@furongjun1999/dsh-memory 0.5.0 → 0.6.0

package/md_cg/bench_e2e_judge.py

@@ -0,0 +1,532 @@
+# -*- coding: utf-8 -*-
+"""端到端干扰池评测：确定性裁决层 vs LLM-as-judge（2026-09-23 · 外部建议里程碑）
+
+评测问题：公开数据集 locomo-zh-500 上，当候选池里「错误答案长得像正确答案」时——
+  arm_base      裸五路融合排序（无裁决对照）
+  arm_firewall  灵枢确定性裁决层（judge_qualification 四态 + judge_ranking 防火墙）
+  arm_llm       LLM-as-judge（同一候选面逐卡资格裁决，语义阅读）
+  三臂谁能把 gold 守在 top1？退化曲线随干扰浓度如何变化？
+
+公平性三原则：
+  1) 同管线：gold 与全部合成干扰走同一 CCG 六要素写入函数，生效条件由节点自身
+     字段机械合成（主体/时间窗口/主题），不看题面——两个裁决器都无题面捷径可抄。
+  2) 同候选：三臂共享同一融合候选（五路 sum RRF）；firewall 与 LLM 都只做资格
+     裁决，裁决机制不同（词面条件四态 vs 语义阅读）。
+  3) 可证伪：合成干扰不设任何「方便防火墙剔除」的标记（不适用条件一律诚实写
+     「无」）——防火墙若无效、LLM 若翻车，都如实报，不粉饰。
+
+干扰家族（每题合成、按家族顺序截断到 level 条；id 前缀 x_ 保证排在 gold 之后，
+  平局兜底偏向 gold——对我方保守）：
+  D2 实体互换  摘要中双方名字互换 + 主体换人：「谁说的/谁做的」答案错，词面几乎同款
+  D3p/D3m 时间错位  同一事实 ±1 年：查询不含时间词时原则上不可区分（诚实边界）
+  D4 否定澄清  「澄清：此前所说「…」并不属实」：否认该事实，词面高度重叠
+  D1 同话题硬负例不合成——语料其余 566 条 turn 即天然干扰池，全 level 在场
+
+浓度梯度：level∈{0,1,2,4} = 每题注入前 level 条合成干扰（500 题全部注入、池共享；
+  level=0 即零干扰锚点池）。家族级归因在 level=4 池上做。
+
+指标：hit@1/5/10、MRR（各臂排序）；top1 置换归因（gold/x_d2/x_d3/x_d4/天然）；
+  gold 被合成干扰压过率；LLM 臂加报 gold-qualified 率 / 弃权率 / 两轮自一致性。
+
+诚实边界：
+  · 本评测为 CCG 化写入（已发布 locomo 主链路成绩是五槽裸串口径），零干扰锚点
+    与 94.6% 只做量级对照，不做逐位复现声明。
+  · 题面为关键词串（与检索同源）；LLM 拿到的查询与引擎完全同面。位置偏差以
+    逐位 qualified 率诊断呈现，v1 不做候选旋转。
+  · LLM temperature=0 仍非严格确定，以两轮独立调用的一致性量化；结果含缓存
+    （--data-root/llm_cache，键含轮次），重跑命中缓存不重复计费。
+
+跑法：
+  python -X utf8 -m md_cg.bench_e2e_judge --quick        # 冒烟 20题×level 0,1×无LLM
+  python -X utf8 -m md_cg.bench_e2e_judge                # 全量 120题×4级×3臂+LLM两轮
+  python -X utf8 -m md_cg.bench_e2e_judge --skip-llm     # 只跑确定性两臂
+环境：DEEPSEEK_API_KEY（LLM 臂）；数据落 --data-root（默认 D:/program/test/e2e_judge）
+"""
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import random
+import re
+import shutil
+import sys
+import time
+import urllib.request
+from concurrent.futures import ThreadPoolExecutor
+
+HERE = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+if HERE not in sys.path:
+    sys.path.insert(0, HERE)
+
+from md_cg.mdcos import MdCGOS                             # noqa: E402
+from md_cg import nodefile                                 # noqa: E402
+
+DATA_SRC = os.path.join(HERE, "data", "benchmarks", "locomo-zh-500")
+DEFAULT_ROOT = r"D:\program\test\e2e_judge"
+FUSION_KW = dict(paths=("lexical", "bucket", "entity", "graph", "fuzzy"),
+                 fusion="sum")
+LEVELS_DEFAULT = (0, 1, 2, 4)
+FAMILIES = ("D2", "D3p", "D3m", "D4")       # 每题合成顺序；level 截断
+FAM_PREFIX = {"D2": "x_d2", "D3p": "x_d3p", "D3m": "x_d3m", "D4": "x_d4"}
+
+
+# ---------------------------------------------------------------- 数据装载
+def load_src():
+    corpus = [json.loads(ln) for ln
+              in open(os.path.join(DATA_SRC, "corpus567.jsonl"),
+                      encoding="utf-8") if ln.strip()]
+    questions = [json.loads(ln) for ln
+                 in open(os.path.join(DATA_SRC, "questions500.jsonl"),
+                         encoding="utf-8") if ln.strip()]
+    return corpus, questions
+
+
+def speakers_by_scene(corpus):
+    m = {}
+    for c in corpus:
+        scene = c["id"].split("_session")[0]
+        m.setdefault(scene, set()).add(c.get("speaker") or "")
+    return {k: sorted(v) for k, v in m.items() if len(v) >= 2}
+
+
+# ---------------------------------------------------------------- CCG 写入面
+def ccg_card(summary, identity, time_s, terms, node_id):
+    """gold 与合成干扰共用同一写入函数（公平性原则 1）。
+
+    生效条件只含节点自身字段（主体/时间窗口/主题），不含任何题面信息；
+    不适用条件一律诚实写「无」——合成干扰不携带可被负条件路剔除的标记。
+    """
+    terms_s = "、".join(terms)
+    lines = [f"{summary}（{identity}，{time_s}）",
+             f"# 功能名：会话记忆 {node_id}",
+             f"# 生效条件：主体={identity}；时间窗口={time_s}；主题={terms_s}",
+             "# 子功能：episodic 会话片段（对话摘要转写）",
+             "# 执行：陈述会话事实",
+             "# 验证方式：据会话记录核对",
+             "# 不适用条件：无"]
+    return "\n".join(lines) + "\n"
+
+
+def _shift_year(time_s, delta):
+    return re.sub(r"(\d{4})", lambda m: str(int(m.group(1)) + delta),
+                  str(time_s), count=1)
+
+
+def _swap_names(text, a, b):
+    return text.replace(a, "\x00").replace(b, a).replace("\x00", b)
+
+
+def synth_for_question(q, corpus_map, spk):
+    """一题的合成干扰（全家族），返回 [(fam, node_id, card), ...]。"""
+    evs = [corpus_map[t] for t in (q.get("evidence_turns") or [])
+           if t in corpus_map]
+    if not evs:
+        return []
+    ev = evs[0]
+    zf = ev.get("zh_fields") or {}
+    summary = str(zf.get("summary") or ev.get("zh") or "")
+    identity = str(zf.get("identity") or ev.get("speaker") or "")
+    time_s = str(zf.get("time") or "")
+    terms = [str(t) for t in (zf.get("terms") or [])]
+    scene = ev["id"].split("_session")[0]
+    pair = spk.get(scene) or []
+    other = next((s for s in pair if s and s != identity), None)
+    qid = q["qid"]
+    out = []
+    if other:                                   # D2 实体互换
+        s2 = _swap_names(summary, identity, other)
+        t2 = [_swap_names(t, identity, other) for t in terms]
+        out.append(("D2", f"{FAM_PREFIX['D2']}_{qid}",
+                    ccg_card(s2, other, time_s, t2, f"{FAM_PREFIX['D2']}_{qid}")))
+    if time_s:                                  # D3p / D3m 时间错位
+        for fam, d in (("D3p", 1), ("D3m", -1)):
+            t3 = _shift_year(time_s, d)
+            nid = f"{FAM_PREFIX[fam]}_{qid}"
+            out.append((fam, nid, ccg_card(summary, identity, t3, terms, nid)))
+    s4 = f"澄清：此前所说「{summary}」并不属实"     # D4 否定澄清
+    nid = f"{FAM_PREFIX['D4']}_{qid}"
+    out.append(("D4", nid, ccg_card(s4, identity, time_s, terms, nid)))
+    return out
+
+
+# ---------------------------------------------------------------- 池构建
+def build_pool(level, corpus, questions, corpus_map, spk, root_dir):
+    """level=每题注入的合成干扰条数（按 FAMILIES 顺序截断）。带指纹缓存。"""
+    synth_all = []          # 全 500 题都合成（池共享、更真实）
+    for q in questions:
+        synth_all.append(synth_for_question(q, corpus_map, spk))
+    n_synth = sum(len(s[:level]) for s in synth_all)
+    fp_src = hashlib.md5(
+        json.dumps([len(corpus), len(questions)]).encode()).hexdigest()[:8]
+    root = os.path.join(root_dir, "pools", f"level{level}")
+    manifest = os.path.join(root, "_manifest.json")
+    want = {"fp": fp_src, "level": level, "corpus": len(corpus),
+            "synth": n_synth}
+    if os.path.isfile(manifest):
+        try:
+            if json.load(open(manifest, encoding="utf-8")) == want \
+               and os.path.isdir(os.path.join(root, "mem")):
+                return MdCGOS(os.path.join(root, "mem")), n_synth
+        except Exception:
+            pass                                   # 陈化即重建
+    if os.path.isdir(root):
+        shutil.rmtree(root)
+    os.makedirs(root, exist_ok=True)
+    cg = MdCGOS(os.path.join(root, "mem"))
+    cards = {}
+    for c in corpus:                              # gold 天然面
+        zf = c.get("zh_fields") or {}
+        nid = c["id"]
+        cards[nid] = ccg_card(str(zf.get("summary") or c.get("zh") or ""),
+                              str(zf.get("identity") or c.get("speaker") or ""),
+                              str(zf.get("time") or ""),
+                              [str(t) for t in (zf.get("terms") or [])], nid)
+    for s in synth_all:
+        for fam, nid, card in s[:level]:
+            assert nid not in cards, f"干扰 id 冲突 {nid}"
+            cards[nid] = card
+    bad = [nid for nid, cd in cards.items()
+           if not nodefile.ccg_completeness(cd)["complete"]]
+    assert not bad, f"CCG 六要素不全（不应发生）: {bad[:3]}"
+    for nid, card in cards.items():
+        cg.add(nid, card, layer="knowledge", verification_basis="data")
+    json.dump(want, open(manifest, "w", encoding="utf-8"))
+    return cg, n_synth
+
+
+# ---------------------------------------------------------------- 检索两臂
+def run_det_arms(cg, q, k=10):
+    gold = set(q.get("evidence_turns") or [])
+    base, _mb = cg.search_rrf(q["question"], k=k, judge=False, **FUSION_KW)
+    fw, _mf = cg.search_rrf(q["question"], k=k, judge=True,
+                            judge_ranking=True, **FUSION_KW)
+    return {"gold": gold,
+            "base": [r[0]["id"] for r in base],
+            "firewall": [r[0]["id"] for r in fw],
+            "fw_states": [((r[2] or {}).get("state") if len(r) > 2 else None)
+                          for r in fw],
+            "base_nodes": [r[0] for r in base],
+            "fw_nodes": [r[0] for r in fw]}
+
+
+# ---------------------------------------------------------------- LLM 裁决
+LLM_SYS = (
+    "你是记忆系统的资格裁决器。给定一个查询和若干条候选记忆卡，逐条判断该卡"
+    "是否有资格作为回答该查询的证据。判定标准：卡的内容真实包含查询所问的事实"
+    "（主题一致且事实成立；澄清/否认类内容不构成该事实的证据；主体或时间与查询"
+    "冲突时同样不合格）。\n"
+    "只输出一行 JSON，不要解释、不要代码围栏，格式：\n"
+    '{"verdicts":[{"no":1,"qualified":true,"reason":"不超过20字"},...]}'
+)
+
+
+def llm_chat(model, base, key, messages, timeout=120, max_tokens=1500,
+             temperature=0.0, extra_payload=None):
+    body = {"model": model, "messages": messages,
+            "temperature": temperature, "max_tokens": max_tokens}
+    if extra_payload:
+        body.update(extra_payload)
+    payload = json.dumps(body).encode("utf-8")
+    req = urllib.request.Request(
+        base.rstrip("/") + "/chat/completions", data=payload,
+        headers={"Content-Type": "application/json",
+                 "Authorization": f"Bearer {key}"})
+    t0 = time.time()
+    with urllib.request.urlopen(req, timeout=timeout) as resp:
+        data = json.loads(resp.read().decode("utf-8"))
+    content = data["choices"][0]["message"].get("content") or ""
+    return content, (data.get("usage") or {}), time.time() - t0
+
+
+def llm_user_prompt(q, cand_nodes):
+    lines = [f"查询：{q['question']}", "", "候选记忆卡："]
+    for i, nd in enumerate(cand_nodes, 1):
+        content = str(nd.get("content") or "")[:400]
+        lines.append(f"【{i}】{content}")
+        lines.append("")
+    lines.append("请逐卡输出资格判定 JSON。")
+    return "\n".join(lines)
+
+
+def parse_verdicts(raw, n):
+    """解析 LLM 裁决 JSON → {no: qualified}；解析失败返回 None（≠空判定）。"""
+    s = raw or ""
+    i, j = s.find("{"), s.rfind("}")
+    verdicts = {}
+    if i >= 0 and j > i:
+        try:
+            obj = json.loads(s[i:j + 1])
+            for v in obj.get("verdicts") or []:
+                no = v.get("no")
+                if isinstance(no, int) and 1 <= no <= n:
+                    verdicts[no] = bool(v.get("qualified"))
+        except ValueError:
+            pass
+    if not verdicts:                            # 解析失败=该题无裁决（计入 err）
+        return None
+    return verdicts
+
+
+def llm_judge_one(cfg, q, cand_nodes, rnd, cache_dir):
+    """一轮 LLM 资格裁决（含缓存）。返回 (qualified_ids, tokens, dt, cached)。"""
+    prompt = llm_user_prompt(q, cand_nodes)
+    key_hex = hashlib.md5(json.dumps(
+        [cfg["model"], cfg["base"], prompt, rnd], ensure_ascii=False)
+        .encode("utf-8")).hexdigest()
+    cpath = os.path.join(cache_dir, key_hex + ".json")
+    if os.path.isfile(cpath):
+        d = json.load(open(cpath, encoding="utf-8"))
+        return (d["qualified"], d["tokens"], 0.0, True)
+    raw, usage, dt = llm_chat(cfg["model"], cfg["base"], cfg["key"],
+                              [{"role": "system", "content": LLM_SYS},
+                               {"role": "user", "content": prompt}],
+                              timeout=cfg["timeout"])
+    v = parse_verdicts(raw, len(cand_nodes))
+    if v is None:
+        raise ValueError(f"LLM 输出不可解析：{raw[:120]!r}")
+    ids = [cand_nodes[no - 1]["id"] for no in sorted(v) if v[no]]
+    d = {"qualified": ids, "raw": raw[:800],
+         "tokens": int(usage.get("total_tokens") or 0), "round": rnd}
+    json.dump(d, open(cpath, "w", encoding="utf-8"), ensure_ascii=False)
+    return ids, d["tokens"], dt, False
+
+
+def llm_reorder(base_ids, qualified):
+    """qualified 优先（保持基序），其余殿后——与防火墙 keep-降权同构。"""
+    qs = [i for i in base_ids if i in set(qualified)]
+    rest = [i for i in base_ids if i not in set(qualified)]
+    return qs + rest
+
+
+# ---------------------------------------------------------------- 指标
+def rank_of(gold, ids):
+    return next((i for i, nid in enumerate(ids, 1) if nid in gold), 0)
+
+
+def classify(nid):
+    for fam, p in FAM_PREFIX.items():
+        if nid.startswith(p + "_"):
+            return fam
+    return "natural" if not nid.startswith("x_") else "x_other"
+
+
+def arm_metrics(rows, key):
+    n = len(rows)
+    st = {"h1": 0, "h5": 0, "h10": 0, "mrr": 0.0,
+          "top1": {"gold": 0, "D2": 0, "D3p": 0, "D3m": 0, "D4": 0,
+                   "natural": 0, "x_other": 0},
+          "gold_outranked_by_synth": 0}
+    for r in rows:
+        ids = r[key]
+        gold = r["gold"]
+        rk = rank_of(gold, ids)
+        if rk == 1:
+            st["h1"] += 1
+        if 0 < rk <= 5:
+            st["h5"] += 1
+        if 0 < rk <= 10:
+            st["h10"] += 1
+        if rk:
+            st["mrr"] += 1.0 / rk
+        st["top1"][classify(ids[0])] += 1 if ids else 0
+        gr = rk
+        synth_before_gold = any(
+            nid.startswith("x_")
+            for nid in ids[:(gr - 1 if gr else len(ids))])
+        if synth_before_gold:
+            st["gold_outranked_by_synth"] += 1
+    return st
+
+
+def fmt_row(name, st, n):
+    t1 = st["top1"]
+    return (f"  {name:<10} hit@1={100*st['h1']/n:5.1f}%  hit@5={100*st['h5']/n:5.1f}%"
+            f"  hit@10={100*st['h10']/n:5.1f}%  MRR={st['mrr']/n:.4f}"
+            f"  top1置换[D2={t1['D2']} D3={t1['D3p']+t1['D3m']} D4={t1['D4']}"
+            f" 天然={t1['natural']}]")
+
+
+# ---------------------------------------------------------------- 主流程
+def main(argv=None):
+    ap = argparse.ArgumentParser(
+        description="端到端干扰池评测：确定性裁决层 vs LLM-as-judge")
+    ap.add_argument("--data-root", default=DEFAULT_ROOT)
+    ap.add_argument("--n", type=int, default=120, help="评测题数（种子采样）")
+    ap.add_argument("--seed", type=int, default=7)
+    ap.add_argument("--levels", default="0,1,2,4")
+    ap.add_argument("--arms", default="base,firewall,llm")
+    ap.add_argument("--rounds", type=int, default=2, help="LLM 独立轮数")
+    ap.add_argument("--model", default="deepseek-chat")
+    ap.add_argument("--base-url", default="https://api.deepseek.com/v1")
+    ap.add_argument("--key", default="")
+    ap.add_argument("--workers", type=int, default=6)
+    ap.add_argument("--timeout", type=int, default=120)
+    ap.add_argument("--k", type=int, default=10)
+    ap.add_argument("--quick", action="store_true", help="冒烟：20题×0,1级×无LLM")
+    ap.add_argument("--skip-llm", action="store_true")
+    a = ap.parse_args(argv)
+
+    if a.quick:
+        a.n, a.levels, a.skip_llm = 20, "0,1", True
+    levels = [int(x) for x in a.levels.split(",") if x.strip() != ""]
+    arms = [x.strip() for x in a.arms.split(",") if x.strip()]
+    use_llm = ("llm" in arms) and not a.skip_llm
+    key = a.key or os.environ.get("DEEPSEEK_API_KEY") or ""
+    if use_llm and not key:
+        print("[error] LLM 臂需要 DEEPSEEK_API_KEY 或 --key（或 --skip-llm）")
+        return 2
+
+    os.makedirs(os.path.join(a.data_root, "results"), exist_ok=True)
+    cache_dir = os.path.join(a.data_root, "llm_cache")
+    os.makedirs(cache_dir, exist_ok=True)
+
+    corpus, questions = load_src()
+    corpus_map = {c["id"]: c for c in corpus}
+    spk = speakers_by_scene(corpus)
+    rnd = random.Random(a.seed)
+    sample = sorted(rnd.sample(questions, min(a.n, len(questions))),
+                    key=lambda q: q["qid"])
+    print(f"[e2e_judge] 语料 {len(corpus)} · 题目 {len(questions)} · "
+          f"采样 {len(sample)}（seed={a.seed}） · levels={levels} · arms={arms}"
+          + (f" · LLM={a.model}×{a.rounds}轮" if use_llm else " · 无LLM"))
+    print(f"[e2e_judge] 数据根：{a.data_root}")
+
+    all_out = {"meta": {"seed": a.seed, "n": len(sample), "levels": levels,
+                        "model": a.model if use_llm else None,
+                        "k": a.k}, "levels": {}}
+    for lv in levels:
+        t0 = time.time()
+        cg, n_synth = build_pool(lv, corpus, questions, corpus_map, spk,
+                                 a.data_root)
+        print(f"\n===== level {lv}（池={len(corpus)+n_synth} 节点，"
+              f"合成干扰 {n_synth} 条）· 建池 {time.time()-t0:.1f}s =====")
+        rows = []
+        for q in sample:
+            det = run_det_arms(cg, q, k=a.k)
+            rows.append({"qid": q["qid"], "qtype": q.get("qtype"),
+                         "question": q["question"], **det})
+        n = len(rows)
+        lv_out = {"pool": len(corpus) + n_synth, "synth": n_synth,
+                  "rows_count": n}
+        for arm in ("base", "firewall"):
+            if arm in arms:
+                st = arm_metrics(rows, arm)
+                print(fmt_row(arm, st, n))
+                lv_out[arm] = st
+        # 四态分布（firewall 诊断面）
+        if "firewall" in arms:
+            sd = {"ACCEPT": 0, "DEFER": 0, "REJECT": 0, "BLINDSPOT": 0}
+            for r in rows:
+                for s in r["fw_states"]:
+                    sd[s] = sd.get(s, 0) + 1
+            tot = sum(sd.values()) or 1
+            print("  firewall四态(top10): " + " ".join(
+                f"{k2}={v}({100*v/tot:.0f}%)" for k2, v in sd.items() if v))
+            lv_out["fw_states"] = sd
+        # LLM 臂
+        if use_llm:
+            cfg = {"model": a.model, "base": a.base_url, "key": key,
+                   "timeout": a.timeout}
+            llm_rows = []
+            t1 = time.time()
+
+            def work(args):
+                r, rd = args
+                qobj = {"qid": r["qid"], "question": r["question"]}
+                try:
+                    qids, toks, dt, cached = llm_judge_one(
+                        cfg, qobj, r["base_nodes"], rd, cache_dir)
+                    return {"qid": r["qid"], "round": rd, "qualified": qids,
+                            "tokens": toks, "latency": dt, "cached": cached,
+                            "order": llm_reorder(r["base"], qids),
+                            "gold": r["gold"], "base": r["base"]}
+                except Exception as exc:                   # noqa: BLE001
+                    return {"qid": r["qid"], "round": rd, "error":
+                            f"{type(exc).__name__}: {exc}", "gold": r["gold"],
+                            "base": r["base"], "order": r["base"],
+                            "qualified": None, "tokens": 0, "latency": 0.0,
+                            "cached": False}
+            tasks = [(r, rd) for rd in range(1, a.rounds + 1) for r in rows]
+            with ThreadPoolExecutor(max_workers=a.workers) as ex:
+                llm_rows = list(ex.map(work, tasks))
+            for rd in range(1, a.rounds + 1):
+                rr = [x for x in llm_rows if x["round"] == rd]
+                for x in rr:
+                    x_ref = next(y for y in rows if y["qid"] == x["qid"])
+                    x["order"] = (llm_reorder(x_ref["base"], x["qualified"])
+                                  if x["qualified"] is not None
+                                  else x_ref["base"])
+                rd_rows = [{"gold": x["gold"], "llm": x["order"]} for x in rr]
+                st = arm_metrics(
+                    [{"gold": d["gold"], "llm": d["llm"]} for d in rd_rows],
+                    "llm")
+                errs = sum(1 for x in rr if x.get("error"))
+                cached = sum(1 for x in rr if x.get("cached"))
+                toks = sum(x["tokens"] for x in rr)
+                print(fmt_row(f"llm(r{rd})", st, n)
+                      + f"  err={errs} cache={cached} tok={toks}")
+                lv_out[f"llm_r{rd}"] = st
+                lv_out[f"llm_r{rd}_diag"] = {"errors": errs,
+                                             "cached": cached, "tokens": toks}
+            # gold-qualified 率 + 弃权率（round1）
+            r1 = {x["qid"]: x for x in llm_rows if x["round"] == 1}
+            gq = ab = 0
+            for x in r1.values():
+                if x["qualified"] is None:
+                    continue
+                if not x["qualified"]:
+                    ab += 1
+                elif set(x["qualified"]) & x["gold"]:
+                    gq += 1
+            print(f"  llm诊断: gold获qualified={100*gq/n:.1f}%  "
+                  f"全弃权={100*ab/n:.1f}%")
+            lv_out["llm_diag"] = {"gold_qualified": gq, "abstain": ab}
+            # 两轮自一致性
+            if a.rounds >= 2:
+                r2 = {x["qid"]: x for x in llm_rows if x["round"] == 2}
+                same_top = agree_v = pair = 0
+                for qid, x1 in r1.items():
+                    x2 = r2.get(qid)
+                    if not x2 or x1["qualified"] is None \
+                       or x2["qualified"] is None:
+                        continue
+                    pair += 1
+                    if x1["order"][:1] == x2["order"][:1]:
+                        same_top += 1
+                    s1, s2 = set(x1["qualified"]), set(x2["qualified"])
+                    if s1 == s2:
+                        agree_v += 1
+                if pair:
+                    print(f"  llm自一致性: top1一致={100*same_top/pair:.1f}%  "
+                          f"qualified集一致={100*agree_v/pair:.1f}%  "
+                          f"(n={pair})")
+                lv_out["llm_selfconsistency"] = {
+                    "top1": same_top, "set": agree_v, "pairs": pair}
+            print(f"  LLM 臂耗时 {time.time()-t1:.1f}s")
+            # 逐题留档（gold 序列化-safe）
+            for x in llm_rows:
+                if isinstance(x.get("gold"), set):
+                    x["gold"] = sorted(x["gold"])
+            lv_out["llm_rows"] = llm_rows
+        all_out["levels"][lv] = lv_out
+        # 行级明细落盘（报告取证用；文件名含采样指纹，防不同规模运行互相覆盖）
+        detail = [{"qid": r["qid"], "qtype": r["qtype"], "gold":
+                   sorted(r["gold"]), "base": r["base"],
+                   "firewall": r["firewall"]} for r in rows]
+        json.dump(detail, open(os.path.join(
+            a.data_root, "results",
+            f"detail_L{lv}_n{len(sample)}_s{a.seed}.json"), "w",
+            encoding="utf-8"), ensure_ascii=False, indent=1)
+
+    out_path = os.path.join(a.data_root, "results",
+                            f"summary_{time.strftime('%Y%m%d_%H%M%S')}.json")
+    json.dump(all_out, open(out_path, "w", encoding="utf-8"),
+              ensure_ascii=False, indent=1)
+    print(f"\n[e2e_judge] 完成。汇总 → {out_path}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())