diff --git a/gateway/sense/calibrate.py b/gateway/sense/calibrate.py
new file mode 100644
index 0000000..cb0bf8b
--- /dev/null
+++ b/gateway/sense/calibrate.py
@@ -0,0 +1,132 @@
+"""split-conformal 阈值校准（T-G3）+ last-good 回退 + 工件读写（D-G3）。
+
+目标（§5 policy.alpha）：P(true>T1 | 判 T1) ≤ α（τ3 同理对 T3）。
+实现：calib 集按分数降序扫描，取**满足精度的最低阈值**（覆盖率最大）；
+无任何满足点 -> 阈值取最高分 + ε（全拒，保守）。
+标签数 < min_labels -> 沿用 last-good（D-G3）-> 再无 -> 内置保守值。
+"""
+from __future__ import annotations
+
+import json
+import time
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+from gateway.sense.store import json_dumps
+
+CONSERVATIVE = {"t1": 0.90, "t3": 0.90}
+THRESHOLD_KIND = "thresholds"
+
+
+def _threshold_for(samples: List[Tuple[float, bool]], alpha: float) -> Optional[float]:
+    """样本 [(score, true==tier)] -> 满足精度 >= 1-α 的最低阈值（并列分数整组判定）。
+
+    按唯一分数降序累积计数；每组边界处检查精度；取满足精度的最大覆盖（最低边界）。
+    无任何满足点 -> None（调用方全拒，保守）。
+    """
+    if not samples:
+        return None
+    groups: Dict[float, List[int]] = {}
+    for score, ok in samples:
+        g = groups.setdefault(score, [0, 0])
+        g[0] += 1
+        g[1] += 1 if ok else 0
+    total = 0
+    correct = 0
+    best: Optional[float] = None
+    for score in sorted(groups, reverse=True):
+        cnt, ok_cnt = groups[score]
+        total += cnt
+        correct += ok_cnt
+        if correct / total >= 1.0 - alpha:
+            best = score
+    return best
+
+
+def compute_thresholds(rows: List[Dict[str, Any]], alpha: float = 0.05,
+                       min_labels: int = 500) -> Dict[str, Any]:
+    """rows: [{probs(JSON 串或 dict), true_tier}] -> {t1, t3, coverage, n, ok}。
+
+    ok=False 表示标签不足（沿用 last-good 的信号）。
+    coverage = 通过阈值的 calib 样本占比（实测覆盖率，§6 晋升门比对 α+2%）。
+    """
+    labeled = []
+    for r in rows:
+        probs = r.get("probs")
+        if isinstance(probs, str):
+            try:
+                probs = json.loads(probs)
+            except json.JSONDecodeError:
+                continue
+        if not isinstance(probs, dict):
+            continue
+        true_tier = r.get("true_tier")
+        if true_tier not in ("T1", "T2", "T3"):
+            continue
+        labeled.append((float(probs.get("t1", 0)), float(probs.get("t3", 0)),
+                        true_tier))
+    n = len(labeled)
+    if n < min_labels:
+        return {"t1": None, "t3": None, "coverage": 0.0, "n": n, "ok": False}
+
+    s1 = [(p1, t == "T1") for p1, _p3, t in labeled]
+    s3 = [(p3, t == "T3") for _p1, p3, t in labeled]
+    t1 = _threshold_for(s1, alpha)
+    t3 = _threshold_for(s3, alpha)
+    if t1 is None or t3 is None:
+        return {"t1": t1, "t3": t3, "coverage": 0.0, "n": n, "ok": False}
+    coverage = (sum(1 for p1, _p3, _t in labeled if p1 >= t1)
+                + sum(1 for _p1, p3, _t in labeled if p3 >= t3)) / (2 * n)
+    return {"t1": round(t1, 4), "t3": round(t3, 4),
+            "coverage": round(coverage, 4), "n": n, "ok": True}
+
+
+def save_thresholds(store, models_dir: str | Path, version: str,
+                    th: Dict[str, Any], activate: bool = False) -> str:
+    """阈值工件落盘 thresholds.json + 登记工件表。返回文件路径。"""
+    d = Path(models_dir) / version
+    d.mkdir(parents=True, exist_ok=True)
+    path = d / "thresholds.json"
+    path.write_text(json_dumps(th), encoding="utf-8")
+    store.register_artifact(version, THRESHOLD_KIND, str(path),
+                            {k: th.get(k) for k in ("t1", "t3", "coverage", "n")},
+                            active=activate)
+    return str(path)
+
+
+def load_active(store, models_dir: str | Path) -> Dict[str, Any]:
+    """active 阈值工件 -> 旧工件扫描回退（last-good）-> 内置保守值（D-G3 回退链）。"""
+    art = store.active_artifact(THRESHOLD_KIND)
+    if art:
+        th = _read_threshold(art)
+        if th is not None:
+            th.setdefault("version", art["version"])
+            return th
+        # active 文件缺失/损坏：按时间倒序扫描同 kind 工件（last-good 语义）
+        for cand in store.list_artifacts(THRESHOLD_KIND):
+            if cand["version"] == art["version"]:
+                continue
+            th = _read_threshold(cand)
+            if th is not None:
+                th.setdefault("version", cand["version"])
+                return th
+    last_good = Path(models_dir) / "last-good.json"
+    if last_good.exists():
+        try:
+            th = json.loads(last_good.read_text(encoding="utf-8"))
+            if th.get("t1") is not None and th.get("t3") is not None:
+                th.setdefault("version", "last-good")
+                return th
+        except (json.JSONDecodeError, OSError):
+            pass
+    return {**CONSERVATIVE, "version": "conservative"}
+
+
+def _read_threshold(art: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+    try:
+        th = json.loads(Path(art["path"]).read_text(encoding="utf-8"))
+        if isinstance(th, dict) and th.get("t1") is not None and th.get("t3") is not None:
+            return th
+    except (json.JSONDecodeError, OSError):
+        pass
+    return None
diff --git a/gateway/sense/labeler.py b/gateway/sense/labeler.py
new file mode 100644
index 0000000..7751c57
--- /dev/null
+++ b/gateway/sense/labeler.py
@@ -0,0 +1,60 @@
+"""夜间 true_tier 推导（T-G3，§4 四条规则，确定性）+ 留存清理。
+
+规则（对每条 true_tier 为空的观察，按序判定）：
+0. human_override 非空 -> 以人工为准（规则 4，优先级最高）。
+1. features.plan_multi（consumer=pipeline 发出 brief 的多步信号）-> true=T3。
+2. outcome∈{escalated,failed,user_retry,timeout} -> true = executed 的下一档
+   （T1→T2→T3，T3 保持 T3）。
+3. outcome∈{ok,verified} -> true = executed（T1 成功即 T1；T2/T3 同理）。
+4. 其余（outcome 为空等）-> 跳过，等下一轮。
+"""
+from __future__ import annotations
+
+import json
+import time
+from typing import Any, Dict, Optional
+
+NEXT_TIER = {"T1": "T2", "T2": "T3", "T3": "T3"}
+SUCCESS = ("ok", "verified")
+FAILURE = ("escalated", "failed", "user_retry", "timeout")
+
+
+def _features(row: Dict[str, Any]) -> Dict[str, Any]:
+    try:
+        obj = json.loads(row.get("features") or "{}")
+        return obj if isinstance(obj, dict) else {}
+    except json.JSONDecodeError:
+        return {}
+
+
+def derive_true_tier(row: Dict[str, Any]) -> Optional[str]:
+    """单行推导（纯函数，测试友好）。返回 None = 暂不判定。"""
+    if row.get("human_override"):
+        return str(row["human_override"])
+    outcome = row.get("outcome") or ""
+    executed = row.get("executed_tier") or "T2"
+    feats = _features(row)
+    if feats.get("plan_multi"):
+        return "T3"                                       # 规则 1（pipeline brief 信号）
+    if outcome in FAILURE:                                # 规则 2
+        return NEXT_TIER.get(executed, "T3")
+    if outcome in SUCCESS:                                # 规则 3
+        return executed
+    return None                                           # 规则 4：等待结果
+
+
+def derive_true_tiers(store, now: Optional[float] = None,
+                      retention_days: int = 180) -> int:
+    """批量推导（§7 签名）：返回回填条数；顺带 180d 留存清理。"""
+    now = time.time() if now is None else now
+    rows = store.all_observations()
+    count = 0
+    for row in rows:
+        if row.get("true_tier"):
+            continue
+        true_tier = derive_true_tier(row)
+        if true_tier:
+            store.set_true_tier(row["request_id"], true_tier)
+            count += 1
+    store.purge_older_than(now - retention_days * 86400)
+    return count
diff --git a/gateway/sense/store.py b/gateway/sense/store.py
index 107a163..d75341f 100644
--- a/gateway/sense/store.py
+++ b/gateway/sense/store.py
@@ -116,6 +116,14 @@ class SenseStore:
                 " WHERE true_tier != ''").fetchone()
         return int(row["c"])
 
+    def all_observations(self, limit: int = 200000) -> List[Dict[str, Any]]:
+        """全量遍历（labeler 夜间推导输入；量大时可分页）。"""
+        with self._lock, self._connect() as conn:
+            rows = conn.execute(
+                "SELECT * FROM tier_observations ORDER BY id LIMIT ?",
+                (limit,)).fetchall()
+        return [dict(r) for r in rows]
+
     def purge_older_than(self, ts: float) -> int:
         """留存清理（D-G6：默认 180 天，夜间任务顺带）。"""
         with self._lock, self._connect() as conn:
@@ -154,6 +162,14 @@ class SenseStore:
                 " ORDER BY created_ts DESC LIMIT 1", (kind,)).fetchone()
         return dict(row) if row else None
 
+    def list_artifacts(self, kind: str) -> List[Dict[str, Any]]:
+        """同 kind 工件按时间倒序（load_active 的 last-good 扫描用）。"""
+        with self._lock, self._connect() as conn:
+            rows = conn.execute(
+                "SELECT * FROM sense_artifacts WHERE kind = ?"
+                " ORDER BY created_ts DESC", (kind,)).fetchall()
+        return [dict(r) for r in rows]
+
     # ---------- 自省 ----------
     def table_names(self) -> List[str]:
         with self._lock, self._connect() as conn:
diff --git a/tests/test_sense_labeler.py b/tests/test_sense_labeler.py
new file mode 100644
index 0000000..f9b9bf9
--- /dev/null
+++ b/tests/test_sense_labeler.py
@@ -0,0 +1,112 @@
+"""标签+校准测试（T-G3）：四条推导规则 / T3 保持 / 180d 清理 / conformal 覆盖 / last-good。"""
+import json
+import time
+
+from gateway.sense.calibrate import (
+    compute_thresholds,
+    load_active,
+    save_thresholds,
+)
+from gateway.sense.labeler import derive_true_tier, derive_true_tiers
+from gateway.sense.store import SenseStore
+
+
+def _obs(rid, executed, outcome, probs=None, features=None, override="", true_tier=""):
+    return {"request_id": rid, "consumer": "pipeline", "decided_tier": executed,
+            "executed_tier": executed, "probs": probs or "{}", "outcome": outcome,
+            "features": features or "{}", "human_override": override,
+            "true_tier": true_tier}
+
+
+# ---------------- derive_true_tier 纯函数 ----------------
+
+def test_rule1_success_t1():
+    assert derive_true_tier(_obs("r", "T1", "ok")) == "T1"
+    assert derive_true_tier(_obs("r", "T1", "verified")) == "T1"
+
+
+def test_rule2_failure_next_tier_and_t3_stays():
+    assert derive_true_tier(_obs("r", "T1", "failed")) == "T2"
+    assert derive_true_tier(_obs("r", "T2", "user_retry")) == "T3"
+    assert derive_true_tier(_obs("r", "T3", "timeout")) == "T3"     # T3 保持 T3
+
+
+def test_rule3_plan_multi_signals_t3():
+    row = _obs("r", "T2", "ok", features=json.dumps({"plan_multi": True}))
+    assert derive_true_tier(row) == "T3"
+
+
+def test_rule4_human_override_wins():
+    row = _obs("r", "T1", "failed", override="T3")
+    assert derive_true_tier(row) == "T3"
+
+
+def test_empty_outcome_skipped():
+    assert derive_true_tier(_obs("r", "T1", "")) is None
+
+
+# ---------------- derive_true_tiers 批量 + 清理 ----------------
+
+def test_batch_derive_and_purge(tmp_path):
+    store = SenseStore.init_db(tmp_path / "s.sqlite3")
+    now = time.time()
+    for rid, outcome in [("a", "ok"), ("b", "failed"), ("c", "")]:
+        store.insert_observation(_obs(rid, "T2", outcome))
+    # 一条 200 天前的旧数据（先推导后清理）
+    store.insert_observation({**_obs("old", "T1", "ok"), "ts": now - 200 * 86400})
+    count = derive_true_tiers(store, now=now, retention_days=180)
+    assert count == 3                                   # a→T2, b→T3, old→T1
+    rows = {r["request_id"]: r for r in store.all_observations()}
+    assert rows["a"]["true_tier"] == "T2"
+    assert rows["b"]["true_tier"] == "T3"
+    assert "old" not in rows                            # 180d 清理生效（推导后删除）
+
+
+# ---------------- compute_thresholds ----------------
+
+def test_conformal_threshold_coverage():
+    """合成分布：T1 高分样本真为 T1 -> τ1 满足精度 >= 1-α；τ3 同理（需真 T3 样本）。"""
+    rows = []
+    for i in range(40):
+        rows.append({"probs": json.dumps({"t1": 0.95 - i * 0.001, "t3": 0.01}),
+                     "true_tier": "T1"})
+    for i in range(10):
+        rows.append({"probs": json.dumps({"t1": 0.3, "t3": 0.3}),
+                     "true_tier": "T2"})
+    for i in range(5):
+        rows.append({"probs": json.dumps({"t3": 0.9, "t1": 0.02}),
+                     "true_tier": "T3"})
+    th = compute_thresholds(rows, alpha=0.05, min_labels=10)
+    assert th["ok"] is True and th["n"] == 55
+    kept = [r for r in rows if json.loads(r["probs"])["t1"] >= th["t1"]]
+    precision = sum(1 for r in kept if r["true_tier"] == "T1") / len(kept)
+    assert precision >= 0.95
+
+
+def test_min_labels_not_met_returns_not_ok():
+    rows = [{"probs": json.dumps({"t1": 0.9, "t3": 0.05}), "true_tier": "T1"}]
+    th = compute_thresholds(rows, alpha=0.05, min_labels=500)
+    assert th["ok"] is False and th["t1"] is None
+
+
+def test_threshold_artifact_roundtrip_and_last_good(tmp_path):
+    """工件读写 + active 切换 + last-good 回退链。"""
+    store = SenseStore.init_db(tmp_path / "s.sqlite3")
+    th = {"t1": 0.8, "t3": 0.85, "coverage": 0.9, "n": 500, "ok": True}
+    path = save_thresholds(store, tmp_path, "v1", th, activate=True)
+    assert "thresholds.json" in path
+    loaded = load_active(store, tmp_path)
+    assert loaded["t1"] == 0.8 and loaded["version"] == "v1"
+    # 保存 v2 不激活 -> 仍 v1；激活 v2 -> 切换；v2 文件损坏 -> 扫描回退 v1
+    save_thresholds(store, tmp_path, "v2", {"t1": 0.7, "t3": 0.7,
+                                            "coverage": 0.9, "n": 500})
+    assert load_active(store, tmp_path)["version"] == "v1"     # v2 未激活
+    store.activate_artifact("v2", "thresholds")
+    assert load_active(store, tmp_path)["version"] == "v2"
+    (tmp_path / "v2" / "thresholds.json").unlink()
+    loaded = load_active(store, tmp_path)
+    assert loaded["version"] == "v1" and loaded["t1"] == 0.8   # last-good 扫描回退
+    # 全新 store（无任何工件登记）-> 保守值
+    fresh_store = SenseStore.init_db(tmp_path / "fresh.sqlite3")
+    fresh = load_active(fresh_store, tmp_path / "fresh")
+    assert fresh["version"] == "conservative"
diff --git a/任务拆解与执行计划.md b/任务拆解与执行计划.md
index 1d0ec45..c539ef6 100644
--- a/任务拆解与执行计划.md
+++ b/任务拆解与执行计划.md
@@ -157,7 +157,7 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 | T-G0 | 骨架：gateway/sense/ 包 + sense.sqlite3 DDL + enabled 门控 | ✅ 完成 | c3efa70 |
 | T-G1 | Embedder：/v1/embeddings 客户端 + int8 量化 + 降级阶梯 | ✅ 完成 | T-G1 |
 | T-G2 | 观察埋点：observer + 三消费方埋点（pipeline/proxy/client） | ✅ 完成 | T-G2 |
-| T-G3 | 标签+校准：夜间 true_tier 推导 + split-conformal + 工件表 | ⬜ 待办 | |
+| T-G3 | 标签+校准：夜间 true_tier 推导 + split-conformal + 工件表 | ✅ 完成 | T-G3 |
 | T-G3b | KnnHead（架构变体 B）：kNN 投票 + conformal-kNN + 按桶分区/封顶/压缩；hybrid fusion 预留（§14） | ⬜ 待办 | |
 | T-G4 | 线性头：离线训练脚本 + LinearHead 纯 Python 推理 + 登记 | ⬜ 待办 | |
 | T-G5 | Grader：决策组合（特征门×概率×conformal）+ /v1/route + 三态 mode | ⬜ 待办 | |