit", lambda *a, **k: (report, parsed, [{"title": "x"}])
+    )
+    applied = {}
+
+    def fake_append(proposals, path=None):
+        applied["proposals"] = proposals
+        return tmp_path / "auto.yaml", len(proposals)
+
+    monkeypatch.setattr(role_audit, "append_auto_terms", fake_append)
+
+    engine = ProactiveEngine(global_config=None)
+    result = ProactiveRunResult()
+    engine._run_role_audit(ProactivePlan(), result)
+    assert result.role_audit_path == str(report)
+    assert [p["term"] for p in applied["proposals"]] == ["surgical tech"]
+
+
+def test_engine_role_audit_disabled(monkeypatch):
+    from gimme_job.proactive.engine import ProactiveEngine, ProactiveRunResult
+    from gimme_job.proactive.plan import ProactivePlan
+
+    def boom(*a, **k):
+        raise AssertionError("run_audit should not be called")
+
+    monkeypatch.setattr(role_audit, "run_audit", boom)
+    plan = ProactivePlan()
+    plan.role_audit.enabled = False
+    engine = ProactiveEngine(global_config=None)
+    result = ProactiveRunResult()
+    engine._run_role_audit(plan, result)
+    assert result.role_audit_path is None
+
+
+def test_engine_role_audit_skips_without_candidates(monkeypatch):
+    from gimme_job.proactive.engine import ProactiveEngine, ProactiveRunResult
+    from gimme_job.proactive.plan import ProactivePlan
+
+    def no_candidates(*a, **k):
+        raise ValueError("no candidates logged for this date")
+
+    monkeypatch.setattr(role_audit, "run_audit", no_candidates)
+    engine = ProactiveEngine(global_config=None)
+    result = ProactiveRunResult()
+    engine._run_role_audit(ProactivePlan(), result)
+    assert result.role_audit_path is None