# -*- coding: utf-8 -*- """灵枢 · 卡校验器 card_validator（自生长闭环第一步 · v1.16） 对 LLM 写的候选卡做机械五查（结构闸门）： 1. 字段完整性：name/domain/level/response{trigger,action,counters}/ condition_space 四要素/content 非空；学科卡必带 edu_level 2. 条件空间四要素：observation_position/observation_tool/ time_window/existence_constraint 全声明 3. counters 冲突：新卡 counters 与现有卡 counters 交叉匹配 （二元组交集 ≥4 视为潜在冲突 → 提示设计者） 4. 学科域/教育层级归属：domain 可识别；edu_level ∈ E1-E5；学科卡必带层级 5. 测试用例可执行性（★ 物理信息基底闸门的前置）：tests 字段存在且 含可观测判据（observable/topic/expected 或 propositions+observable） 输出：{verdict: pass/needs_revision/fail, checks, pending_designer} pass → pending_designer（人工终裁，方向闸门） needs_revision → 打回 LLM 重写（带检查报告） fail → 拒绝（结构致命缺失） 设计边界：校验器只做机械检查（结构/格式/冲突信号），不裁决知识真理性—— 真理由物理信息基底裁决（测试用例实际执行），方向由设计者终裁。 """ import json import os import sys HERE = os.path.dirname(os.path.abspath(__file__)) # 教育层级 EDU_LEVELS = {"E1", "E2", "E3", "E4", "E5"} # 条件空间四要素 CS_FIELDS = ["observation_position", "observation_tool", "time_window", "existence_constraint"] # 学科卡必带字段（response 三件套） RESP_FIELDS = ["trigger", "action", "counters"] # 卡必带 state_attributes SA_REQUIRED = ["name", "domain", "level", "response"] def _bigram_set(text): """二元组集合（去非中文/字母数字）。""" import re t = re.sub(r"[^\u4e00-\u9fffA-Za-z0-9]", "", text or "") return {t[i:i + 2] for i in range(len(t) - 1)} def load_existing_counters(dex=None): """现有 138 卡 counters 文本集合（供冲突检测）。""" out = [] if dex is None: return out try: from aeis_core import MemoryLayer for n in dex.store.query_nodes(layer=MemoryLayer.KNOWLEDGE, limit=500): sa = n.state_attributes if not sa.get("name"): continue c = (sa.get("response") or {}).get("counters", "") or "" if c: out.append((sa.get("name"), c)) except Exception: pass return out def validate(card, dex=None, existing_counters=None): """五查机械校验。card: dict（LLM 写的候选卡）或 ConditionDex 节点 dict。 返回 {verdict, checks: [{check, passed, detail}], pending_designer}。""" checks = [] # ---------- 1. 字段完整性 ---------- missing = [] for f in SA_REQUIRED: if not card.get(f): missing.append(f) resp = card.get("response") or {} for f in RESP_FIELDS: if not resp.get(f): missing.append(f"response.{f}") if not (card.get("content") or card.get("claim")): missing.append("content/claim") # 学科卡（domain 不是协议层）必带 edu_level is_protocol = card.get("domain") in ( "智能论", "条件论", "存在论", "信息论", "负反馈系统", "认知双过程", "情感-信任连接", "自我增强回路", "自我意识", "决策论", "博弈论", "统计学", "概率论", "形式逻辑", "科学方法论", "控制论", "机制设计", "时空记忆图", "情感情绪仿真", "价值理论与AI对齐", "复杂系统与涌现", "自我实现预言", "智能推荐算法", "预测误差学习", "演化论与模因", "政治经济学", "热力学/耗散结构", "人类观察者") if not is_protocol and not card.get("edu_level"): missing.append("edu_level（学科卡必带教育层级）") checks.append({ "check": "1.字段完整性", "passed": not missing, "detail": f"缺失: {missing}" if missing else "全部就位", }) # ---------- 2. 条件空间四要素 ---------- cs = card.get("condition_space") or card.get("cs") or {} if isinstance(cs, str): try: cs = json.loads(cs) except Exception: cs = {} cs_missing = [f for f in CS_FIELDS if not cs.get(f)] checks.append({ "check": "2.条件空间四要素", "passed": not cs_missing, "detail": f"缺失: {cs_missing}" if cs_missing else "四要素齐备", }) # ---------- 3. counters 冲突（与现有卡） ---------- conflict = [] new_counters = resp.get("counters", "") or "" if existing_counters is None: existing_counters = load_existing_counters(dex) new_cb = _bigram_set(new_counters) for name, c in existing_counters: if name == card.get("name"): continue inter = len(new_cb & _bigram_set(c)) if inter >= 4: conflict.append(f"{name}(交集{inter})") checks.append({ "check": "3.counters冲突", "passed": not conflict, "detail": f"潜在冲突: {conflict[:3]}" if conflict else "无冲突信号", }) # ---------- 4. 学科域/教育层级归属 ---------- dom = card.get("domain", "") edu = card.get("edu_level") dom_issues = [] if not dom: dom_issues.append("domain 为空") if is_protocol and edu is not None and edu not in EDU_LEVELS: dom_issues.append(f"协议层卡不应带 edu_level={edu}") if not is_protocol: if edu not in EDU_LEVELS: dom_issues.append(f"edu_level={edu} 不在 E1-E5") checks.append({ "check": "4.学科域/层级归属", "passed": not dom_issues, "detail": "; ".join(dom_issues) if dom_issues else f"domain={dom} edu={edu or '通用'}", }) # ---------- 5. 测试用例可执行性（物理基底闸门 · v1.16 升级为真执行） ---------- # 格式检查（可观测判据存在）是前置；tests.executable（lang/code）存在时 # **真跑**——代码运行/编译结果 = 物理信息基底裁决（爸爸外部参照）。 tests = card.get("tests") or card.get("test_cases") tests_ok = True tests_detail = "无 tests 字段" if tests: if isinstance(tests, dict): t_keys = [k for k in ("topic", "expected", "observable", "propositions") if tests.get(k)] tests_ok = bool(t_keys) tests_detail = f"tests 含: {t_keys}" if t_keys else "tests 字段缺可观测判据" # 真执行：tests.executable 存在 → 物理基底校准（代码类卡） ex = tests.get("executable") if ex and isinstance(ex, dict) and ex.get("lang") and ex.get("code"): try: from code_test_runner import CodeTestRunner run_r = CodeTestRunner().run_card({"name": card.get("name"), "tests": tests}) status = run_r.get("status", "error") if status == "pass": tests_ok = True tests_detail = (f"物理基底真跑 PASS: {run_r.get('detail','')}" f"{' | out: '+run_r['output'][:40] if run_r.get('output') else ''}") elif status == "env_missing": tests_ok = True # 环境缺失不判卡错（D-005 诚实降级）， # 但标记待环境——格式仍通过，设计者终裁时可见 tests_detail = f"物理基底：环境缺失（{run_r.get('detail','')[:50]}）——待环境" else: tests_ok = False tests_detail = f"物理基底真跑 FAIL: {run_r.get('detail','')[:120]}" except Exception as e: tests_ok = False tests_detail = f"物理基底执行异常: {str(e)[:100]}" elif isinstance(tests, list) and tests: tests_ok = any(isinstance(t, dict) and ("observable" in t or "expected" in t) for t in tests) tests_detail = "tests 列表含可观测判据" if tests_ok else "tests 列表缺可观测判据" else: tests_ok = False checks.append({ "check": "5.测试用例可执行性", "passed": tests_ok, "detail": tests_detail, }) # ---------- 汇总 ---------- passed_all = all(c["passed"] for c in checks) fail_count = sum(1 for c in checks if not c["passed"]) if fail_count == 0: verdict = "pass" elif fail_count <= 2 and not any(c["check"].startswith("1.") for c in checks if not c["passed"]): verdict = "needs_revision" else: verdict = "fail" return { "verdict": verdict, "checks": checks, "pending_designer": verdict == "pass", "summary": f"{verdict}: {sum(1 for c in checks if c['passed'])}/5 通过", } # ---------------- 自测 ---------------- if __name__ == "__main__": sys.stdout.reconfigure(encoding="utf-8") GOOD = { "name": "测试卡", "domain": "物理学", "level": 2, "edu_level": "E2", "response": {"trigger": "涉及测试议题", "action": "以测试知识回应", "counters": "越级使用未学概念"}, "condition_space": {"observation_position": "测试位", "observation_tool": "测试工具", "time_window": [0.0, 9999999999.0], "existence_constraint": "测试约束"}, "content": "测试内容", "tests": {"topic": "测试主题", "observable": "可观测判据"}, } BAD = {"name": "残缺卡", "domain": "未知领域"} for label, c in [("GOOD", GOOD), ("BAD", BAD)]: r = validate(c) print(f"[{label}] {r['summary']}") for ch in r["checks"]: mark = "✓" if ch["passed"] else "✗" print(f" {mark} {ch['check']}: {ch['detail']}") print(f" pending_designer={r['pending_designer']}")