diff --git a/app/service/risk/profile_l3.py b/app/service/risk/profile_l3.py
new file mode 100644
index 0000000..7fd508f
--- /dev/null
+++ b/app/service/risk/profile_l3.py
@@ -0,0 +1,164 @@
+"""L3 监测画像写入（B3 · PRD FR-7 / R-05 最小写入）。
+
+合并规则（防降级，PRD FR-7）：monitor_tier 取最高档（normal < watch < high），
+monitor_tags 追加合并不覆盖，risk_score 取 max，computed_at 每次写当前时间
+（列 NOT NULL 必须显式赋值），last_alert_id 联动最新预警。
+
+并发：进程内锁按 customer_id 串行（多进程部署换 Redis SET NX，接口不变）；
+锁超时降级与跨进程竞态由 IntegrityError 兜底——重读合并后再 update，不丢更新。
+"""
+
+from __future__ import annotations
+
+import logging
+from datetime import datetime
+from typing import Any
+
+from sqlalchemy.exc import IntegrityError
+
+from app.repository.risk_repository import RiskRepository
+from app.service.risk.alert_service import _run_locked  # 同包复用聚合锁；B7 收敛至公共原语
+
+logger = logging.getLogger(__name__)
+
+TIER_ORDER = ("normal", "watch", "high")
+ALERT_TYPE_TIER = {
+    "aml": "high",
+    "pattern": "watch",
+    "large_amount": "watch",
+    "freq_trade": "watch",
+    "suitability": "normal",
+}
+AML_PENDING_TAG = "aml_hit_pending_review"
+
+
+def tier_of(alert_type: str) -> str:
+    """alert_type → L3 档位映射（PRD FR-7 固定映射；未知类型拒绝写入）。"""
+    tier = ALERT_TYPE_TIER.get(alert_type)
+    if tier is None:
+        raise ValueError(f"unknown alert_type for L3 mapping: {alert_type}")
+    return tier
+
+
+def highest_tier(a: str, b: str) -> str:
+    """取最高档（防降级核心原语）。"""
+    return a if TIER_ORDER.index(a) >= TIER_ORDER.index(b) else b
+
+
+def merge_l3(
+    existing: dict[str, Any] | None,
+    *,
+    mapped_tier: str,
+    risk_score: int | None,
+    monitor_tags: list[str],
+    last_alert_id: str | None,
+    score_dimensions: dict[str, Any] | None = None,
+    computed_at: datetime | None = None,
+) -> dict[str, Any]:
+    """纯函数：existing 行（repo.get_l3 输出）与新事件合并后的 L3 行。
+
+    existing 为 None 表示新客户首写；risk_score 双方均空时保持 NULL。
+    """
+    if existing is None:
+        tier, old_score, old_tags, old_dims = "normal", None, [], {}
+    else:
+        tier = existing.get("monitor_tier") or "normal"
+        old_score = existing.get("risk_score")
+        old_tags = existing.get("monitor_tags") or []
+        old_dims = existing.get("score_dimensions") or {}
+
+    scores = [int(s) for s in (old_score, risk_score) if s is not None]
+    return {
+        "monitor_tier": highest_tier(tier, mapped_tier),
+        "risk_score": max(scores) if scores else None,
+        "monitor_tags": sorted(set(old_tags) | set(monitor_tags)),
+        "score_dimensions": score_dimensions if score_dimensions is not None else old_dims,
+        "last_alert_id": last_alert_id,
+        "computed_at": computed_at or datetime.now(),
+    }
+
+
+def upsert_profile_l3(
+    customer_id: str,
+    alert_type: str,
+    risk_score: int | None = None,
+    monitor_tags: list[str] | None = None,
+    last_alert_id: str | None = None,
+    score_dimensions: dict[str, Any] | None = None,
+    risk_repo: RiskRepository | None = None,
+    computed_at: datetime | None = None,
+) -> dict[str, Any]:
+    """预警事件 → L3 upsert（B2 预警落库后调用；aml 自动追加待复核标签）。
+
+    返回合并后的 L3 行（含 customer_id）。
+    """
+    repo = risk_repo or RiskRepository()
+    mapped_tier = tier_of(alert_type)
+    tags = list(monitor_tags or [])
+    if alert_type == "aml" and AML_PENDING_TAG not in tags:
+        tags.append(AML_PENDING_TAG)
+
+    def _write(locked: bool) -> dict[str, Any]:
+        existing = repo.get_l3(customer_id)
+        merged = merge_l3(
+            existing,
+            mapped_tier=mapped_tier,
+            risk_score=risk_score,
+            monitor_tags=tags,
+            last_alert_id=last_alert_id,
+            score_dimensions=score_dimensions,
+            computed_at=computed_at,
+        )
+        try:
+            if existing is None:
+                repo.insert_l3(
+                    customer_id,
+                    merged["monitor_tier"],
+                    merged["risk_score"],
+                    merged["score_dimensions"],
+                    merged["monitor_tags"],
+                    merged["last_alert_id"],
+                    merged["computed_at"],
+                )
+            else:
+                repo.update_l3(
+                    customer_id,
+                    merged["monitor_tier"],
+                    merged["risk_score"],
+                    merged["score_dimensions"],
+                    merged["monitor_tags"],
+                    merged["last_alert_id"],
+                    merged["computed_at"],
+                )
+        except IntegrityError:
+            # 首写竞态（锁超时降级/跨进程）：另一线程已 insert，重读合并转更新
+            if existing is not None:
+                raise
+            raced = repo.get_l3(customer_id)
+            merged = merge_l3(
+                raced,
+                mapped_tier=mapped_tier,
+                risk_score=risk_score,
+                monitor_tags=tags,
+                last_alert_id=last_alert_id,
+                score_dimensions=score_dimensions,
+                computed_at=computed_at,
+            )
+            repo.update_l3(
+                customer_id,
+                merged["monitor_tier"],
+                merged["risk_score"],
+                merged["score_dimensions"],
+                merged["monitor_tags"],
+                merged["last_alert_id"],
+                merged["computed_at"],
+            )
+        merged["customer_id"] = customer_id
+        return merged
+
+    return _run_locked(f"l3:{customer_id}", _write)
+
+
+def get_profile_l3(customer_id: str, risk_repo: RiskRepository | None = None) -> dict[str, Any] | None:
+    """L3 只读薄封装（对话线/引擎复用；Redis 缓存待 B7 lifespan 一并接入）。"""
+    return (risk_repo or RiskRepository()).get_l3(customer_id)
diff --git a/tests/test_profile_l3.py b/tests/test_profile_l3.py
new file mode 100644
index 0000000..1a9fca9
--- /dev/null
+++ b/tests/test_profile_l3.py
@@ -0,0 +1,247 @@
+"""profile_l3 单测（B3 · 最高档合并防降级 / AML 后大额不回落 / 并发首单）。
+
+sqlite StaticPool 单连接共享内存库（同 test_alert_service 模式）；
+IntegrityError 兜底用 monkeypatch 模拟跨进程首写竞态。
+"""
+
+from datetime import datetime
+
+import pytest
+from sqlalchemy import create_engine, text
+from sqlalchemy.pool import StaticPool
+
+from app.repository.risk_repository import RiskRepository
+from app.service.risk.profile_l3 import (
+    AML_PENDING_TAG,
+    ALERT_TYPE_TIER,
+    get_profile_l3,
+    highest_tier,
+    merge_l3,
+    tier_of,
+    upsert_profile_l3,
+)
+
+
+@pytest.fixture()
+def env():
+    engine = create_engine(
+        "sqlite://",
+        poolclass=StaticPool,
+        connect_args={"check_same_thread": False},
+    )
+    with engine.begin() as conn:
+        conn.execute(
+            text(
+                """
+                CREATE TABLE customer_profile_l3 (
+                    customer_id VARCHAR(64) PRIMARY KEY,
+                    monitor_tier VARCHAR(16) NOT NULL,
+                    risk_score INTEGER,
+                    score_dimensions TEXT,
+                    monitor_tags TEXT,
+                    last_alert_id VARCHAR(64),
+                    computed_at TIMESTAMP NOT NULL,
+                    updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+                )
+                """
+            )
+        )
+    repo = RiskRepository(engine=engine)
+    yield repo, engine
+    engine.dispose()
+
+
+def _upsert(repo, cid, alert_type, score=None, alert_id=None, tags=None, dims=None):
+    return upsert_profile_l3(
+        cid,
+        alert_type,
+        risk_score=score,
+        monitor_tags=tags,
+        last_alert_id=alert_id,
+        score_dimensions=dims,
+        risk_repo=repo,
+    )
+
+
+def _row(engine, cid):
+    with engine.connect() as conn:
+        return conn.execute(
+            text(
+                "SELECT monitor_tier, risk_score, monitor_tags, last_alert_id, computed_at"
+                " FROM customer_profile_l3 WHERE customer_id = :cid"
+            ),
+            {"cid": cid},
+        ).mappings().one()
+
+
+# ---------- 映射与纯函数 ----------
+
+
+@pytest.mark.parametrize(("alert_type", "tier"), sorted(ALERT_TYPE_TIER.items()))
+def test_alert_type_tier_mapping(alert_type, tier):
+    assert tier_of(alert_type) == tier
+
+
+def test_unknown_alert_type_rejected(env):
+    repo, _ = env
+    with pytest.raises(ValueError):
+        _upsert(repo, "C1", "unknown_type", 70)
+
+
+def test_highest_tier_order():
+    assert highest_tier("normal", "watch") == "watch"
+    assert highest_tier("watch", "high") == "high"
+    assert highest_tier("high", "normal") == "high"
+    assert highest_tier("normal", "normal") == "normal"
+
+
+def test_merge_new_customer():
+    merged = merge_l3(None, mapped_tier="watch", risk_score=70, monitor_tags=["t1"],
+                      last_alert_id="ALT-1")
+    assert merged["monitor_tier"] == "watch"
+    assert merged["risk_score"] == 70
+    assert merged["monitor_tags"] == ["t1"]
+    assert merged["last_alert_id"] == "ALT-1"
+    assert merged["computed_at"] is not None
+
+
+def test_merge_both_scores_none_stays_none():
+    existing = {"monitor_tier": "watch", "risk_score": None, "monitor_tags": [],
+                "score_dimensions": {}}
+    merged = merge_l3(existing, mapped_tier="watch", risk_score=None, monitor_tags=[],
+                      last_alert_id="ALT-2")
+    assert merged["risk_score"] is None
+
+
+# ---------- upsert：首写与合并 ----------
+
+
+def test_first_event_inserts(env):
+    repo, engine = env
+    merged = _upsert(repo, "C1", "large_amount", 70, alert_id="ALT-1")
+    assert merged["monitor_tier"] == "watch" and merged["customer_id"] == "C1"
+    row = _row(engine, "C1")
+    assert row["monitor_tier"] == "watch" and row["risk_score"] == 70
+    assert row["last_alert_id"] == "ALT-1" and row["computed_at"] is not None
+    assert get_profile_l3("C1", risk_repo=repo)["monitor_tier"] == "watch"
+
+
+def test_normal_upgrades_to_watch(env):
+    repo, _ = env
+    _upsert(repo, "C1", "suitability", 90, alert_id="ALT-0")
+    merged = _upsert(repo, "C1", "pattern", 80, alert_id="ALT-1")
+    assert merged["monitor_tier"] == "watch"  # normal → watch 升档
+
+
+def test_watch_does_not_degrade_to_normal(env):
+    """ Suitability 映射 normal：已 watch 的客户不被 suitability 事件拉低。"""
+    repo, engine = env
+    _upsert(repo, "C1", "pattern", 80, alert_id="ALT-1")
+    merged = _upsert(repo, "C1", "suitability", 90, alert_id="ALT-2")
+    assert merged["monitor_tier"] == "watch"
+    assert _row(engine, "C1")["monitor_tier"] == "watch"
+
+
+def test_aml_marks_high_with_pending_review_tag(env):
+    repo, engine = env
+    merged = _upsert(repo, "C1", "aml", 95, alert_id="ALT-1")
+    assert merged["monitor_tier"] == "high"
+    assert AML_PENDING_TAG in merged["monitor_tags"]
+    assert AML_PENDING_TAG in _row(engine, "C1")["monitor_tags"]
+
+
+def test_large_amount_after_aml_does_not_fall_back(env):
+    """B3 验收：AML 后大额 → tier 仍 high、tags 并集、score 取 max。"""
+    repo, engine = env
+    _upsert(repo, "C1", "aml", 95, alert_id="ALT-1")
+    merged = _upsert(repo, "C1", "large_amount", 70, alert_id="ALT-2", tags=["manual_review"])
+    assert merged["monitor_tier"] == "high"
+    assert merged["risk_score"] == 95  # 取 max，不降
+    assert set(merged["monitor_tags"]) == {AML_PENDING_TAG, "manual_review"}
+    assert merged["last_alert_id"] == "ALT-2"  # 联动最新
+    row = _row(engine, "C1")
+    assert row["monitor_tier"] == "high" and row["risk_score"] == 95
+
+
+def test_tags_accumulate_not_overwrite(env):
+    repo, _ = env
+    _upsert(repo, "C1", "pattern", 80, alert_id="ALT-1", tags=["freq"])
+    merged = _upsert(repo, "C1", "pattern", 80, alert_id="ALT-2", tags=["freq"])
+    assert merged["monitor_tags"] == ["freq"]  # set 合并去重
+
+
+def test_score_dimensions_replaced_only_when_passed(env):
+    repo, _ = env
+    _upsert(repo, "C1", "aml", 95, alert_id="ALT-1", dims={"amount": 1})
+    merged = _upsert(repo, "C1", "large_amount", 70, alert_id="ALT-2")
+    assert merged["score_dimensions"] == {"amount": 1}  # 未传保留旧值
+    merged = _upsert(repo, "C1", "large_amount", 70, alert_id="ALT-3", dims={"amount": 2})
+    assert merged["score_dimensions"] == {"amount": 2}  # 传入则替换
+
+
+# ---------- 并发与竞态 ----------
+
+
+def test_concurrent_first_upsert_single_row(env):
+    """并发冒烟：两线程同客户首单（aml + 大额）→ 1 行、high、tags 并集。"""
+    import json
+    from threading import Thread
+
+    repo, engine = env
+    errors = []
+
+    def worker(alert_type, alert_id):
+        try:
+            _upsert(repo, "C1", alert_type, 70, alert_id=alert_id)
+        except Exception as exc:  # pragma: no cover
+            errors.append(exc)
+
+    threads = [
+        Thread(target=worker, args=("aml", "ALT-A")),
+        Thread(target=worker, args=("large_amount", "ALT-B")),
+    ]
+    for t in threads:
+        t.start()
+    for t in threads:
+        t.join()
+    assert not errors, errors
+    with engine.connect() as conn:
+        count, payload = conn.execute(
+            text(
+                "SELECT COUNT(*), GROUP_CONCAT(monitor_tags) FROM customer_profile_l3"
+                " WHERE customer_id = 'C1'"
+            )
+        ).fetchone()
+    assert count == 1
+    tags = json.loads(payload)
+    assert AML_PENDING_TAG in tags
+
+
+def test_integrity_error_falls_back_to_remerge(env, monkeypatch):
+    """跨进程竞态（评审 P2-7③）：首读 None、insert 撞主键 → 重读合并转更新，不丢对方写入。"""
+    repo, engine = env
+    from sqlalchemy.exc import IntegrityError
+
+    # 对方进程已写入 high（模拟 AML 先落库）
+    repo.insert_l3("C1", "high", 95, {}, [AML_PENDING_TAG], "ALT-AML", datetime.now())
+
+    calls = {"n": 0}
+    orig_get = type(repo).get_l3
+
+    def racing_get(self, cid):
+        calls["n"] += 1
+        if calls["n"] == 1:
+            return None  # 本进程读写在对方 insert 之前发生
+        return orig_get(self, cid)
+
+    monkeypatch.setattr(type(repo), "get_l3", racing_get)
+    try:
+        merged = _upsert(repo, "C1", "large_amount", 70, alert_id="ALT-2")
+    finally:
+        monkeypatch.undo()
+
+    assert merged["monitor_tier"] == "high"  # 重读后合并，不降级
+    assert merged["risk_score"] == 95
+    assert merged["last_alert_id"] == "ALT-2"
+    row = _row(engine, "C1")
+    assert row["monitor_tier"] == "high" and row["last_alert_id"] == "ALT-2"