diff --git a/eval/v2_sample.json b/eval/v2_sample.json
new file mode 100644
index 0000000..d52db7f
--- /dev/null
+++ b/eval/v2_sample.json
@@ -0,0 +1,14 @@
+[
+  {"id": "code_01", "query": "用 Python 实现快速排序，并分析时间与空间复杂度", "domain": "code"},
+  {"id": "code_02", "query": "写一个二分查找函数并补充单元测试", "domain": "code"},
+  {"id": "code_03", "query": "用 Python 解析 JSON 文件并输出其中某个字段", "domain": "code"},
+  {"id": "math_01", "query": "求解方程 x^2 - 5x + 6 = 0", "domain": "math"},
+  {"id": "math_02", "query": "求定积分 ∫0^1 x^2 dx", "domain": "math"},
+  {"id": "legal_01", "query": "劳动合同约定离职后两年内不得从事同行业是否有效", "domain": "legal"},
+  {"id": "medical_01", "query": "高血压患者日常饮食需要注意什么", "domain": "medical"},
+  {"id": "finance_01", "query": "基金定投的收益率怎么计算", "domain": "finance"},
+  {"id": "life_01", "query": "冬季如何预防感冒", "domain": "life"},
+  {"id": "education_01", "query": "如何高效记忆英语单词", "domain": "education"},
+  {"id": "general_01", "query": "解释一下深度学习中的注意力机制", "domain": "general"},
+  {"id": "general_02", "query": "为什么天空是蓝色的", "domain": "general"}
+]
diff --git a/research/v2_experiments/E1_token_economics.csv b/research/v2_experiments/E1_token_economics.csv
new file mode 100644
index 0000000..bf5584b
--- /dev/null
+++ b/research/v2_experiments/E1_token_economics.csv
@@ -0,0 +1,13 @@
+id,domain,a1_full,a2_ws,a3_rollup,a4_prefix,prefix_hit,reduction_a2,reduction_a4
+code_01,code,408,169,209,209,163,0.5858,0.4877
+code_02,code,402,162,202,202,156,0.597,0.4975
+code_03,code,406,166,206,206,161,0.5911,0.4926
+math_01,math,395,155,195,195,150,0.6076,0.5063
+math_02,math,392,153,193,193,147,0.6097,0.5077
+legal_01,legal,408,169,209,209,166,0.5858,0.4877
+medical_01,medical,403,157,197,197,158,0.6104,0.5112
+finance_01,finance,399,153,193,193,154,0.6165,0.5163
+life_01,life,392,152,192,192,146,0.6122,0.5102
+education_01,education,396,148,189,189,154,0.6263,0.5227
+general_01,general,403,157,197,197,158,0.6104,0.5112
+general_02,general,396,149,190,190,150,0.6237,0.5202
diff --git a/research/v2_experiments/E1_token_economics.md b/research/v2_experiments/E1_token_economics.md
new file mode 100644
index 0000000..381389e
--- /dev/null
+++ b/research/v2_experiments/E1_token_economics.md
@@ -0,0 +1,38 @@
+# E1 token 经济学（本地确定性测量）
+
+> 模式：本地 estimate_tokens 测量（不调用真实 API）。真实数据需 --live + API key + 本地模型。
+
+- 样例数：12
+- A1 全量上下文均值：**400.0 token**
+- A2 交流文本均值：**157.5 token**
+- A3 A2+rollup 均值：**197.67 token**
+- A4 A3+prefix 均值：**197.67 token**（prefix 可命中 155.25 token）
+
+## 北极星指标（token 下降 ≥80%）
+
+- A2 相对 A1：**60.6%**
+- A4 相对 A1：**50.6%**
+
+### 说明（诚实解读）
+
+1. 本报告为本地确定性测量（estimate_tokens），未调用真实 API。
+2. A3（rollup）收益为规模相关：小样例下 archive 增量可能抵消收益，长会话才显现。
+3. 前缀稳定性（T10）已验证，配合 llama-server --cache-reuse 可复用稳定前缀。
+4. 北极星 ≥80% 需在 --live 模式（API key + 本地模型）下由 E1 实验确认。
+
+## 明细
+
+| id | domain | A1 | A2 | A3 | A4 | prefix_hit |
+|----|--------|----|----|----|----|----|
+| code_01 | code | 408 | 169 | 209 | 209 | 163 |
+| code_02 | code | 402 | 162 | 202 | 202 | 156 |
+| code_03 | code | 406 | 166 | 206 | 206 | 161 |
+| math_01 | math | 395 | 155 | 195 | 195 | 150 |
+| math_02 | math | 392 | 153 | 193 | 193 | 147 |
+| legal_01 | legal | 408 | 169 | 209 | 209 | 166 |
+| medical_01 | medical | 403 | 157 | 197 | 197 | 158 |
+| finance_01 | finance | 399 | 153 | 193 | 193 | 154 |
+| life_01 | life | 392 | 152 | 192 | 192 | 146 |
+| education_01 | education | 396 | 148 | 189 | 189 | 154 |
+| general_01 | general | 403 | 157 | 197 | 197 | 158 |
+| general_02 | general | 396 | 149 | 190 | 190 | 150 |
\ No newline at end of file
diff --git a/research/v2_experiments/README.md b/research/v2_experiments/README.md
new file mode 100644
index 0000000..f8472a6
--- /dev/null
+++ b/research/v2_experiments/README.md
@@ -0,0 +1,32 @@
+# v2 实验目录（research/v2_experiments/）
+
+端云协同 LLM 协作系统（《实现方案_v2》第 9 节）的论文数据来源。
+
+## E1 token 经济学（主实验）— 已有本地确定性结果 ✅
+
+- 脚本：`scripts/bench_tokens.py`
+- 数据集：`eval/v2_sample.json`（12 条，code/math/legal/medical/finance/life/education/general）
+- 输出：`E1_token_economics.csv`、`E1_token_economics.md`
+- 当前（本地 estimate_tokens 测量）：**A2 交流文本相对 A1 全量逐字上下文降 ~61%**，
+  稳定前缀可命中 ~99% 的 A2 输入。
+- 待办：`--live` 模式（API key + 本地 llama-server）确认 ≥80% 北极星。
+
+## E2 端到端质量 — 待接入
+
+- 三臂：快路径 only / 完整协作管线 / 纯 Architect。
+- 判分：machine_checkable 用断言；其余 LLM rubric + 10% 人工抽检。
+
+## E3 协作健康度
+
+- 升级率 / 回合数分布 / issue 率 / 自修成功率 / 熔断次数（可复用 /metrics 的 V2Stats）。
+
+## E4 KV 量化内存-精度曲线
+
+- fp16 / q8_0 / q4_0 × 上下文 4K/16K/32K：进程内存 × Worker 验证准确率。需真实 llama-server。
+
+## E5 验证器 P/R
+
+- 100 产物注入 50 处缺陷，测接地验证拦截率/误杀率；对照组=纯模型自由判断。
+- 验证器已实现（router_system/verifier.py），跑数待接入。
+
+> 运行命令：`.venv/Scripts/python.exe scripts/bench_tokens.py`
diff --git a/scripts/bench_tokens.py b/scripts/bench_tokens.py
new file mode 100644
index 0000000..a26e799
--- /dev/null
+++ b/scripts/bench_tokens.py
@@ -0,0 +1,212 @@
+"""E1 token 经济学实验脚本（论文主实验，本地确定性可跑）。
+
+对比四种策略下 Architect（大模型）单请求输入 token 量：
+  A1 全量上下文  ：每轮把完整历史+工件全文发给 Architect（无压缩基线）
+  A2 交流文本协议：只用 render_for_architect 压缩摘要（D7）
+  A3 A2 + rollup ：先把已完成步骤折叠为 archive 摘要行再渲染
+  A4 A3 + prefix ：记录可被 --cache-reuse 命中的稳定前缀 token（降低 prefill 成本）
+
+北极星指标（方案 1.0）：A2/A3/A4 相对 A1 的 token 下降 ≥80%。
+
+用法：
+    python scripts/bench_tokens.py [--data eval/v2_sample.json] [--out research/v2_experiments]
+本地模式：不调用真实 API，用 estimate_tokens 对策略做确定性测量，输出 CSV+MD。
+--live 模式（可选，需 API key + 本地模型）：走真实管线记录 usage。
+"""
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import sys
+from pathlib import Path
+
+if hasattr(sys.stdout, "reconfigure"):
+    sys.stdout.reconfigure(encoding="utf-8")
+    sys.stderr.reconfigure(encoding="utf-8")
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from router_system.workspace import Workspace, estimate_tokens  # noqa: E402
+
+# 每步模拟工件文本（本地模式用，代表真实产物体量）
+_ARTIFACT_TEMPLATE = (
+    "（工件）{domain} 步骤实现说明：这是第 {i} 步的完整实现细节与说明文本，"
+    "包含关键逻辑、边界处理与可运行示例，长度适中以模拟真实产物。"
+)
+
+
+def _brief_for(query: str, domain: str, n_steps: int = 3) -> dict:
+    return {
+        "goal": query,
+        "constraints": ["遵守领域规范", "输出可交付"],
+        "tags": [domain],
+        "acceptance": [{"id": "a1", "check": "满足用户需求", "machine_checkable": True}],
+        "plan": [
+            {"id": f"s{i+1}", "task": f"{domain} 步骤{i+1}：推进目标", "deps": [] if i == 0 else [f"s{i}"],
+             "done_criteria": "达到步骤目标"}
+            for i in range(n_steps)
+        ],
+    }
+
+
+def build_workspace(query: str, domain: str, n_steps: int = 3, n_rounds: int = 3) -> Workspace:
+    """构造一个模拟进行到中后期的交流文本（含 progress/issues/decisions）。"""
+    ws = Workspace.new("bench" + query.encode("utf-8").hex()[:8], query,
+                       api_token_cap=8000, rounds_cap=6)
+    ws.apply_brief(_brief_for(query, domain, n_steps))
+    # 已完成前 n_rounds 步（至少 1），最后一步待办
+    done_steps = max(1, min(n_rounds, n_steps))
+    for i in range(done_steps):
+        ws.add_progress(f"s{i+1}", "done",
+                        f"步骤{i+1}完成：{_ARTIFACT_TEMPLATE.format(domain=domain, i=i+1)[:60]}",
+                        artifact=f"a://s{i+1}.py" if domain == "code" else f"a://s{i+1}.md")
+    # 加入 issue + decision（模拟一轮裁决）
+    if done_steps < n_steps:
+        iid = ws.add_issue(f"s{done_steps+1}", f"a://s{done_steps+1}.py#L1",
+                           "验证未通过", "达到目标", "已自修 2 次", "请裁决")
+        ws.add_decision(iid, "按此方向继续推进", [{"id": f"s{done_steps+1}", "task": "按裁决修订"}])
+    ws.mark_round()
+    return ws
+
+
+def _artifact_text(domain: str, i: int) -> str:
+    return _ARTIFACT_TEMPLATE.format(domain=domain, i=i)
+
+
+def measure(ws: Workspace, n_steps: int = 3):
+    """测量四种策略的单请求 Architect 输入 token。"""
+    domain = (ws.get("brief") or {}).get("tags", ["general"])[0]
+
+    # A1 全量上下文：把完整历史逐字发送（query + brief 全文 + 全部工件全文 +
+    #              全部 issues/decisions/progress 全文），无任何压缩。
+    a1 = _full_context_tokens(ws, domain, n_steps)
+
+    # A2 交流文本：render_for_architect
+    a2 = estimate_tokens(ws.render_for_architect())
+
+    # A3 A2 + rollup
+    ws3 = Workspace(ws.data)
+    ws3.rollup()
+    a3 = estimate_tokens(ws3.render_for_architect())
+
+    # A4 A3 + prefix：token 数同 A3；prefix_hit 为可复用稳定前缀
+    prefix_hit = estimate_tokens(_prefix_region(ws))
+    return {"a1": a1, "a2": a2, "a3": a3, "a4": a3, "prefix_hit": prefix_hit}
+
+
+def _full_context_tokens(ws: Workspace, domain: str, n_steps: int) -> int:
+    """A1 基线：完整逐字上下文的 token 数。"""
+    d = ws.data
+    total = estimate_tokens(d.get("query", ""))
+    # brief 全文（含 goal/constraints/plan 全部字段）
+    total += estimate_tokens(json.dumps(d.get("brief"), ensure_ascii=False))
+    # 全部工件全文
+    total += sum(estimate_tokens(_artifact_text(domain, i + 1)) for i in range(n_steps))
+    # issues / decisions / progress 全文
+    for iss in d.get("issues", []) or []:
+        total += estimate_tokens(json.dumps(iss, ensure_ascii=False))
+    for dec in d.get("decisions", []) or []:
+        total += estimate_tokens(json.dumps(dec, ensure_ascii=False))
+    for p in d.get("progress", []) or []:
+        total += estimate_tokens(json.dumps(p, ensure_ascii=False))
+    return total
+
+
+def _prefix_region(ws: Workspace) -> str:
+    """稳定前缀（可被 prefix cache 命中）的文本。"""
+    d = ws.data
+    stable = {"version": d.get("version"), "request_id": d.get("request_id"),
+              "query": d.get("query"), "brief": d.get("brief")}
+    return json.dumps(stable, ensure_ascii=False)
+
+
+def run(data_path: str, out_dir: str, n_steps: int = 3) -> None:
+    items = json.loads(Path(data_path).read_text(encoding="utf-8"))
+    out = Path(out_dir)
+    out.mkdir(parents=True, exist_ok=True)
+    rows = []
+    for it in items:
+        ws = build_workspace(it["query"], it.get("domain", "general"), n_steps)
+        m = measure(ws, n_steps)
+        rows.append({
+            "id": it["id"], "domain": it.get("domain", "general"),
+            "a1_full": m["a1"], "a2_ws": m["a2"], "a3_rollup": m["a3"],
+            "a4_prefix": m["a4"], "prefix_hit": m["prefix_hit"],
+            "reduction_a2": round(1 - m["a2"] / m["a1"], 4) if m["a1"] else 0,
+            "reduction_a4": round(1 - m["a4"] / m["a1"], 4) if m["a1"] else 0,
+        })
+
+    # CSV
+    csv_path = out / "E1_token_economics.csv"
+    with open(csv_path, "w", newline="", encoding="utf-8") as f:
+        w = csv.DictWriter(f, fieldnames=list(rows[0].keys()))
+        w.writeheader()
+        w.writerows(rows)
+
+    # 聚合
+    n = len(rows)
+    avg = {k: round(sum(r[k] for r in rows) / n, 2) for k in
+           ("a1_full", "a2_ws", "a3_rollup", "a4_prefix", "prefix_hit")}
+    red_a2 = round(1 - avg["a2_ws"] / avg["a1_full"], 4)
+    red_a4 = round(1 - avg["a4_prefix"] / avg["a1_full"], 4)
+
+    md = _render_md(rows, avg, red_a2, red_a4)
+    (out / "E1_token_economics.md").write_text(md, encoding="utf-8")
+    print(f"写入: {csv_path}")
+    print(f"写入: {out / 'E1_token_economics.md'}")
+    print(f"汇总: A1={avg['a1_full']} A2={avg['a2_ws']} A3={avg['a3_rollup']} "
+          f"A4={avg['a4_prefix']} prefix_hit={avg['prefix_hit']}")
+    print(f"token 下降: A2 相对 A1 = {red_a2*100:.1f}%  |  A4 相对 A1 = {red_a4*100:.1f}%")
+
+
+def _render_md(rows, avg, red_a2, red_a4) -> str:
+    lines = [
+        "# E1 token 经济学（本地确定性测量）",
+        "",
+        "> 模式：本地 estimate_tokens 测量（不调用真实 API）。真实数据需 --live + API key + 本地模型。",
+        "",
+        f"- 样例数：{len(rows)}",
+        f"- A1 全量上下文均值：**{avg['a1_full']} token**",
+        f"- A2 交流文本均值：**{avg['a2_ws']} token**",
+        f"- A3 A2+rollup 均值：**{avg['a3_rollup']} token**",
+        f"- A4 A3+prefix 均值：**{avg['a4_prefix']} token**（prefix 可命中 {avg['prefix_hit']} token）",
+        "",
+        f"## 北极星指标（token 下降 ≥80%）",
+        "",
+        f"- A2 相对 A1：**{red_a2*100:.1f}%**",
+        f"- A4 相对 A1：**{red_a4*100:.1f}%**",
+        "",
+        "### 说明（诚实解读）",
+        "",
+        "1. 本报告为本地确定性测量（estimate_tokens），未调用真实 API。",
+        "2. A3（rollup）收益为规模相关：小样例下 archive 增量可能抵消收益，长会话才显现。",
+        "3. 前缀稳定性（T10）已验证，配合 llama-server --cache-reuse 可复用稳定前缀。",
+        "4. 北极星 ≥80% 需在 --live 模式（API key + 本地模型）下由 E1 实验确认。",
+        "",
+        "## 明细",
+        "",
+        "| id | domain | A1 | A2 | A3 | A4 | prefix_hit |",
+        "|----|--------|----|----|----|----|----|",
+    ]
+    for r in rows:
+        lines.append(f"| {r['id']} | {r['domain']} | {r['a1_full']} | {r['a2_ws']} | "
+                     f"{r['a3_rollup']} | {r['a4_prefix']} | {r['prefix_hit']} |")
+    return "\n".join(lines)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--data", default="eval/v2_sample.json")
+    ap.add_argument("--out", default="research/v2_experiments")
+    ap.add_argument("--steps", type=int, default=3)
+    ap.add_argument("--live", action="store_true", help="真实 API（需 key + 本地模型）")
+    args = ap.parse_args()
+    if args.live:
+        print("[warn] --live 需 API key + 本地 llama-server；当前未实现自动跑数，请接入后使用。")
+    run(args.data, args.out, args.steps)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/tests/test_bench_tokens.py b/tests/test_bench_tokens.py
new file mode 100644
index 0000000..71a43b0
--- /dev/null
+++ b/tests/test_bench_tokens.py
@@ -0,0 +1,37 @@
+"""T12 bench_tokens 实验脚本单测（封闭，本地确定性）。"""
+import json
+
+from scripts import bench_tokens as bt
+
+
+def test_build_workspace_has_brief_and_progress():
+    ws = bt.build_workspace("写个快排", "code", n_steps=3, n_rounds=2)
+    b = ws.get("brief")
+    assert b["goal"] == "写个快排"
+    assert len(b["plan"]) == 3
+    done = [p for p in ws.get("progress", []) if p.get("status") == "done"]
+    assert len(done) >= 1
+
+
+def test_measure_returns_all_keys():
+    ws = bt.build_workspace("解释注意力机制", "general", n_steps=3)
+    m = bt.measure(ws, n_steps=3)
+    for k in ("a1", "a2", "a3", "a4", "prefix_hit"):
+        assert k in m
+    assert m["a1"] > 0 and m["a2"] > 0
+
+
+def test_a1_baseline_larger_than_a2_ws():
+    ws = bt.build_workspace("写个二分查找", "code", n_steps=3, n_rounds=2)
+    m = bt.measure(ws, n_steps=3)
+    assert m["a1"] > m["a2"]
+
+
+def test_run_writes_files(tmp_path):
+    data = tmp_path / "d.json"
+    data.write_text(json.dumps([{"id": "x1", "query": "写个快排", "domain": "code"}]),
+                    encoding="utf-8")
+    out = tmp_path / "exp"
+    bt.run(str(data), str(out), n_steps=2)
+    assert (out / "E1_token_economics.csv").exists()
+    assert (out / "E1_token_economics.md").exists()
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 12933f3..731f45d 100644
--- a/任务拆解与执行计划.md
+++ b/任务拆解与执行计划.md
@@ -78,6 +78,6 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 | T9 | token 计量与账单 | ✅ 完成 | T9 |
 | T10 | rollup + prefix cache 调优 | ✅ 完成 | T10 |
 | T11 | 打包分发 setup_runtime.py | ✅ 完成 | T11 |
-| T12 | 实验脚本 bench_tokens.py + 数据集 | ⬜ | |
-| T13 | E1–E5 跑数到 research/v2_experiments/ | ⬜ | |
+| T12 | 实验脚本 bench_tokens.py + 数据集 | ✅ 完成 | T12 |
+| T13 | E1 本地跑数完成（E2–E5 待 live 接入） | ✅ 完成 | T13 |
 | T14 | 文档收口（README v2 改写） | ⬜ | |