diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index bd37810a..df3ff37e 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -123,6 +123,22 @@ jobs:
       - name: Check AGENTS.md references
         run: python scripts/check_agent_docs.py ../AGENTS.md
 
+  # ── 前后端评分口径一致性（#112）──
+  # 前端 recommendation.ts 镜像的两个门槛数字（82/45）与
+  # app/core/scoring_caliber.py CONFIG 对比；#110 的学费不允许再交一次。
+  scoring-mirror:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: backend
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
+      - name: Check scoring caliber mirror
+        run: python scripts/check_scoring_mirror.py
+
   # ── 依赖安全扫描 ──
   security-scan:
     runs-on: ubuntu-latest
diff --git a/AGENTS.md b/AGENTS.md
index eec3c28a..32daee68 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -60,14 +60,19 @@ api/v1/ ──► services/ ──► repositories/ ──► models/ ──►
 `IntegrityError` 等异常类。
 
 **机器强制的边界**：`make layering`（等价于 `cd backend && uv run python
-scripts/check_layering.py`，CI 作业 `backend-layering`）只作用于 `app/api/v1/`，且只认两条：
-禁止 `import sqlalchemy`（`sqlalchemy.ext.asyncio` / `sqlalchemy.exc` 除外），禁止直接写
-`select/insert/update/delete` 构造与 `db.execute/db.add/db.scalars/db.scalar`。上表其余禁止项
-它查不出——`from app.models import <ORMModel>` 与 api 层合法要用的 Enum、依赖注入同模块，
-自动判定会误报，只能人工看。`app/mcp/` 是与 `api/v1` 平级的入口层，不在机器检查范围内，
-按上表 api 行的纪律人工把关（当前只调 services 与 `DailyReportRepository`；OAuth
-provider 的业务编排在 `app/services/mcp_oauth_service.py`，同意页 REST 端点在
-`app/api/v1/mcp_oauth.py`）。
+scripts/check_layering.py`，CI 作业 `backend-layering`）作用于两个目录，共三条规则。
+`app/api/v1/` 认两条：禁止 `import sqlalchemy`（`sqlalchemy.ext.asyncio` /
+`sqlalchemy.exc` 除外），禁止直接写 `select/insert/update/delete` 构造与
+`db.execute/db.add/db.scalars/db.scalar`。`app/repositories/` 认一条（#113）：
+禁止 import `app.services.*`（函数内局部 import 同样抓）；唯一存量豁免是
+`analysis_repo.py` 的两个评分编排 import（评分上移重构单 #122 清空，
+allowlist 只减不增）。跨层共享的口径常量（`scoring_caliber.py`、
+`tag_normalization.py`）已落 `app/core/`，repositories 从 core 取用是合法向下依赖。
+上表其余禁止项机器查不出——`from app.models import <ORMModel>` 与 api 层合法要用的
+Enum、依赖注入同模块，自动判定会误报，只能人工看。`app/mcp/` 是与 `api/v1` 平级的
+入口层，不在机器检查范围内，按上表 api 行的纪律人工把关（当前只调 services 与
+`DailyReportRepository`；OAuth provider 的业务编排在
+`app/services/mcp_oauth_service.py`，同意页 REST 端点在 `app/api/v1/mcp_oauth.py`）。
 
 ### 2.3 迁移与高风险变更
 
@@ -105,10 +110,11 @@ make test-backend                         # 或直接跑全量
 - Backend tests: `cd backend && uv run python -m pytest <relevant-tests> -q` —— `-k` 是 pytest 的**全局**过滤器，`pytest a.py b.py -k x` 会同时筛两个文件，不限于 a.py；要只筛某个文件就分开跑（依赖由 uv 管理：`uv sync` 安装，事实源为 pyproject.toml + uv.lock）
 - Shell scripts: `bash -n <script>`
 - Layering: `make layering`
+- Scoring mirror: `cd backend && uv run python scripts/check_scoring_mirror.py`（前端镜像口径对比 CONFIG，CI 作业 `scoring-mirror`）
 - AGENTS.md 引用: `cd backend && uv run python scripts/check_agent_docs.py`（校验本文引用的路径与 make 目标，CI 作业 `agent-docs`）
 - Frontend type check: `cd frontend && npx tsc --noEmit`
 - Frontend tests: `cd frontend && npm run test:coverage`（含覆盖率门禁，对应 CI `frontend-tests`）
-- Full quality gate: `make lint`（ruff + 分层检查 + 前端类型检查）
+- Full quality gate: `make lint`（ruff + 分层检查 + 口径一致性 + 前端类型检查）
 
 ## 四、怎么交
 
@@ -161,7 +167,7 @@ docs: 跑测前必读补上 OAuth 隔离测试为什么必须 python -m
 
 ### 4.4 Verifier 契约
 
-CI 六项检查（types / tests / lint / layering / agent-docs / security-scan）是机器
+CI 七项检查（types / tests / lint / layering / scoring-mirror / agent-docs / security-scan）是机器
 Verifier，全绿是合并前提；预存红必须显式处置并留记录。PR 描述须附 Worker 证据（本地
 验证命令 + 结果，见 PR 模板 Verification 段）。承载行为变更的 PR 还需在模板 Verifier
 verdict 段落记录独立复核结论（复核者不得是同一实现过程）。
diff --git a/Makefile b/Makefile
index faac568f..e4d233bb 100644
--- a/Makefile
+++ b/Makefile
@@ -9,12 +9,12 @@
 #   make setup          安装后端 + 前端依赖
 #   make dev            docker compose 起全栈
 #   make test           跑后端 + 前端测试
-#   make lint           完整质量门（ruff + 分层 + 前端类型检查）
+#   make lint           完整质量门（ruff + 分层 + 口径一致 + 前端类型检查）
 # =============================================================================
 
 .DEFAULT_GOAL := help
 .PHONY: help setup dev dev-db test test-backend test-frontend \
-        lint lint-backend lint-frontend layering backup
+        lint lint-backend lint-frontend layering scoring-mirror backup
 
 help:  ## 列出所有可用目标
 	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) \
@@ -38,14 +38,17 @@ test-backend:  ## 后端全量 pytest（自动起一次性 PG 容器，跑完即
 test-frontend:  ## 前端 vitest + 覆盖率门禁（对应 CI frontend-tests）
 	cd frontend && npm run test:coverage
 
-lint: lint-backend layering lint-frontend  ## 完整质量门（ruff + 分层 + tsc）
+lint: lint-backend layering scoring-mirror lint-frontend  ## 完整质量门（ruff + 分层 + 口径一致 + tsc）
 
 lint-backend:  ## 后端 ruff 全量检查 + 格式检查（手动质量门）
 	cd backend && uv run ruff check . && uv run ruff format --check .
 
-layering:  ## 分层纪律检查：api/v1 禁止直接 ORM 查询 / 禁用 import
+layering:  ## 分层纪律检查：api/v1 禁 ORM 查询；repositories 禁 services 逆向 import
 	cd backend && uv run python scripts/check_layering.py
 
+scoring-mirror:  ## 前后端评分口径一致性（前端镜像 82/45 对比 CONFIG，#112）
+	cd backend && uv run python scripts/check_scoring_mirror.py
+
 lint-frontend:  ## 前端类型检查（AGENTS.md 首选，比 npm run lint 稳）
 	cd frontend && npx tsc --noEmit
 
diff --git a/backend/app/api/v1/contents.py b/backend/app/api/v1/contents.py
index b33fd13a..e02331d6 100644
--- a/backend/app/api/v1/contents.py
+++ b/backend/app/api/v1/contents.py
@@ -147,9 +147,7 @@ async def list_contents(
     tag: str | None = Query(None, description="按标签筛选（规范化键，大小写不敏感；命中 content 或任一分析的标签）"),
     include_trend_sources: bool = Query(False, description="Include榜单/趋势源 such as DouyinHot"),
     hours: int | None = Query(None, description="Time range in hours, e.g. 24, 48, 168"),
-    sort_by: str = Query(
-        "created_at", pattern=r"^(created_at|published_at|crawled_at|curation_score)$"
-    ),
+    sort_by: str = Query("created_at", pattern=r"^(created_at|published_at|crawled_at|curation_score)$"),
     sort_order: str = Query("desc", pattern=r"^(asc|desc)$"),
     admin_view: bool = Query(False, description="Return management fields; admin only"),
     current_user: User | None = Depends(get_optional_current_user),
@@ -416,9 +414,9 @@ async def content_tag_facets(
     from collections import Counter
     from datetime import timedelta
 
+    from app.core.tag_normalization import normalize_tag_list
     from app.repositories.ignored_repo import IgnoredRepo
     from app.services.json_cache import get_cached_json, set_cached_json
-    from app.services.tag_normalization import normalize_tag_list
 
     scope_q = (q or "").strip()
     cache_key = (
diff --git a/backend/app/services/scoring_caliber.py b/backend/app/core/scoring_caliber.py
similarity index 100%
rename from backend/app/services/scoring_caliber.py
rename to backend/app/core/scoring_caliber.py
diff --git a/backend/app/services/tag_normalization.py b/backend/app/core/tag_normalization.py
similarity index 100%
rename from backend/app/services/tag_normalization.py
rename to backend/app/core/tag_normalization.py
diff --git a/backend/app/repositories/content_repo.py b/backend/app/repositories/content_repo.py
index f93031c9..015f66d8 100644
--- a/backend/app/repositories/content_repo.py
+++ b/backend/app/repositories/content_repo.py
@@ -594,8 +594,8 @@ async def list_paginated_with_analyses(
         )
 
         # ── 服务端推荐等级 / 标签筛选 ────────────────────────────────
+        from app.core.tag_normalization import normalize_tag_key
         from app.models.analysis import AiAnalysis
-        from app.services.tag_normalization import normalize_tag_key
 
         # 两者口径与前端展示保持一致：等级取最新一条分析；标签命中
         # content.tags 或任一分析的 tags（对应前端 getItemTags 的合并）。
@@ -866,8 +866,8 @@ async def list_for_today_picks(
         (``owner_user_id IS NULL``) or content owned by that user. ``None``
         means no visibility filter (batch/internal callers).
         """
+        from app.core.scoring_caliber import CONFIG as SCORING_CONFIG
         from app.models.analysis import AiAnalysis
-        from app.services.scoring_engine import CONFIG as SCORING_CONFIG
 
         cutoff = naive_utc_now() - timedelta(hours=hours)
         risk_threshold = float(SCORING_CONFIG["risk_threshold"])
@@ -908,8 +908,8 @@ async def list_for_report_window(
         口径与 today-picks / DuckDB query_today_picks 对齐：剔除事件附属内容
         与已忽略内容 (``exclude_ids``)。
         """
+        from app.core.scoring_caliber import CONFIG as SCORING_CONFIG
         from app.models.analysis import AiAnalysis
-        from app.services.scoring_engine import CONFIG as SCORING_CONFIG
 
         risk_threshold = float(SCORING_CONFIG["risk_threshold"])
         latest_analysis_id = self._latest_analysis_id_subquery(AiAnalysis)
diff --git a/backend/app/services/analysis.py b/backend/app/services/analysis.py
index 4cc6a1eb..f95212c5 100644
--- a/backend/app/services/analysis.py
+++ b/backend/app/services/analysis.py
@@ -324,8 +324,8 @@ def _build_analysis_record(
     标签与推荐等级在此收口（幂等）：lite 级联与本地回退路径不经过
     _normalize_analysis_result，统一在构造处保证存储口径一致。
     """
+    from app.core.tag_normalization import normalize_tag_list
     from app.services.recommendation_level import classify_recommendation_level
-    from app.services.tag_normalization import normalize_tag_list
 
     scores = result.get("scores", {})
     curation = result.get("curation", {})
diff --git a/backend/app/services/analysis_normalize.py b/backend/app/services/analysis_normalize.py
index da105b3a..3725557c 100644
--- a/backend/app/services/analysis_normalize.py
+++ b/backend/app/services/analysis_normalize.py
@@ -20,8 +20,8 @@
 import json
 from typing import Any
 
+from app.core.tag_normalization import normalize_tag_list
 from app.services.recommendation_level import classify_recommendation_level
-from app.services.tag_normalization import normalize_tag_list
 
 
 def _detect_lang(title: str, content: str) -> str:
diff --git a/backend/app/services/content_pipeline.py b/backend/app/services/content_pipeline.py
index 12240551..3188a59d 100644
--- a/backend/app/services/content_pipeline.py
+++ b/backend/app/services/content_pipeline.py
@@ -19,6 +19,7 @@
 from sqlalchemy.ext.asyncio import AsyncSession
 
 from app.core.config import settings
+from app.core.tag_normalization import normalize_tag_list
 from app.models.content import ContentItem, ContentStatus
 from app.models.source import Source, SourceStatus, SourceType
 
@@ -31,7 +32,6 @@
 from app.services.llm_pre_filter import apply_pre_filter
 from app.services.scraper_http import build_scraper_client_kwargs
 from app.services.scrapers import get_scraper_cls
-from app.services.tag_normalization import normalize_tag_list
 from app.utils.url_safety import UnsafeUrlError, ensure_public_hostname
 
 logger = logging.getLogger(__name__)
diff --git a/backend/app/services/llm/prompts/analysis.py b/backend/app/services/llm/prompts/analysis.py
index 6997a101..4e0dec00 100644
--- a/backend/app/services/llm/prompts/analysis.py
+++ b/backend/app/services/llm/prompts/analysis.py
@@ -4,11 +4,11 @@
 Used by:
     - app.services.analysis.analyze_content
 
-评分口径(加权/门槛/风险扣分)一律从 app.services.scoring_caliber 渲染,
+评分口径(加权/门槛/风险扣分)一律从 app.core.scoring_caliber 渲染,
 本文件不得手写数字——tests/test_scoring_caliber.py 锁死漂移(#110)。
 """
 
-from app.services.scoring_caliber import (
+from app.core.scoring_caliber import (
     curation_threshold_text,
     curation_weights_text,
     dimension_anchor_section,
diff --git a/backend/app/services/scoring_engine.py b/backend/app/services/scoring_engine.py
index 3f810344..4cf7b78d 100644
--- a/backend/app/services/scoring_engine.py
+++ b/backend/app/services/scoring_engine.py
@@ -19,16 +19,16 @@
 import math
 from datetime import UTC, datetime
 
+from app.core.scoring_caliber import CONFIG
 from app.schemas.scoring import (
     ScoreBreakdownResponse,
     ScoreResultItem,
     ScoringRequestItem,
     ScoringResponse,
 )
-from app.services.scoring_caliber import CONFIG
 
 # ── Tunable configuration ────────────────────────────────────────────
-# 唯一声明处已迁至 app/services/scoring_caliber.py(#110):prompt 教学与
+# 唯一声明处为 app/core/scoring_caliber.py(#110,#113):prompt 教学与
 # picks 查询共享同一份口径,tests/test_scoring_caliber.py 锁死漂移。
 
 
diff --git a/backend/scripts/check_layering.py b/backend/scripts/check_layering.py
index 8bb9ff1c..bcfd3be8 100644
--- a/backend/scripts/check_layering.py
+++ b/backend/scripts/check_layering.py
@@ -1,7 +1,9 @@
 #!/usr/bin/env python3
 """分层纪律的机器强制检查（对应 AGENTS.md「Layering Discipline」）。
 
-只机器化 AGENTS.md 里*符号级、可无歧义判定*的两条禁止项，作用于 `app/api/v1/`：
+机器化 AGENTS.md 里*符号级、可无歧义判定*的禁止项，作用于两个目录：
+
+**app/api/v1/**（两条）：
 
 1. 禁止 import sqlalchemy —— 例外：`sqlalchemy.ext.asyncio`（AsyncSession 类型注解）
    与 `sqlalchemy.exc`（IntegrityError / OperationalError 等异常类，供 try/except）。
@@ -9,13 +11,20 @@
    以及 `db.execute(...) / db.add(...) / db.scalars(...) / db.scalar(...)`。
    这些必须下沉到 `repositories/`。
 
+**app/repositories/**（一条，#113）：
+
+3. 禁止 import `app.services.*` —— 依赖方向 repositories→services 逆向
+   （函数内局部 import 同样抓：#113 的违例全部藏在函数体内）。
+   已知存量豁免走 `REPO_SERVICES_IMPORT_ALLOWLIST`（按文件+模块粒度，
+   引用重构 issue，棘轮式收缩，只减不增）。
+
 **不检查**的项（因为是符号级且和允许项同模块，自动判定会大量误报）：
 `from app.models import <ORMModel>` 与 `<ORMModel>` 混在 Enum / User 依赖注入里的情况，
 仍靠 AGENTS.md 2.2 分层表「禁止」列的人工评审兜底。
 
 用法：
-    python scripts/check_layering.py            # 检查默认目录
-    python scripts/check_layering.py app/api/v1 # 指定目录
+    python scripts/check_layering.py            # 检查 api/v1 + repositories
+    python scripts/check_layering.py app/api/v1 # 只检查指定目录（api 规则）
 
 命中违规时打印文件:行 + 说明并以退出码 1 结束，供 CI 阻断。
 """
@@ -28,11 +37,20 @@
 
 # 检查目标目录（相对 backend/）。
 API_DIR = Path("app/api/v1")
+REPO_DIR = Path("app/repositories")
 
 # 刻意豁免的文件名（分层的合法低层边界）。
 # _db_write.py 不再需要豁免——SQLite PRAGMA 调用已随 SQLite 分支移除。
 ALLOWLIST: set[str] = set()
 
+# repositories→services 禁令的存量豁免（#113 收口时遗留，#122 重构后清空）。
+# 键 = 文件名，值 = 允许 import 的 app.services 模块前缀集合；
+# 新增条目必须附带 issue 引号，且只允许删除、不允许新增（棘轮）。
+REPO_SERVICES_IMPORT_ALLOWLIST: dict[str, set[str]] = {
+    # analysis_repo 评分编排上移 services 的重构单：#122
+    "analysis_repo.py": {"app.services.feedback_signal", "app.services.scoring_engine"},
+}
+
 # 允许的 sqlalchemy 子模块前缀（值对象 / 类型注解 / 异常类）。
 ALLOWED_SQLALCHEMY_PREFIXES = ("sqlalchemy.ext.asyncio", "sqlalchemy.exc")
 
@@ -95,16 +113,54 @@ def check_file(path: Path) -> list[str]:
     return violations
 
 
+def check_repo_services_imports(path: Path) -> list[str]:
+    """repositories 层文件：禁止 import app.services.*（allowlist 内的前缀除外）。
+
+    ast.walk 覆盖模块级与函数体内的 import——#113 的违例全部藏在函数体里，
+    只查模块级会漏。
+    """
+    allowed_prefixes = REPO_SERVICES_IMPORT_ALLOWLIST.get(path.name, set())
+    violations: list[str] = []
+    tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
+
+    def _violation(lineno: int, desc: str) -> None:
+        if any(desc == prefix or desc.startswith(prefix + ".") for prefix in allowed_prefixes):
+            return
+        violations.append(
+            f"{path}:{lineno}: 禁止 `{desc}`——repositories 层不得 import services"
+            f"（业务编排属于 service，口径常量在 app/core/；存量豁免只减不增）"
+        )
+
+    for node in ast.walk(tree):
+        if isinstance(node, ast.ImportFrom):
+            module = node.module or ""
+            if module == "app.services" or module.startswith("app.services."):
+                _violation(node.lineno, module)
+        elif isinstance(node, ast.Import):
+            for alias in node.names:
+                if alias.name == "app.services" or alias.name.startswith("app.services."):
+                    _violation(node.lineno, alias.name)
+
+    return violations
+
+
 def main(argv: list[str]) -> int:
-    target = Path(argv[1]) if len(argv) > 1 else API_DIR
-    if not target.exists():
-        print(f"ERROR: 目录不存在: {target}（请在 backend/ 下运行）", file=sys.stderr)
-        return 2
+    # legacy 单目录模式只跑 api 规则（保持既有用法兼容）；默认双目录全查。
+    targets = [(Path(argv[1]), "api")] if len(argv) > 1 else [(API_DIR, "api"), (REPO_DIR, "repo")]
 
     all_violations: list[str] = []
-    files = sorted(p for p in target.glob("*.py") if p.name not in ALLOWLIST)
-    for path in files:
-        all_violations.extend(check_file(path))
+    file_count = 0
+    for target, kind in targets:
+        if not target.exists():
+            print(f"ERROR: 目录不存在: {target}（请在 backend/ 下运行）", file=sys.stderr)
+            return 2
+        files = sorted(p for p in target.glob("*.py") if p.name not in ALLOWLIST)
+        file_count += len(files)
+        for path in files:
+            if kind == "api":
+                all_violations.extend(check_file(path))
+            else:
+                all_violations.extend(check_repo_services_imports(path))
 
     if all_violations:
         print("分层违规（详见 AGENTS.md「Layering Discipline」）：", file=sys.stderr)
@@ -116,7 +172,8 @@ def main(argv: list[str]) -> int:
         )
         return 1
 
-    print(f"分层检查通过：{len(files)} 个 api/v1 文件无直接 ORM 查询 / 禁用 import。")
+    scopes = " + ".join(str(t) for t, _ in targets)
+    print(f"分层检查通过：{scopes} 共 {file_count} 个文件，无禁用 import / 逆向依赖。")
     return 0
 
 
diff --git a/backend/scripts/check_scoring_mirror.py b/backend/scripts/check_scoring_mirror.py
new file mode 100644
index 00000000..2d64a777
--- /dev/null
+++ b/backend/scripts/check_scoring_mirror.py
@@ -0,0 +1,107 @@
+#!/usr/bin/env python3
+"""前后端评分口径一致性检查（#112）：前端镜像数字与 CONFIG 的漂移门禁。
+
+`frontend/src/lib/recommendation.ts` 硬编码了两个镜像后端 CONFIG 的展示门槛：
+
+- `RISK_HARD_EXCLUDE` ↔ `CONFIG["risk_threshold"]`（硬排除线）
+- `QUALITY_GATE_MIN` ↔ `CONFIG["quality_gate_min"]`（过薄下限）
+
+#110 的教训是「同口径数字的第二份手写副本会静默漂移」；#111 给后端 Python 侧
+建了正则防线（tests/test_scoring_caliber.py），本脚本是同一防线的跨树延伸：
+CONFIG 调整而前端忘了跟，CI 即红。
+
+`RISK_HIGH` 是前端自创的展示档位（后端无对应数字，软起点 45 / 硬排除 82 之间
+的展示分档），**不参与对比**——但必须带着「前端自有」标记注释存在：标记被删
+说明有人动了这个数字却没意识到它是豁免项，同样红。
+
+方案取舍见 #112：接口下发被否决（为一年一变的常量引入 fetch/加载态/失败面
+不成比例），注释豁免单独不足（漂移只有门禁拦得住）。
+
+用法：
+    python scripts/check_scoring_mirror.py   # 在 backend/ 下运行
+
+不一致时打印 前端值 vs CONFIG 值 并以退出码 1 结束，供 CI 阻断。
+"""
+
+from __future__ import annotations
+
+import re
+import sys
+from pathlib import Path
+
+BACKEND_ROOT = Path(__file__).resolve().parents[1]
+REPO_ROOT = BACKEND_ROOT.parent
+FRONTEND_TS = REPO_ROOT / "frontend" / "src" / "lib" / "recommendation.ts"
+
+# 前端常量名 → CONFIG 键（参与一致性对比的镜像对）。
+MIRROR_PAIRS: dict[str, str] = {
+    "RISK_HARD_EXCLUDE": "risk_threshold",
+    "QUALITY_GATE_MIN": "quality_gate_min",
+}
+
+# 豁免档位：不对比，但行内必须带这个标记（防止无声地改掉豁免语义）。
+EXEMPT_CONSTANTS: dict[str, str] = {"RISK_HIGH": "前端自有"}
+
+_CONST_RE = r"^\s*const\s+{name}\s*=\s*(\d+)\s*;"
+
+
+def extract_const(source: str, name: str) -> tuple[int, int] | None:
+    """返回 (值, 行号)；找不到返回 None。"""
+    for lineno, line in enumerate(source.splitlines(), start=1):
+        m = re.match(_CONST_RE.format(name=name), line)
+        if m:
+            return int(m.group(1)), lineno
+    return None
+
+
+def run_check(ts_path: Path, config: dict) -> list[str]:
+    """返回违规描述列表（空列表表示一致）。config 传入以便测试注入。"""
+    source = ts_path.read_text(encoding="utf-8")
+    violations: list[str] = []
+
+    for const_name, config_key in MIRROR_PAIRS.items():
+        found = extract_const(source, const_name)
+        if found is None:
+            violations.append(f"{ts_path}: 找不到 `{const_name}` 常量声明（改名请同步本脚本）")
+            continue
+        value, lineno = found
+        expected = config[config_key]
+        if value != expected:
+            violations.append(
+                f"{ts_path}:{lineno}: `{const_name} = {value}` 与 CONFIG['{config_key}'] = {expected} "
+                f"漂移——调口径时前端镜像必须同步（或反向：后端改了口径，前端解释文案已失真）"
+            )
+
+    for const_name, marker in EXEMPT_CONSTANTS.items():
+        found = extract_const(source, const_name)
+        if found is None:
+            continue  # 常量被删除（功能移除）不算漂移
+        _, lineno = found
+        line = source.splitlines()[lineno - 1]
+        if marker not in line:
+            violations.append(
+                f"{ts_path}:{lineno}: `{const_name}` 行内缺少「{marker}」标记——"
+                f"它是豁免对比的前端展示档位，改它或去标记都必须是有意识的决定"
+            )
+
+    return violations
+
+
+def main() -> int:
+    sys.path.insert(0, str(BACKEND_ROOT))
+    from app.core.scoring_caliber import CONFIG  # noqa: E402 — sys.path 就绪后才能 import
+
+    violations = run_check(FRONTEND_TS, CONFIG)
+    if violations:
+        print("评分口径前后端不一致（详见 scripts/check_scoring_mirror.py 与 #112）：", file=sys.stderr)
+        for line in violations:
+            print(f"  {line}", file=sys.stderr)
+        return 1
+
+    mirrors = "、".join(f"{c}={CONFIG[k]}" for c, k in MIRROR_PAIRS.items())
+    print(f"口径一致性通过：前端镜像 {mirrors} 与 CONFIG 一致；豁免档位标记在位。")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/backend/scripts/scoring_anchor_ab.py b/backend/scripts/scoring_anchor_ab.py
index 5eb2c551..89f28040 100644
--- a/backend/scripts/scoring_anchor_ab.py
+++ b/backend/scripts/scoring_anchor_ab.py
@@ -33,11 +33,11 @@
 from sqlalchemy import or_, select  # noqa: E402
 
 from app.core.database import async_session  # noqa: E402
+from app.core.scoring_caliber import CONFIG, dimension_anchor_section  # noqa: E402
 from app.models.content import ContentItem  # noqa: E402
 from app.services.analysis import _select_analysis_prompts  # noqa: E402
 from app.services.llm import call_llm_json  # noqa: E402
 from app.services.llm.prompts.analysis import ANALYSIS_PROMPT_EN  # noqa: E402
-from app.services.scoring_caliber import CONFIG, dimension_anchor_section  # noqa: E402
 from app.utils.prompt_safety import sanitize_prompt_input  # noqa: E402
 
 logger = logging.getLogger("scoring_anchor_ab")
diff --git a/backend/tests/test_check_layering.py b/backend/tests/test_check_layering.py
new file mode 100644
index 00000000..120beadb
--- /dev/null
+++ b/backend/tests/test_check_layering.py
@@ -0,0 +1,123 @@
+"""check_layering.py 的回归测试：repositories→services 禁令自身的正确性机器化。
+
+#113 落地的门禁。它一旦误报，agent 会学会忽略检查结果；它一旦漏报
+（比如只查模块级 import、漏掉函数体内的局部 import——#113 的违例全部藏在
+函数体里），逆向依赖就回到静默存续。两者都比没有这道门禁更糟，用测试钉住。
+
+用例分三组：
+- **应杀**：合成违例文件（模块级 / 函数内局部 / `import app.services.x` 三形态）
+  必须被抓到。
+- **豁免**：allowlist 内的前缀放行、前缀外的仍杀（棘轮语义——豁免是
+  模块粒度不是文件粒度，analysis_repo 若新增第三个 services 依赖照样红）。
+- **契约**：对真实仓库默认调用退出码为 0（CI 依赖命令行 + 退出码契约）。
+
+本文件是纯 stdlib + subprocess，不碰数据库（外层 conftest 的 DB fixture 与它无关）。
+背景：#113（范围 2→5 处的取证）、#122（allowlist 清空的重构单）。
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+BACKEND_ROOT = Path(__file__).resolve().parents[1]
+SCRIPT = BACKEND_ROOT / "scripts" / "check_layering.py"
+
+
+def run_script(*args: str) -> tuple[int, str]:
+    """跑命令行契约（退出码 + 输出），cwd 固定 backend/ 让相对目录生效。"""
+    proc = subprocess.run(
+        [sys.executable, str(SCRIPT), *args],
+        capture_output=True,
+        text=True,
+        cwd=str(BACKEND_ROOT),
+    )
+    return proc.returncode, proc.stdout + proc.stderr
+
+
+def load_cl():
+    """按路径加载 scripts/check_layering.py（scripts 不是包）。"""
+    import importlib.util
+
+    spec = importlib.util.spec_from_file_location("check_layering_under_test", SCRIPT)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+@pytest.fixture(scope="module")
+def cl():
+    return load_cl()
+
+
+def repo_violations(tmp_path: Path, source: str, cl) -> list[str]:
+    """写入临时文件跑 check_repo_services_imports。"""
+    target = tmp_path / "some_repo.py"
+    target.write_text(source, encoding="utf-8")
+    return cl.check_repo_services_imports(target)
+
+
+# ── 应杀 ──
+
+
+def test_module_level_from_services_is_caught(cl, tmp_path: Path) -> None:
+    src = "from app.services.scoring_engine import score_items\n"
+    assert repo_violations(tmp_path, src, cl)
+
+
+def test_function_local_from_services_is_caught(cl, tmp_path: Path) -> None:
+    """#113 的违例形态：藏在函数体内的局部 import。"""
+    src = (
+        "def f():\n"
+        "    from app.services.scoring_engine import CONFIG as SCORING_CONFIG\n"
+        "    return SCORING_CONFIG\n"
+    )
+    assert repo_violations(tmp_path, src, cl)
+
+
+def test_plain_import_services_is_caught(cl, tmp_path: Path) -> None:
+    src = "import app.services.feedback_signal\n"
+    assert repo_violations(tmp_path, src, cl)
+
+
+def test_non_services_import_is_fine(cl, tmp_path: Path) -> None:
+    """core/models/repositories 内部依赖不受影响。"""
+    src = (
+        "from app.core.scoring_caliber import CONFIG\n"
+        "from app.core.tag_normalization import normalize_tag_key\n"
+        "from app.repositories.base import BaseRepository\n"
+    )
+    assert repo_violations(tmp_path, src, cl) == []
+
+
+# ── 豁免（棘轮语义）──
+
+
+def test_allowlisted_module_passes_and_prefix_escape_is_caught(cl, tmp_path: Path) -> None:
+    """豁免按 文件名+模块前缀 生效：allowlist 外的 services 依赖仍然杀。"""
+    allowed = "from app.services.scoring_engine import score_items\n"
+    escaped = "from app.services.duckdb_service import run_query\n"
+    # 同名文件才走 analysis_repo 的豁免
+    target = tmp_path / "analysis_repo.py"
+    target.write_text(allowed + escaped, encoding="utf-8")
+    violations = cl.check_repo_services_imports(target)
+    assert len(violations) == 1
+    assert "duckdb_service" in violations[0]
+
+
+# ── 契约 ──
+
+
+def test_real_repo_default_invocation_passes() -> None:
+    """真实仓库默认调用（api/v1 + repositories）必须退出码 0。"""
+    code, out = run_script()
+    assert code == 0, f"真实仓库分层检查未通过：\n{out}"
+
+
+def test_allowlist_only_shrinks() -> None:
+    """allowlist 只减不增：当前仅剩 analysis_repo 一条（#122 清空）。"""
+    cl = load_cl()
+    assert set(cl.REPO_SERVICES_IMPORT_ALLOWLIST) == {"analysis_repo.py"}
diff --git a/backend/tests/test_check_scoring_mirror.py b/backend/tests/test_check_scoring_mirror.py
new file mode 100644
index 00000000..76519c97
--- /dev/null
+++ b/backend/tests/test_check_scoring_mirror.py
@@ -0,0 +1,111 @@
+"""check_scoring_mirror.py 的回归测试：口径门禁自身的正确性机器化。
+
+#112 落地的门禁（CI 作业 scoring-mirror）。它一旦误报，agent 会学会忽略
+检查结果；它一旦漏报，前端镜像数字回到静默漂移——#110 的学费就白交了。
+
+用例分三组：
+- **应杀**：镜像数字漂移（82→84）、豁免标记被删，都必须红。
+- **放行**：豁免档位带标记改值不红（它是前端自有，改值是合法决定）；
+  常量整体移除不算漂移。
+- **契约**：对真实仓库命令行调用退出码 0（CI 依赖命令行 + 退出码契约）。
+
+本文件纯 stdlib + subprocess，不碰数据库（外层 conftest 的 DB fixture 与它无关）。
+"""
+
+from __future__ import annotations
+
+import subprocess
+import sys
+from pathlib import Path
+
+import pytest
+
+BACKEND_ROOT = Path(__file__).resolve().parents[1]
+SCRIPT = BACKEND_ROOT / "scripts" / "check_scoring_mirror.py"
+
+REAL_TS = BACKEND_ROOT.parent / "frontend" / "src" / "lib" / "recommendation.ts"
+
+BASE_CONFIG = {"risk_threshold": 82, "quality_gate_min": 45}
+
+BASE_SOURCE = """const RISK_HARD_EXCLUDE = 82; // = CONFIG.risk_threshold（镜像，CI 对比锁死）
+const RISK_HIGH = 75; // 前端自有展示档位：后端无对应数字，不参与口径对比
+const QUALITY_GATE_MIN = 45; // = CONFIG.quality_gate_min（镜像，CI 对比锁死）
+"""
+
+
+def load_mod():
+    import importlib.util
+
+    spec = importlib.util.spec_from_file_location("check_scoring_mirror_under_test", SCRIPT)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+@pytest.fixture(scope="module")
+def mod():
+    return load_mod()
+
+
+def write_ts(tmp_path: Path, source: str) -> Path:
+    p = tmp_path / "recommendation.ts"
+    p.write_text(source, encoding="utf-8")
+    return p
+
+
+# ── 应杀 ──
+
+
+def test_mirror_drift_is_caught(mod, tmp_path: Path) -> None:
+    """后端调到 84 而前端还是 82 → 红（或反向手改前端）。"""
+    src = BASE_SOURCE.replace("RISK_HARD_EXCLUDE = 82", "RISK_HARD_EXCLUDE = 84")
+    violations = mod.run_check(write_ts(tmp_path, src), BASE_CONFIG)
+    assert len(violations) == 1
+    assert "RISK_HARD_EXCLUDE" in violations[0]
+
+
+def test_exempt_marker_removed_is_caught(mod, tmp_path: Path) -> None:
+    """RISK_HIGH 行的「前端自有」标记被删 → 红（防止无意识动豁免语义）。"""
+    src = BASE_SOURCE.replace("// 前端自有展示档位：后端无对应数字，不参与口径对比", "")
+    violations = mod.run_check(write_ts(tmp_path, src), BASE_CONFIG)
+    assert any("RISK_HIGH" in v for v in violations)
+
+
+def test_missing_const_is_caught(mod, tmp_path: Path) -> None:
+    src = "const QUALITY_GATE_MIN = 45; // = CONFIG.quality_gate_min（镜像，CI 对比锁死）\n"
+    violations = mod.run_check(write_ts(tmp_path, src), BASE_CONFIG)
+    assert any("RISK_HARD_EXCLUDE" in v for v in violations)
+
+
+# ── 放行 ──
+
+
+def test_exempt_value_change_with_marker_passes(mod, tmp_path: Path) -> None:
+    """豁免档位带标记改值：前端自有，合法决定，不红。"""
+    src = BASE_SOURCE.replace("RISK_HIGH = 75", "RISK_HIGH = 78")
+    assert mod.run_check(write_ts(tmp_path, src), BASE_CONFIG) == []
+
+
+def test_base_source_passes(mod, tmp_path: Path) -> None:
+    assert mod.run_check(write_ts(tmp_path, BASE_SOURCE), BASE_CONFIG) == []
+
+
+# ── 契约 ──
+
+
+def test_real_repo_invocation_passes() -> None:
+    """真实仓库命令行调用必须退出码 0（真实 CONFIG × 真实 recommendation.ts）。"""
+    proc = subprocess.run(
+        [sys.executable, str(SCRIPT)],
+        capture_output=True,
+        text=True,
+        cwd=str(BACKEND_ROOT),
+    )
+    assert proc.returncode == 0, f"真实仓库口径一致性未通过：\n{proc.stdout}{proc.stderr}"
+
+
+def test_real_ts_exists_and_matches_shape(mod) -> None:
+    """真实 recommendation.ts 三个常量都在（防重构后测试对象悄悄消失）。"""
+    source = REAL_TS.read_text(encoding="utf-8")
+    for name in ("RISK_HARD_EXCLUDE", "RISK_HIGH", "QUALITY_GATE_MIN"):
+        assert mod.extract_const(source, name) is not None, f"{name} 不在真实前端文件里"
diff --git a/backend/tests/test_scoring_anchor_ab.py b/backend/tests/test_scoring_anchor_ab.py
index fbdf884f..91cf2992 100644
--- a/backend/tests/test_scoring_anchor_ab.py
+++ b/backend/tests/test_scoring_anchor_ab.py
@@ -10,8 +10,8 @@
 import importlib.util
 import os
 
+from app.core.scoring_caliber import dimension_anchor_section
 from app.services.llm.prompts.analysis import ANALYSIS_PROMPT, ANALYSIS_PROMPT_EN
-from app.services.scoring_caliber import dimension_anchor_section
 
 _SCRIPT_PATH = os.path.join(
     os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
diff --git a/backend/tests/test_scoring_caliber.py b/backend/tests/test_scoring_caliber.py
index 507ce310..f04051b2 100644
--- a/backend/tests/test_scoring_caliber.py
+++ b/backend/tests/test_scoring_caliber.py
@@ -14,9 +14,7 @@
 import inspect
 import re
 
-from app.services import scoring_engine
-from app.services.llm.prompts import analysis as analysis_prompts
-from app.services.scoring_caliber import (
+from app.core.scoring_caliber import (
     ANCHOR_BANDS,
     CONFIG,
     DIMENSION_ANCHORS_EN,
@@ -26,6 +24,8 @@
     dimension_anchor_lines,
     dimension_anchor_section,
 )
+from app.services import scoring_engine
+from app.services.llm.prompts import analysis as analysis_prompts
 
 SIX_DIM_KEYS = (
     "w_info_density",
diff --git a/backend/tests/test_tag_normalization.py b/backend/tests/test_tag_normalization.py
index d128e9ae..f06d80bb 100644
--- a/backend/tests/test_tag_normalization.py
+++ b/backend/tests/test_tag_normalization.py
@@ -1,8 +1,8 @@
-"""标签规范化（services/tag_normalization.py）单元测试。"""
+"""标签规范化（core/tag_normalization.py）单元测试。"""
 
 from __future__ import annotations
 
-from app.services.tag_normalization import normalize_tag_key, normalize_tag_list, split_compound_tags
+from app.core.tag_normalization import normalize_tag_key, normalize_tag_list, split_compound_tags
 
 
 def test_split_compound_tags_variants():
diff --git a/frontend/src/lib/recommendation.ts b/frontend/src/lib/recommendation.ts
index 45fea7d0..85175953 100644
--- a/frontend/src/lib/recommendation.ts
+++ b/frontend/src/lib/recommendation.ts
@@ -39,14 +39,15 @@ function isDefaultScoreProfile(analysis: ContentAnalysis): boolean {
   return defaultLikeCount >= 5 && !hasCuration && !hasTextSignal;
 }
 
-// 与后端评分引擎对齐的门槛（backend/app/services/scoring_engine.py CONFIG）：
-// risk_threshold=82 硬排除；quality_gate_min=45 内容过薄下限。
+// 与后端评分口径对齐的门槛（唯一声明处 backend/app/core/scoring_caliber.py CONFIG）：
 // 规则顺序：先过统一的风险/质量/创作价值门槛，再判断推荐类型，
 // 避免「质量 10 + 创作价值 90」被判为强烈建议写、「风险 95 + 热度 90」
 // 被判为适合蹭热点这类后端根本不会收录的推荐口径。
-const RISK_HARD_EXCLUDE = 82;
-const RISK_HIGH = 75;
-const QUALITY_GATE_MIN = 45;
+// 两个镜像数字由 CI 门禁锁死漂移（backend/scripts/check_scoring_mirror.py，#112）：
+// 后端调口径而这里没跟，CI 即红。
+const RISK_HARD_EXCLUDE = 82; // = CONFIG.risk_threshold（镜像，CI 对比锁死）
+const RISK_HIGH = 75; // 前端自有展示档位：后端无对应数字（软起点 45 / 硬排除 82 之间），不参与口径对比
+const QUALITY_GATE_MIN = 45; // = CONFIG.quality_gate_min（镜像，CI 对比锁死）
 
 export function explainRecommendation(analysis: ContentAnalysis | null | undefined): RecommendationDecision {
   if (!analysis) {