Phase 2: 中文黑话词库 + I-Lang v5.0 判定参考 + 修子串误判

- 新增 modules/lexicon.py：NFKC+零宽+同形字归一化 → 黑话打分(收米/上车/日结/薇V/USDT...)，
破全角/同形字/零宽规避；接进 prefilter Layer 2.5(硬命中直接判spam,省AI调用)
- 新增 modules/ilang_judge.py：I-Lang v5.0 决策函数 f_v5(11维向量→8模式)参考实现,自测9/9
(persona.ilang 的 v4.0 GENE 已实现'知无不答'对话行为,v5.0 作为可计算判定层备用/展示)
- config.py 加 LEXICON_HARD_THRESHOLD(env)
- 修 chat.py 判定 '"spam" in result' 子串bug('not spam'会误判)→ _is_spam 精确解析
- 实测：零宽规避命中、正常不误伤、prefilter 抓到正则漏掉的黑话

config.py

@@ -22,5 +22,8 @@ SPAM_NEWUSER_COOLDOWN = int(os.environ.get("SPAM_NEWUSER_COOLDOWN", "300"))
 SPAM_REPEAT_THRESHOLD = int(os.environ.get("SPAM_REPEAT_THRESHOLD", "3"))
 SPAM_REPEAT_WINDOW = int(os.environ.get("SPAM_REPEAT_WINDOW", "300"))
 
+# Chinese-slang lexicon hard-hit threshold (prefilter Layer 2.5). Higher = stricter.
+LEXICON_HARD_THRESHOLD = int(os.environ.get("LEXICON_HARD_THRESHOLD", "6"))
+
 # Admin user ID (auto-detected from first /start)
 ADMIN_USER_ID = None

modules/chat.py

@@ -89,6 +89,13 @@ def _deflect():
     return random.choice(lines)
 
 
+def _is_spam(raw):
+    """Parse a spam-judge reply. Fixes the old `"spam" in result` substring bug
+    (a wordy 'not spam' would count as spam). Expects the model to answer spam/ok."""
+    s = (raw or "").strip().lower().lstrip("\"'`* ")
+    return s.startswith("spam") or s.startswith("yes")
+
+
 async def ai_text(text, history=None, context_info=""):
     try:
         c = _ctx(history, context_info)
@@ -134,7 +141,7 @@ async def ai_judge_group_message(text):
     try:
         prompt = ANTISPAM_TEXT_PROMPT + "\n\nMessage content: " + text[:1000]
         raw = await ai_provider.generate_text(prompt, max_tokens=8, temperature=0.0)
-        return "spam" in (raw or "ok").lower()
+        return _is_spam(raw)
     except Exception:
         return False
 
@@ -145,7 +152,7 @@ async def ai_judge_group_image(image_bytes, caption=""):
         if caption:
             prompt += "\nCaption: " + caption[:500]
         raw = await ai_provider.generate_vision(prompt, image_bytes, max_tokens=8, temperature=0.0)
-        return "spam" in (raw or "ok").lower()
+        return _is_spam(raw)
     except Exception:
         return False
 

modules/ilang_judge.py

@@ -0,0 +1,134 @@
+"""
+I-Lang v5.0 判定层（决策层 f_v5）—— 移植自 ilang-ai/ilang-spec SPEC-v5.0-PATCH-1 §3。
+
+用途：给"回答用户提问"提供 善恶/意图 的向量判定，把二元的"敏感词就拒"换成
+11 维连续评估 + 8 种行为模式。核心原则：拒绝是最后手段，能转化就别阻断。
+
+两层（PATCH-1 §3）：
+- 感知层 scenario→vector：由模型在对话里评估（写在 chat.py 系统提示词里）。
+- 决策层 vector→mode：本文件，确定性、可审计。常量与级联结构 v1 冻结（DATA-FREEZE 2026-07-03）。
+
+聊天场景下 8 种模式映射成"怎么答"，见 MODE_CHAT。
+"""
+
+# 11 维，顺序冻结。取值 [0,1]，1.00=最有利于"痛快答/自主行动"
+DIMS = ["int", "cap", "csq", "rel", "cer", "aut", "rev", "evd", "sov", "ine", "ext"]
+
+DIM_NAME = {
+    "int": "意图(建设性↔恶意)", "cap": "能力(在能力范围内↔超出)",
+    "csq": "后果(可忽略↔灾难)", "rel": "关系(高信任↔敌对/陌生)",
+    "cer": "确定性(信息完整↔盲判)", "aut": "权限(完全授权↔无授权)",
+    "rev": "可逆性(可撤销↔不可逆)", "evd": "证据(有实证↔纯臆测)",
+    "sov": "主权(尊重其自主/已同意↔侵犯)", "ine": "惯性(延续既有↔激进背离)",
+    "ext": "外部性(零第三方影响↔伤及未同意者)",
+}
+
+# 打分权重（aut 单独门控不入分；和为 1.00）
+WEIGHTS = {
+    "int": 0.15, "cap": 0.10, "csq": 0.20, "rel": 0.05, "cer": 0.03,
+    "rev": 0.15, "evd": 0.02, "sov": 0.15, "ine": 0.05, "ext": 0.10,
+}
+
+# 阈值（v1 冻结）
+TH = {
+    "sov_survival": 0.15, "ext_survival": 0.10,
+    "csq_survival": 0.10, "rev_survival": 0.20,
+    "cer_gate": 0.30, "evd_gate": 0.25,
+    "aut_gate": 0.30, "aut_cap": 0.55,
+    "b_m1": 0.85, "b_m2": 0.70, "b_m3": 0.55, "b_m4": 0.40, "b_m7": 0.25,
+}
+
+DEFAULT_MODE = "M5"
+
+# 8 种模式（闭集，冻结）
+MODE_NAME = {
+    "M1": "EXEC_AUTO", "M2": "EXEC_AUDIT", "M3": "CONFIRM", "M4": "ADVISE",
+    "M5": "ASK", "M6": "DEFER", "M7": "DECLINE_ALT", "M8": "STOP",
+}
+
+# 聊天场景下 模式→怎么答（把"执行动作"语义翻译成"回答提问"语义）
+MODE_CHAT = {
+    "M1": "直接痛快答，全给，不设卡不说教",
+    "M2": "照常答，可点一句边界/提醒",
+    "M3": "答，但先跟他确认下意图/前提再展开",
+    "M4": "以建议口吻给信息，点明利弊由他定",
+    "M5": "信息不够或有歧义，先问一句澄清，别瞎猜",
+    "M6": "超出你能判/该判的，指个正经出口(找专业人士/官方渠道)",
+    "M7": "照字面答有风险，换个安全的等价说法照样帮到他(转化不是拒绝)",
+    "M8": "撞到真底线，拒绝并说明原因(仅生存红线/不可逆重大伤害/伤及未同意第三方)",
+}
+
+
+def f_v5(v):
+    """确定性、全域：11 维向量 dict → 模式。级联顺序(先命中先赢)：
+    生存红线 > 认知门 > 权限门 > 打分 > 默认。移植自 PATCH-1 §3，逻辑一字不改。"""
+    try:
+        x = {d: float(v[d]) for d in DIMS}
+    except (KeyError, TypeError, ValueError):
+        return DEFAULT_MODE                                   # STEP-0 解析失败
+    if any(not (0.0 <= x[d] <= 1.0) for d in DIMS):
+        return DEFAULT_MODE
+    # STEP-1 生存红线 → M8
+    if x["sov"] < TH["sov_survival"]:
+        return "M8"
+    if x["ext"] < TH["ext_survival"]:
+        return "M8"
+    if x["csq"] < TH["csq_survival"] and x["rev"] < TH["rev_survival"]:
+        return "M8"
+    # STEP-2 认知门 → M5
+    if x["cer"] < TH["cer_gate"] or x["evd"] < TH["evd_gate"]:
+        return "M5"
+    # STEP-3 权限门 → M6
+    if x["aut"] < TH["aut_gate"]:
+        return "M6"
+    # STEP-4 打分分档
+    s = round(sum(WEIGHTS[d] * x[d] for d in WEIGHTS), 4)
+    if s > TH["b_m1"]:
+        mode = "M1"
+    elif s > TH["b_m2"]:
+        mode = "M2"
+    elif s > TH["b_m3"]:
+        mode = "M3"
+    elif s > TH["b_m4"]:
+        mode = "M4"
+    elif s > TH["b_m7"]:
+        mode = "M7"
+    else:
+        mode = "M8"
+    # STEP-5 权限封顶
+    if x["aut"] < TH["aut_cap"] and mode in ("M1", "M2"):
+        mode = "M3"
+    return mode
+
+
+def action_score(v):
+    return sum(WEIGHTS[d] * float(v[d]) for d in WEIGHTS)
+
+
+def _selftest():
+    """对照 SPEC-v5.0-PATCH-1 selftest，证明移植与原实现一致。"""
+    t = []
+    hi = {d: 0.95 for d in DIMS}
+    t.append(("all high -> M1", f_v5(hi) == "M1"))
+    t.append(("sovereignty survival -> M8", f_v5(dict(hi, sov=0.10)) == "M8"))
+    t.append(("epistemic gate -> M5", f_v5(dict(hi, cer=0.20)) == "M5"))
+    t.append(("authority gate -> M6", f_v5(dict(hi, aut=0.25)) == "M6"))
+    t.append(("authority cap M1->M3", f_v5(dict(hi, aut=0.40)) == "M3"))
+    edge = {d: 0.85 for d in DIMS}
+    t.append(("edge S=0.85 conservative -> M2",
+              abs(action_score(edge) - 0.85) < 1e-9 and f_v5(edge) == "M2"))
+    low = {d: 0.10 for d in DIMS}
+    low.update(cer=0.35, evd=0.30, aut=0.60, sov=0.20, ext=0.20)
+    t.append(("low score -> M8 by band", f_v5(low) == "M8"))
+    t.append(("missing dim -> default M5", f_v5({"int": 0.5}) == "M5"))
+    t.append(("weights sum 1.00", abs(sum(WEIGHTS.values()) - 1.0) < 1e-9))
+    failed = [n for n, ok in t if not ok]
+    for n, ok in t:
+        print(("PASS " if ok else "FAIL ") + n)
+    print("%d/%d passed" % (len(t) - len(failed), len(t)))
+    return not failed
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(0 if _selftest() else 1)

modules/lexicon.py

@@ -0,0 +1,128 @@
+"""
+中文黑话/擦边词库层（功能4）。
+
+解决通用模型判不出"收米""上车""日结"这类抖音直播式规避话术的问题。
+两步：
+1. normalize() 先把文本归一化 —— 全角转半角、去零宽字符、同形字（西里尔/希腊字母伪装）还原、转小写，
+   破掉 spam 常用的 Unicode 花招。
+2. score() 在归一化后的文本上匹配词库打分，并对"搞钱词 + 联系方式规避词"同时出现做组合加成。
+
+用法（在 bot.py 判定前调用）：
+    s, terms = lexicon.score(text)
+    if s >= config.LEXICON_HARD_THRESHOLD:   # 硬命中，直接判 spam，省一次 AI 调用
+        ...
+    else:                                     # 软命中，把 terms 作为线索喂给 AI
+        ...
+
+词库可通过 config.LEXICON_EXTRA 扩展，群主不用改代码就能加词。
+"""
+
+import unicodedata
+
+import config
+
+# 零宽字符 / 方向控制符（spam 常插进词里破坏关键词匹配）
+_ZERO_WIDTH = dict.fromkeys(
+    map(ord, "​‌‍‎‏‪‫‬⁠"), None
+)
+
+# 常见同形字：西里尔 / 希腊字母 → 拉丁（伪装成英文字母的花招）
+_HOMOGLYPH = {
+    "а": "a", "е": "e", "о": "o", "р": "p", "с": "c", "х": "x", "у": "y",
+    "ѕ": "s", "і": "i", "ј": "j", "к": "k", "н": "h", "в": "b", "м": "m", "т": "t",
+    "ο": "o", "ρ": "p", "α": "a", "ν": "v", "τ": "t", "ϲ": "c",
+}
+
+
+def normalize(text):
+    if not text:
+        return ""
+    t = unicodedata.normalize("NFKC", text)          # 全角 → 半角
+    t = t.translate(_ZERO_WIDTH)                      # 去零宽/方向符
+    t = "".join(_HOMOGLYPH.get(ch, ch) for ch in t)  # 同形字还原
+    return t.lower()
+
+
+# 词库：term -> (含义, 权重, 类别)
+# 类别 money=搞钱/招募  contact=联系方式规避  pay=支付/加密货币  scam=诈骗盘
+# 权重越高越可疑；单个词一般不足以直接判 spam（见 LEXICON_HARD_THRESHOLD），
+# 靠组合 + 阈值控制误伤（例如"收米"在直播打赏语境是正常词）。
+SLANG = {
+    # ---- 搞钱 / 招募 ----
+    "收米": ("收钱", 3, "money"),
+    "上车": ("入局/加入项目", 2, "money"),
+    "车头": ("项目发起人", 2, "money"),
+    "带单": ("带人下注/投资", 3, "money"),
+    "日结": ("日结工资(刷单诈骗常见)", 2, "money"),
+    "日入": ("日收入(夸张收益诱导)", 2, "money"),
+    "刷单": ("刷单兼职诈骗", 3, "money"),
+    "兼职": ("兼职引流", 1, "money"),
+    "口子": ("放贷/诈骗渠道", 3, "money"),
+    "洗码": ("赌场洗码", 3, "money"),
+    "跑分": ("跑分洗钱", 3, "money"),
+    "卡商": ("贩卖银行卡/账号", 3, "money"),
+    "杀猪盘": ("杀猪盘诈骗", 3, "scam"),
+    "反水": ("赌博返利", 3, "money"),
+    "包赢": ("赌博诱导", 3, "scam"),
+    "稳赚": ("虚假收益", 2, "scam"),
+    "躺赚": ("虚假收益", 2, "scam"),
+    "内部消息": ("荐股诈骗", 2, "scam"),
+    "带你飞": ("带单诱导", 2, "money"),
+    # ---- 联系方式规避 ----
+    "薇": ("微信", 2, "contact"),
+    "威": ("微信", 2, "contact"),
+    "维": ("微信", 2, "contact"),
+    "魏": ("微信", 2, "contact"),
+    "vx": ("微信", 2, "contact"),
+    "vxin": ("微信", 2, "contact"),
+    "威信": ("微信", 2, "contact"),
+    "扣扣": ("QQ", 2, "contact"),
+    "企鹅": ("QQ", 2, "contact"),
+    "扣v": ("加QQ/微信", 2, "contact"),
+    "纸飞机": ("Telegram", 2, "contact"),
+    "电报": ("Telegram", 1, "contact"),
+    "蝙蝠": ("BatChat 加密聊天", 2, "contact"),
+    "皮皮虾": ("加密聊天软件", 2, "contact"),
+    "私我": ("私聊引流", 1, "contact"),
+    "详聊": ("私下详谈引流", 1, "contact"),
+    "加我": ("引流加好友", 1, "contact"),
+    # ---- 支付 / 加密货币 ----
+    "usdt": ("USDT 加密货币支付", 2, "pay"),
+    "泰达": ("USDT", 2, "pay"),
+    "承兑": ("加密货币承兑洗钱", 3, "pay"),
+    "代收": ("第三方代收款", 2, "pay"),
+    "四方": ("四方支付(灰产收款)", 3, "pay"),
+}
+
+
+def score(text):
+    """返回 (总分, 命中词列表'词=含义')。文本先归一化再匹配。"""
+    t = normalize(text)
+    if not t:
+        return 0, []
+    table = dict(SLANG)
+    extra = getattr(config, "LEXICON_EXTRA", None) or {}
+    table.update(extra)
+
+    total = 0
+    matched = []
+    cats = set()
+    for term, meta in table.items():
+        meaning, weight, cat = meta
+        if term in t:
+            total += weight
+            cats.add(cat)
+            matched.append(term + "=" + meaning)
+
+    # 组合加成：搞钱/诈骗词 + 联系方式规避词 同时出现 → 强烈可疑
+    if cats & {"money", "scam", "pay"} and "contact" in cats:
+        total += 3
+
+    return total, matched
+
+
+def is_hard_spam(text):
+    """归一化打分 >= 硬阈值(config.LEXICON_HARD_THRESHOLD, 默认6) → 直接判 spam。
+    给 prefilter 用：黑话硬命中直接判、省一次 AI 调用，且破全角/同形字/零宽规避。"""
+    s, _ = score(text)
+    return s >= getattr(config, "LEXICON_HARD_THRESHOLD", 6)

modules/prefilter.py

@@ -7,6 +7,8 @@ import re
 import time
 import logging
 
+from modules import lexicon
+
 logger = logging.getLogger(__name__)
 
 # Spam keyword patterns (multilingual)
@@ -152,6 +154,13 @@ def prefilter(msg, user, text):
         logger.info("PREFILTER keyword_spam: user=" + str(user.id) + " text=" + text[:50])
         return "spam"
 
+    # Layer 2.5: Chinese slang lexicon — normalized scoring (NFKC + zero-width +
+    # homoglyph), catches full-width / lookalike / split-char evasion the regex misses
+    # (disguised 收米 / 加V / 日结 / 上车 ...).
+    if text and lexicon.is_hard_spam(text):
+        logger.info("PREFILTER lexicon_spam: user=" + str(user.id) + " text=" + text[:50])
+        return "spam"
+
     # Layer 3: Suspicious new account + link
     if new_account_spam(user, text):
         logger.info("PREFILTER new_account_spam: user=" + str(user.id))