diff --git a/src/nsc/runtime/models.py b/src/nsc/runtime/models.py
index 71438f5..5be01c4 100644
--- a/src/nsc/runtime/models.py
+++ b/src/nsc/runtime/models.py
@@ -2,18 +2,37 @@
 
 tier → 模型 的映射在 config/models.yaml（配置是生成物，路由策略是资产）。
 成本统计走 litellm.completion_cost；Langfuse trace 在配置缺失时静默降级。
+SW-01：每次调用的 prompt/response 落 SQLite transcripts（Lab ADR-0001 §接口），
+best-effort——写库失败静默降级，绝不影响路由本身。
 """
 
 from __future__ import annotations
 
+import json
 import os
+import sqlite3
 import time
 from dataclasses import dataclass
+from datetime import UTC, datetime
 from pathlib import Path
 from typing import Any
 
 import yaml
 
+_TRANSCRIPT_SCHEMA = """
+CREATE TABLE IF NOT EXISTS transcripts (
+  ts            TEXT NOT NULL,
+  caller        TEXT NOT NULL,
+  model         TEXT NOT NULL,
+  prompt        TEXT NOT NULL,
+  response      TEXT NOT NULL,
+  tokens_in     INTEGER NOT NULL DEFAULT 0,
+  tokens_out    INTEGER NOT NULL DEFAULT 0,
+  cost_usd      REAL NOT NULL DEFAULT 0,
+  experiment_id TEXT NOT NULL DEFAULT ''
+)
+"""
+
 
 @dataclass(slots=True)
 class LLMResult:
@@ -36,7 +55,13 @@ def _has_json(text: str) -> bool:
 class ModelRouter:
     """按 tier 路由到具体模型，带重试与成本统计。"""
 
-    def __init__(self, config_path: str | Path = "config/models.yaml") -> None:
+    def __init__(
+        self,
+        config_path: str | Path = "config/models.yaml",
+        *,
+        transcript_db: str | Path | None = None,
+        experiment_id: str = "",
+    ) -> None:
         cfg = yaml.safe_load(Path(config_path).read_text("utf-8"))
         self.tiers: dict[str, dict[str, Any]] = cfg.get("tiers", {})
         self.budgets: dict[str, float] = cfg.get("budgets", {})
@@ -44,6 +69,64 @@ def __init__(self, config_path: str | Path = "config/models.yaml") -> None:
         self.attempts: int = int(retry.get("attempts", 3))
         # litellm 不认识的新模型（如 LongCat-2.0）用配置价兜底成本统计。
         self.cost_per_mtok: dict[str, float] = cfg.get("cost_usd_per_mtok", {}) or {}
+        # SW-01 transcript 台账：库路径/实验号可用环境变量接线（Lab subprocess 场景）。
+        self.transcript_db = Path(
+            transcript_db or os.environ.get("NSC_TRANSCRIPT_DB") or "out/transcripts.db"
+        )
+        self.experiment_id = experiment_id or os.environ.get("NSC_EXPERIMENT_ID", "")
+        self._tconn: sqlite3.Connection | None = None
+
+    def _transcript_conn(self) -> sqlite3.Connection | None:
+        """懒建连接；任何 IO 异常 → 返回 None（本功能 best-effort）。"""
+        if self._tconn is None:
+            try:
+                self.transcript_db.parent.mkdir(parents=True, exist_ok=True)
+                # timeout=0:transcripts 是 best-effort 记账,库被并发写锁住时立刻
+                # 失败走静默路径,绝不为它阻塞路由(SW-01 review:busy timeout 会加路由延迟)
+                self._tconn = sqlite3.connect(str(self.transcript_db), timeout=0.0)
+                self._tconn.execute(_TRANSCRIPT_SCHEMA)
+                self._tconn.commit()
+            except Exception:
+                self._tconn = None
+        return self._tconn
+
+    def _record_transcript(
+        self,
+        tier: str,
+        model_id: str,
+        messages: list[dict[str, str]],
+        text: str,
+        tokens_in: int,
+        tokens_out: int,
+        cost: float,
+    ) -> None:
+        conn = self._transcript_conn()
+        if conn is None:
+            return
+        try:
+            conn.execute(
+                "INSERT INTO transcripts (ts, caller, model, prompt, response,"
+                " tokens_in, tokens_out, cost_usd, experiment_id)"
+                " VALUES (?,?,?,?,?,?,?,?,?)",
+                (
+                    datetime.now(UTC).isoformat(),
+                    tier,
+                    model_id,
+                    json.dumps(messages, ensure_ascii=False),
+                    text,
+                    tokens_in,
+                    tokens_out,
+                    cost,
+                    self.experiment_id,
+                ),
+            )
+            conn.commit()
+        except Exception:
+            # best-effort:失败即回滚并弃置连接,防止半开事务长期持锁拖垮后续写入
+            try:
+                conn.rollback()
+            finally:
+                self._tconn = None
 
     def resolve(self, tier: str) -> dict[str, Any]:
         if tier not in self.tiers:
@@ -119,6 +202,7 @@ def complete(
                 tokens_in * float(self.cost_per_mtok.get("input", 0.0))
                 + tokens_out * float(self.cost_per_mtok.get("output", 0.0))
             ) / 1_000_000
+        self._record_transcript(tier, cfg["model"], messages, text, tokens_in, tokens_out, cost)
         return LLMResult(
             text=text,
             model_id=cfg["model"],
diff --git a/tests/test_transcripts.py b/tests/test_transcripts.py
new file mode 100644
index 0000000..ef25a9d
--- /dev/null
+++ b/tests/test_transcripts.py
@@ -0,0 +1,104 @@
+"""SW-01 transcript 持久化：ModelRouter 每次 LLM 调用的 prompt/response 落 SQLite。
+
+表结构对齐 Lab 仓 ADR-0001 §接口：(ts, caller, model, prompt, response,
+tokens_in, tokens_out, cost_usd, experiment_id)。库位置默认 out/transcripts.db，
+可用 NSC_TRANSCRIPT_DB / 构造参数改道（Lab 经 subprocess 调 SW 时靠环境变量接线）。
+"""
+
+from __future__ import annotations
+
+import json
+import sqlite3
+import types
+from pathlib import Path
+
+import pytest
+
+
+class _FakeResp:
+    """litellm.completion 的最小桩（无网络）。"""
+
+    def __init__(self) -> None:
+        msg = types.SimpleNamespace(content='{"ok": 1}', reasoning_content="")
+        self.choices = [types.SimpleNamespace(message=msg)]
+        self.usage = types.SimpleNamespace(prompt_tokens=11, completion_tokens=7)
+
+
+@pytest.fixture()
+def fake_litellm(monkeypatch):
+    calls: list[dict] = []
+
+    def _completion(**kwargs):
+        calls.append(kwargs)
+        return _FakeResp()
+
+    import litellm
+
+    monkeypatch.setattr(litellm, "completion", _completion)
+    return calls
+
+
+def _rows(db: Path) -> list[dict]:
+    conn = sqlite3.connect(str(db))
+    try:
+        cur = conn.execute(
+            "SELECT ts, caller, model, prompt, response, tokens_in, tokens_out,"
+            " cost_usd, experiment_id FROM transcripts"
+        )
+        cols = [c[0] for c in cur.description or []]
+        return [dict(zip(cols, r, strict=True)) for r in cur.fetchall()]
+    finally:
+        conn.close()
+
+
+def test_complete_writes_transcript_row(tmp_path, fake_litellm, monkeypatch):
+    from nsc.runtime.models import ModelRouter
+
+    db = tmp_path / "t.db"
+    monkeypatch.setenv("NSC_TRANSCRIPT_DB", str(db))
+    router = ModelRouter(experiment_id="exp-42")
+    res = router.complete(
+        "tier_bulk",
+        [
+            {"role": "system", "content": "sys"},
+            {"role": "user", "content": "hello"},
+        ],
+        json_mode=True,
+        seed=1,
+    )
+    assert res.text == '{"ok": 1}'
+
+    rows = _rows(db)
+    assert len(rows) == 1
+    row = rows[0]
+    assert row["caller"] == "tier_bulk"
+    # 不硬编码具体模型 ID(models.yaml 会演化):断言落库值 == 路由实际选中的模型
+    assert row["model"] == router.resolve("tier_bulk")["model"]
+    assert json.loads(row["prompt"]) == [
+        {"role": "system", "content": "sys"},
+        {"role": "user", "content": "hello"},
+    ]
+    assert row["response"] == '{"ok": 1}'
+    assert row["tokens_in"] == 11
+    assert row["tokens_out"] == 7
+    assert row["experiment_id"] == "exp-42"
+    assert row["cost_usd"] > 0  # cost_per_mtok 兜底价
+
+
+def test_transcript_db_via_constructor(tmp_path, fake_litellm, monkeypatch):
+    from nsc.runtime.models import ModelRouter
+
+    monkeypatch.delenv("NSC_TRANSCRIPT_DB", raising=False)
+    db = tmp_path / "ctor.db"
+    router = ModelRouter(transcript_db=db)
+    router.complete("tier_bulk", [{"role": "user", "content": "hi"}])
+    assert len(_rows(db)) == 1
+
+
+def test_transcript_failure_never_breaks_routing(tmp_path, fake_litellm, monkeypatch):
+    """transcript 是 best-effort 台账：写库失败不得影响 LLM 路由本身。"""
+    from nsc.runtime.models import ModelRouter
+
+    router = ModelRouter(transcript_db=tmp_path)  # 目录当 db 路径 → 打不开
+    res = router.complete("tier_bulk", [{"role": "user", "content": "hi"}])
+    assert res.text == '{"ok": 1}'