@hupan56/wlkj 3.3.13 → 3.3.15

package/templates/qoder/scripts/validation/eval/gradient_matrix.py

@@ -1,261 +1,261 @@
-"""gradient_matrix.py — T1.0 低模型梯度评测尺子（飞轮 ruler）。
-
-证卖点：工作流的 KG grounding 让【弱模型】不编造、召回真实体，追上强模型裸跑。
-
-矩阵 = {弱/中/强模型} × {裸, +grounding}，每格跨 10 个业务主题算：
-  - grounding 精度(P)  : 模型列出的"具体名称"里有多少是真 KG 实体（反幻觉）
-  - 召回(R)            : gold 实体被答出多少
-  - 幻觉率(H=1-P)      : 编造的占比
-  - token              : 单次总 token
-
-跑法（容器内）：
-  docker exec -i wlinkj-workspace-backend-1 python /app/scripts/validation/eval/gradient_matrix.py
-  ... --themes 考勤,排班 --json   # 子集/机器可读
-"""
-import os
-import re
-import sys
-import json
-import time
-import urllib.request
-import urllib.error
-
-# 容器内 /app 在 sys.path（from app.db import）
-for _p in ("/app", os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))):
-    if os.path.isdir(_p) and _p not in sys.path:
-        sys.path.insert(0, _p)
-
-from sqlalchemy import text  # noqa: E402
-from app.db import SessionLocal  # noqa: E402
-
-# ── 配置 ─────────────────────────────────────────────────────────────
-CHAT_URL = os.environ.get("ALI_CHAT_BASE", "").rstrip("/") + "/chat/completions"
-CHAT_KEY = os.environ.get("DASHSCOPE_API_KEY", "")
-MODELS = {"weak": "qwen-turbo", "mid": "qwen-plus", "strong": "qwen-max"}
-DEFAULT_THEMES = ["考勤", "排班", "加班", "车辆", "报警", "保险", "能耗", "预算", "导出", "审批"]
-RETRIEVE_K = 45  # +grounding 喂的检索条数（FW1: 30→45 覆盖 gold-40，解 R 封顶 30/40=0.75）
-SYS = "你是车辆物联网系统(ICS)的代码/业务知识助手。只回答你确信真实存在的系统对象，绝不编造名称。"
-
-
-def chat(model, prompt, max_tokens=600, timeout=90):
-    """同步 chat，返 (text, total_tokens)。失败重试 3 次。"""
-    body = json.dumps({
-        "model": model,
-        "messages": [{"role": "system", "content": SYS},
-                     {"role": "user", "content": prompt}],
-        "temperature": 0.0,
-        "max_tokens": max_tokens,
-    }).encode()
-    for attempt in range(3):
-        try:
-            req = urllib.request.Request(
-                CHAT_URL, data=body,
-                headers={"Authorization": f"Bearer {CHAT_KEY}", "Content-Type": "application/json"})
-            resp = urllib.request.urlopen(req, timeout=timeout)
-            d = json.loads(resp.read())
-            return d["choices"][0]["message"]["content"], int(d.get("usage", {}).get("total_tokens", 0))
-        except (urllib.error.URLError, TimeoutError, OSError, KeyError) as e:
-            if attempt < 2:
-                time.sleep(1.0 * (attempt + 1))
-            else:
-                return f"[ERR {repr(e)[:80]}]", 0
-
-
-def load_universe_and_gold(themes):
-    """载入实体名称全集(精度校验) + 每主题 gold 实体(召回校验)。"""
-    s = SessionLocal()
-    rows = _q(s, "SELECT canonical, cn FROM entities WHERE canonical IS NOT NULL")
-    UNI = set()
-    for canon, cn in rows:
-        if canon:
-            UNI.add(canon.strip())
-        if cn:
-            UNI.add(cn.strip())
-    GOLD = {}
-    for th in themes:
-        rs = _q(s, "SELECT canonical, cn, type FROM entities "
-                   "WHERE (cn ILIKE :p OR canonical ILIKE :p) LIMIT 40", {"p": f"%{th}%"})
-        GOLD[th] = [{"canonical": c, "cn": n, "type": t} for c, n, t in rs]
-    s.close()
-    return UNI, GOLD
-
-
-def retrieve_entities(theme, k=RETRIEVE_K):
-    """模拟工作流 KG 检索（wl-search 工具会做的事）：返回真实体 dict 列表。"""
-    s = SessionLocal()
-    rs = _q(s, "SELECT canonical, cn, type FROM entities "
-               "WHERE (cn ILIKE :p OR canonical ILIKE :p) LIMIT :k",
-            {"p": f"%{theme}%", "k": k})
-    s.close()
-    return [{"canonical": c, "cn": n, "type": t} for c, n, t in rs]
-
-
-def ctx_from_entities(ents):
-    """把检索实体格式化成喂给 LLM 的上下文文本。"""
-    out = []
-    for e in ents:
-        c, n, t = e["canonical"], e["cn"], e["type"]
-        out.append(f"{t}: {c}" + (f" ({n})" if n else ""))
-    return "\n".join(out)
-
-
-def _q(s, sql, params=None):
-    try:
-        return s.execute(text(sql), params or {}).fetchall()
-    except Exception:
-        try:
-            s.rollback()
-        except Exception:
-            pass
-        return []
-
-
-# ── prompt ───────────────────────────────────────────────────────────
-def prompt_bare(theme):
-    return (f"在我们的车辆物联网系统(ICS)里，找出和「{theme}」相关的具体对象"
-            f"（数据库表/Java类/接口/按钮/页面等）。\n"
-            f"请逐行列出你【确信真实存在】的具体名称，每行一个，不要解释，不要编造。"
-            f"不确定就少列。")
-
-
-def prompt_scaffold(theme, ctx):
-    return (f"在我们的车辆物联网系统(ICS)里，找出和「{theme}」相关的具体对象"
-            f"（数据库表/Java类/接口/按钮/页面等）。\n"
-            f"知识图谱已检索到以下相关实体（均为真实存在）：\n{ctx}\n\n"
-            f"请把以上检索结果里的实体【尽量全部逐行列出】，每行一个，"
-            f"直接用检索结果里的名称，不要自行筛选/合并/省略，不要解释。检索到几个就列几个。")
-
-
-# ── 评分 ─────────────────────────────────────────────────────────────
-_LINE_CLEAN = re.compile(r"^[\-\*\d\.\)、\s]+")
-_CODE_TOK = re.compile(r"[A-Za-z_][\w./\-:]{2,}")
-
-
-def _clean_lines(ans):
-    """'每行一个' 答案 → 清洗后的短行列表（模型声称存在的对象，每行一个）。"""
-    out = []
-    for l in ans.splitlines():
-        l = _LINE_CLEAN.sub("", l).strip()
-        if l and len(l) <= 60:
-            out.append(l)
-    return out
-
-
-def score(ans, gold, UNI, ground_cn):
-    """grounding 精度：每行算'真实'当其含某真代码标识符(∈UNI) 或 命中本格 ground_cn(中文)。
-    ground_cn = 该格可作'真实'依据的中文名集合(bare=gold, scaffold=retrieved)，bare/scaffold 同口径公平。
-    """
-    lines = _clean_lines(ans)
-    real_n = 0
-    for l in lines:
-        if set(_CODE_TOK.findall(l)) & UNI:
-            real_n += 1
-        elif any(g and g in l for g in ground_cn):
-            real_n += 1
-    precision = (real_n / len(lines)) if lines else 0.0
-    hit = 0
-    for g in gold:
-        c = g.get("canonical") or ""
-        n = g.get("cn") or ""
-        if (c and c in ans) or (n and n in ans):
-            hit += 1
-    recall = (hit / len(gold)) if gold else 0.0
-    return {
-        "precision": round(precision, 3),
-        "recall": round(recall, 3),
-        "halluc": round(1 - precision, 3),
-        "n_lines": len(lines),
-        "n_real": real_n,
-        "gold_hit": hit,
-        "gold_n": len(gold),
-    }
-
-
-def aggregate(rows):
-    n = len(rows) or 1
-    return {
-        "P": round(sum(r["precision"] for r in rows) / n, 3),
-        "R": round(sum(r["recall"] for r in rows) / n, 3),
-        "H": round(sum(r["halluc"] for r in rows) / n, 3),
-        "tokens": round(sum(r["tokens"] for r in rows) / n, 0),
-        "n": len(rows),
-    }
-
-
-def run(themes, verbose=True):
-    def log(*a):
-        if verbose:
-            print(*a)
-
-    if not CHAT_KEY:
-        raise RuntimeError("缺 DASHSCOPE_API_KEY")
-    log("[ruler] 载入 KG 全集 + gold…")
-    UNI, GOLD = load_universe_and_gold(themes)
-    log(f"[ruler] 实体名称全集 {len(UNI)} 条")
-    for th in themes:
-        log(f"  theme {th}: gold={len(GOLD[th])}")
-
-    matrix = {}
-    t0 = time.time()
-    for tier, model in MODELS.items():
-        for mode in ("bare", "scaffold"):
-            rows = []
-            for th in themes:
-                gold = GOLD[th]
-                if mode == "bare":
-                    ground_cn = {g["canonical"] for g in gold} | {g["cn"] for g in gold if g["cn"]}
-                    p = prompt_bare(th)
-                else:
-                    ents = retrieve_entities(th)
-                    ground_cn = {e["canonical"] for e in ents} | {e["cn"] for e in ents if e["cn"]}
-                    p = prompt_scaffold(th, ctx_from_entities(ents))
-                ans, tok = chat(model, p)
-                sc = score(ans, gold, UNI, ground_cn)
-                sc["tokens"] = tok
-                sc["theme"] = th
-                rows.append(sc)
-                log(f"  [{tier}/{mode}] {th}: P={sc['precision']:.2f} R={sc['recall']:.2f} "
-                    f"H={sc['halluc']:.2f} tok={tok} lines={sc['n_lines']}")
-            matrix[f"{tier}/{mode}"] = aggregate(rows)
-            log(f"[{tier}/{mode}] agg P={matrix[f'{tier}/{mode}']['P']:.3f} "
-                f"R={matrix[f'{tier}/{mode}']['R']:.3f} H={matrix[f'{tier}/{mode}']['H']:.3f} "
-                f"tok={matrix[f'{tier}/{mode}']['tokens']:.0f}")
-    log(f"[ruler] 总耗时 {int(time.time()-t0)}s")
-    return {"matrix": matrix, "themes": themes, "models": MODELS}
-
-
-def format_table(res):
-    m = res["matrix"]
-    lines = ["## 低模型梯度矩阵（grounding 精度P / 召回R / 幻觉H / token）",
-             "", "| 模型档 | 裸 P | 裸 R | 裸 H | 裸tok | +grounding P | +grounding R | +grounding H | +ground tok |",
-             "|---|---|---|---|---|---|---|---|---|"]
-    for tier in MODELS:
-        b = m[f"{tier}/bare"]
-        f = m[f"{tier}/scaffold"]
-        lines.append(f"| {tier}({MODELS[tier]}) | {b['P']:.2f} | {b['R']:.2f} | {b['H']:.2f} | {b['tokens']:.0f} | "
-                     f"{f['P']:.2f} | {f['R']:.2f} | {f['H']:.2f} | {f['tokens']:.0f} |")
-    # 卖点提炼
-    lines.append("")
-    weak_gain_p = m["weak/scaffold"]["P"] - m["weak/bare"]["P"]
-    weak_catchup = m["weak/scaffold"]["P"] - m["strong/bare"]["P"]
-    lines.append(f"- 弱模型 +grounding 精度提升: **{weak_gain_p*100:+.1f}pp**（反幻觉增益）")
-    lines.append(f"- 弱模型+grounding vs 强模型裸 精度差: **{weak_catchup*100:+.1f}pp**（>0=追上/反超=卖点成立）")
-    return "\n".join(lines)
-
-
-def main():
-    import argparse
-    ap = argparse.ArgumentParser()
-    ap.add_argument("--themes", default=",".join(DEFAULT_THEMES))
-    ap.add_argument("--json", action="store_true")
-    args = ap.parse_args()
-    themes = [t.strip() for t in args.themes.split(",") if t.strip()]
-    res = run(themes)
-    if args.json:
-        print(json.dumps(res, ensure_ascii=False, indent=2))
-    else:
-        print("\n" + format_table(res))
-
-
-if __name__ == "__main__":
-    main()
+"""gradient_matrix.py — T1.0 低模型梯度评测尺子（飞轮 ruler）。
+
+证卖点：工作流的 KG grounding 让【弱模型】不编造、召回真实体，追上强模型裸跑。
+
+矩阵 = {弱/中/强模型} × {裸, +grounding}，每格跨 10 个业务主题算：
+  - grounding 精度(P)  : 模型列出的"具体名称"里有多少是真 KG 实体（反幻觉）
+  - 召回(R)            : gold 实体被答出多少
+  - 幻觉率(H=1-P)      : 编造的占比
+  - token              : 单次总 token
+
+跑法（容器内）：
+  docker exec -i wlinkj-workspace-backend-1 python /app/scripts/validation/eval/gradient_matrix.py
+  ... --themes 考勤,排班 --json   # 子集/机器可读
+"""
+import os
+import re
+import sys
+import json
+import time
+import urllib.request
+import urllib.error
+
+# 容器内 /app 在 sys.path（from app.db import）
+for _p in ("/app", os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "..", "..", ".."))):
+    if os.path.isdir(_p) and _p not in sys.path:
+        sys.path.insert(0, _p)
+
+from sqlalchemy import text  # noqa: E402
+from app.db import SessionLocal  # noqa: E402
+
+# ── 配置 ─────────────────────────────────────────────────────────────
+CHAT_URL = os.environ.get("ALI_CHAT_BASE", "").rstrip("/") + "/chat/completions"
+CHAT_KEY = os.environ.get("DASHSCOPE_API_KEY", "")
+MODELS = {"weak": "qwen-turbo", "mid": "qwen-plus", "strong": "qwen-max"}
+DEFAULT_THEMES = ["考勤", "排班", "加班", "车辆", "报警", "保险", "能耗", "预算", "导出", "审批"]
+RETRIEVE_K = 45  # +grounding 喂的检索条数（FW1: 30→45 覆盖 gold-40，解 R 封顶 30/40=0.75）
+SYS = "你是车辆物联网系统(ICS)的代码/业务知识助手。只回答你确信真实存在的系统对象，绝不编造名称。"
+
+
+def chat(model, prompt, max_tokens=600, timeout=90):
+    """同步 chat，返 (text, total_tokens)。失败重试 3 次。"""
+    body = json.dumps({
+        "model": model,
+        "messages": [{"role": "system", "content": SYS},
+                     {"role": "user", "content": prompt}],
+        "temperature": 0.0,
+        "max_tokens": max_tokens,
+    }).encode()
+    for attempt in range(3):
+        try:
+            req = urllib.request.Request(
+                CHAT_URL, data=body,
+                headers={"Authorization": f"Bearer {CHAT_KEY}", "Content-Type": "application/json"})
+            resp = urllib.request.urlopen(req, timeout=timeout)
+            d = json.loads(resp.read())
+            return d["choices"][0]["message"]["content"], int(d.get("usage", {}).get("total_tokens", 0))
+        except (urllib.error.URLError, TimeoutError, OSError, KeyError) as e:
+            if attempt < 2:
+                time.sleep(1.0 * (attempt + 1))
+            else:
+                return f"[ERR {repr(e)[:80]}]", 0
+
+
+def load_universe_and_gold(themes):
+    """载入实体名称全集(精度校验) + 每主题 gold 实体(召回校验)。"""
+    s = SessionLocal()
+    rows = _q(s, "SELECT canonical, cn FROM entities WHERE canonical IS NOT NULL")
+    UNI = set()
+    for canon, cn in rows:
+        if canon:
+            UNI.add(canon.strip())
+        if cn:
+            UNI.add(cn.strip())
+    GOLD = {}
+    for th in themes:
+        rs = _q(s, "SELECT canonical, cn, type FROM entities "
+                   "WHERE (cn ILIKE :p OR canonical ILIKE :p) LIMIT 40", {"p": f"%{th}%"})
+        GOLD[th] = [{"canonical": c, "cn": n, "type": t} for c, n, t in rs]
+    s.close()
+    return UNI, GOLD
+
+
+def retrieve_entities(theme, k=RETRIEVE_K):
+    """模拟工作流 KG 检索（wl-search 工具会做的事）：返回真实体 dict 列表。"""
+    s = SessionLocal()
+    rs = _q(s, "SELECT canonical, cn, type FROM entities "
+               "WHERE (cn ILIKE :p OR canonical ILIKE :p) LIMIT :k",
+            {"p": f"%{theme}%", "k": k})
+    s.close()
+    return [{"canonical": c, "cn": n, "type": t} for c, n, t in rs]
+
+
+def ctx_from_entities(ents):
+    """把检索实体格式化成喂给 LLM 的上下文文本。"""
+    out = []
+    for e in ents:
+        c, n, t = e["canonical"], e["cn"], e["type"]
+        out.append(f"{t}: {c}" + (f" ({n})" if n else ""))
+    return "\n".join(out)
+
+
+def _q(s, sql, params=None):
+    try:
+        return s.execute(text(sql), params or {}).fetchall()
+    except Exception:
+        try:
+            s.rollback()
+        except Exception:
+            pass
+        return []
+
+
+# ── prompt ───────────────────────────────────────────────────────────
+def prompt_bare(theme):
+    return (f"在我们的车辆物联网系统(ICS)里，找出和「{theme}」相关的具体对象"
+            f"（数据库表/Java类/接口/按钮/页面等）。\n"
+            f"请逐行列出你【确信真实存在】的具体名称，每行一个，不要解释，不要编造。"
+            f"不确定就少列。")
+
+
+def prompt_scaffold(theme, ctx):
+    return (f"在我们的车辆物联网系统(ICS)里，找出和「{theme}」相关的具体对象"
+            f"（数据库表/Java类/接口/按钮/页面等）。\n"
+            f"知识图谱已检索到以下相关实体（均为真实存在）：\n{ctx}\n\n"
+            f"请把以上检索结果里的实体【尽量全部逐行列出】，每行一个，"
+            f"直接用检索结果里的名称，不要自行筛选/合并/省略，不要解释。检索到几个就列几个。")
+
+
+# ── 评分 ─────────────────────────────────────────────────────────────
+_LINE_CLEAN = re.compile(r"^[\-\*\d\.\)、\s]+")
+_CODE_TOK = re.compile(r"[A-Za-z_][\w./\-:]{2,}")
+
+
+def _clean_lines(ans):
+    """'每行一个' 答案 → 清洗后的短行列表（模型声称存在的对象，每行一个）。"""
+    out = []
+    for l in ans.splitlines():
+        l = _LINE_CLEAN.sub("", l).strip()
+        if l and len(l) <= 60:
+            out.append(l)
+    return out
+
+
+def score(ans, gold, UNI, ground_cn):
+    """grounding 精度：每行算'真实'当其含某真代码标识符(∈UNI) 或 命中本格 ground_cn(中文)。
+    ground_cn = 该格可作'真实'依据的中文名集合(bare=gold, scaffold=retrieved)，bare/scaffold 同口径公平。
+    """
+    lines = _clean_lines(ans)
+    real_n = 0
+    for l in lines:
+        if set(_CODE_TOK.findall(l)) & UNI:
+            real_n += 1
+        elif any(g and g in l for g in ground_cn):
+            real_n += 1
+    precision = (real_n / len(lines)) if lines else 0.0
+    hit = 0
+    for g in gold:
+        c = g.get("canonical") or ""
+        n = g.get("cn") or ""
+        if (c and c in ans) or (n and n in ans):
+            hit += 1
+    recall = (hit / len(gold)) if gold else 0.0
+    return {
+        "precision": round(precision, 3),
+        "recall": round(recall, 3),
+        "halluc": round(1 - precision, 3),
+        "n_lines": len(lines),
+        "n_real": real_n,
+        "gold_hit": hit,
+        "gold_n": len(gold),
+    }
+
+
+def aggregate(rows):
+    n = len(rows) or 1
+    return {
+        "P": round(sum(r["precision"] for r in rows) / n, 3),
+        "R": round(sum(r["recall"] for r in rows) / n, 3),
+        "H": round(sum(r["halluc"] for r in rows) / n, 3),
+        "tokens": round(sum(r["tokens"] for r in rows) / n, 0),
+        "n": len(rows),
+    }
+
+
+def run(themes, verbose=True):
+    def log(*a):
+        if verbose:
+            print(*a)
+
+    if not CHAT_KEY:
+        raise RuntimeError("缺 DASHSCOPE_API_KEY")
+    log("[ruler] 载入 KG 全集 + gold…")
+    UNI, GOLD = load_universe_and_gold(themes)
+    log(f"[ruler] 实体名称全集 {len(UNI)} 条")
+    for th in themes:
+        log(f"  theme {th}: gold={len(GOLD[th])}")
+
+    matrix = {}
+    t0 = time.time()
+    for tier, model in MODELS.items():
+        for mode in ("bare", "scaffold"):
+            rows = []
+            for th in themes:
+                gold = GOLD[th]
+                if mode == "bare":
+                    ground_cn = {g["canonical"] for g in gold} | {g["cn"] for g in gold if g["cn"]}
+                    p = prompt_bare(th)
+                else:
+                    ents = retrieve_entities(th)
+                    ground_cn = {e["canonical"] for e in ents} | {e["cn"] for e in ents if e["cn"]}
+                    p = prompt_scaffold(th, ctx_from_entities(ents))
+                ans, tok = chat(model, p)
+                sc = score(ans, gold, UNI, ground_cn)
+                sc["tokens"] = tok
+                sc["theme"] = th
+                rows.append(sc)
+                log(f"  [{tier}/{mode}] {th}: P={sc['precision']:.2f} R={sc['recall']:.2f} "
+                    f"H={sc['halluc']:.2f} tok={tok} lines={sc['n_lines']}")
+            matrix[f"{tier}/{mode}"] = aggregate(rows)
+            log(f"[{tier}/{mode}] agg P={matrix[f'{tier}/{mode}']['P']:.3f} "
+                f"R={matrix[f'{tier}/{mode}']['R']:.3f} H={matrix[f'{tier}/{mode}']['H']:.3f} "
+                f"tok={matrix[f'{tier}/{mode}']['tokens']:.0f}")
+    log(f"[ruler] 总耗时 {int(time.time()-t0)}s")
+    return {"matrix": matrix, "themes": themes, "models": MODELS}
+
+
+def format_table(res):
+    m = res["matrix"]
+    lines = ["## 低模型梯度矩阵（grounding 精度P / 召回R / 幻觉H / token）",
+             "", "| 模型档 | 裸 P | 裸 R | 裸 H | 裸tok | +grounding P | +grounding R | +grounding H | +ground tok |",
+             "|---|---|---|---|---|---|---|---|---|"]
+    for tier in MODELS:
+        b = m[f"{tier}/bare"]
+        f = m[f"{tier}/scaffold"]
+        lines.append(f"| {tier}({MODELS[tier]}) | {b['P']:.2f} | {b['R']:.2f} | {b['H']:.2f} | {b['tokens']:.0f} | "
+                     f"{f['P']:.2f} | {f['R']:.2f} | {f['H']:.2f} | {f['tokens']:.0f} |")
+    # 卖点提炼
+    lines.append("")
+    weak_gain_p = m["weak/scaffold"]["P"] - m["weak/bare"]["P"]
+    weak_catchup = m["weak/scaffold"]["P"] - m["strong/bare"]["P"]
+    lines.append(f"- 弱模型 +grounding 精度提升: **{weak_gain_p*100:+.1f}pp**（反幻觉增益）")
+    lines.append(f"- 弱模型+grounding vs 强模型裸 精度差: **{weak_catchup*100:+.1f}pp**（>0=追上/反超=卖点成立）")
+    return "\n".join(lines)
+
+
+def main():
+    import argparse
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--themes", default=",".join(DEFAULT_THEMES))
+    ap.add_argument("--json", action="store_true")
+    args = ap.parse_args()
+    themes = [t.strip() for t in args.themes.split(",") if t.strip()]
+    res = run(themes)
+    if args.json:
+        print(json.dumps(res, ensure_ascii=False, indent=2))
+    else:
+        print("\n" + format_table(res))
+
+
+if __name__ == "__main__":
+    main()

package/templates/qoder/scripts/validation/eval/gradient_matrix_baseline_2026-07-21.json

@@ -1,33 +1,33 @@
-{
-  "task": "T1.0 低模型梯度评测基线",
-  "date": "2026-07-21",
-  "ruler": "scripts/validation/eval/gradient_matrix.py",
-  "env": "云 10.89.7.120 容器 wlinkj-workspace-backend-1, DashScope compatible-mode, KG entities=81255 edges=96965",
-  "models": {"weak": "qwen-turbo", "mid": "qwen-plus", "strong": "qwen-max"},
-  "themes": ["考勤", "排班", "加班", "车辆", "报警", "保险", "能耗", "预算", "导出", "审批"],
-  "wall_seconds": 138,
-  "chat_calls": 60,
-  "matrix": {
-    "weak/bare":     {"P": 0.02,  "R": 0.00, "H": 0.98, "tokens": 210},
-    "weak/scaffold": {"P": 0.961, "R": 0.815, "H": 0.039, "tokens": 929},
-    "mid/bare":      {"P": 0.00,  "R": 0.00, "H": 1.00, "tokens": 134},
-    "mid/scaffold":  {"P": 0.986, "R": 0.823, "H": 0.014, "tokens": 859},
-    "strong/bare":   {"P": 0.032, "R": 0.00, "H": 0.968, "tokens": 136},
-    "strong/scaffold":{"P": 1.000,"R": 0.796, "H": 0.000, "tokens": 844}
-  },
-  "selling_point_proof": {
-    "all_bare_hallucinate": "三档裸跑 P 0-0.032 / R 0 / H 0.97-1.0 — 连 qwen-max 无 grounding 召回 0",
-    "weak_plus_workflow_vs_strong_bare_pp": 92.9,
-    "tier_converges_with_grounding": "+grounding 后三档 P 0.961/0.986/1.000, R 0.815/0.823/0.796 — 模型档位几乎不影响质量"
-  },
-  "flywheel_round1_gap": {
-    "shortboard": "召回天花板 R≈0.80 (ILIKE top-30 未覆盖全 gold + 模型偶过滤)",
-    "lever": "检索升级: 真实 rag_search 语义 + 1跳图扩展 (映射 T1.4)",
-    "target_R": 0.90
-  },
-  "honest_caveats": [
-    "token 维度本尺子不适用: +grounding 必然加 token(喂上下文换 P 0->0.96); §9.1 的 token<=裸×0.8 指命令层去重/缓存(T1.2/T1.3/T1.7 真命令transcript量)",
-    "🖥️ 宿主真梯度(QoderWork Haiku/Sonnet/Opus × 真命令) 待用户跑 (SPEC §12 V1.1); 本代理用 DashScope turbo/plus/max 同构证卖点方向"
-  ],
-  "status": "尺子就绪 + API 代理基线完成; 🖥️ 宿主梯度列入 §12"
-}
+{
+  "task": "T1.0 低模型梯度评测基线",
+  "date": "2026-07-21",
+  "ruler": "scripts/validation/eval/gradient_matrix.py",
+  "env": "云 10.89.7.120 容器 wlinkj-workspace-backend-1, DashScope compatible-mode, KG entities=81255 edges=96965",
+  "models": {"weak": "qwen-turbo", "mid": "qwen-plus", "strong": "qwen-max"},
+  "themes": ["考勤", "排班", "加班", "车辆", "报警", "保险", "能耗", "预算", "导出", "审批"],
+  "wall_seconds": 138,
+  "chat_calls": 60,
+  "matrix": {
+    "weak/bare":     {"P": 0.02,  "R": 0.00, "H": 0.98, "tokens": 210},
+    "weak/scaffold": {"P": 0.961, "R": 0.815, "H": 0.039, "tokens": 929},
+    "mid/bare":      {"P": 0.00,  "R": 0.00, "H": 1.00, "tokens": 134},
+    "mid/scaffold":  {"P": 0.986, "R": 0.823, "H": 0.014, "tokens": 859},
+    "strong/bare":   {"P": 0.032, "R": 0.00, "H": 0.968, "tokens": 136},
+    "strong/scaffold":{"P": 1.000,"R": 0.796, "H": 0.000, "tokens": 844}
+  },
+  "selling_point_proof": {
+    "all_bare_hallucinate": "三档裸跑 P 0-0.032 / R 0 / H 0.97-1.0 — 连 qwen-max 无 grounding 召回 0",
+    "weak_plus_workflow_vs_strong_bare_pp": 92.9,
+    "tier_converges_with_grounding": "+grounding 后三档 P 0.961/0.986/1.000, R 0.815/0.823/0.796 — 模型档位几乎不影响质量"
+  },
+  "flywheel_round1_gap": {
+    "shortboard": "召回天花板 R≈0.80 (ILIKE top-30 未覆盖全 gold + 模型偶过滤)",
+    "lever": "检索升级: 真实 rag_search 语义 + 1跳图扩展 (映射 T1.4)",
+    "target_R": 0.90
+  },
+  "honest_caveats": [
+    "token 维度本尺子不适用: +grounding 必然加 token(喂上下文换 P 0->0.96); §9.1 的 token<=裸×0.8 指命令层去重/缓存(T1.2/T1.3/T1.7 真命令transcript量)",
+    "🖥️ 宿主真梯度(QoderWork Haiku/Sonnet/Opus × 真命令) 待用户跑 (SPEC §12 V1.1); 本代理用 DashScope turbo/plus/max 同构证卖点方向"
+  ],
+  "status": "尺子就绪 + API 代理基线完成; 🖥️ 宿主梯度列入 §12"
+}