# -*- coding: utf-8 -*- """md_cg · 第 12 篇：独立元认知（观察自身认知的二阶单元） 理论出处（本仓原文）： · 情绪 = 信息差二阶变化 d²D/dt²（`docs/theory/智能的公理化基石.md` §十一，:412-510） · 情感 = 信任二阶变化 d²T/dt²（同文档 :469-508） · 五大单元：元认知是反思/验证通道的**外部观察者**（同文档 :540-556） · 推论三「局部不可知」：盲区即知识（同文档 :277-288） · 信任 P_trust / P_gap（同文档 :393-408） · 导航税：显式状态记录（`docs/theory/智能的认知过程.md`:84-86） 覆盖： A 轨迹：D(t) / dD/dt / d²D/dt² / 情绪 / 四态分布 B 校准：期望 vs 实际 / gap / 分箱 / ECE / 确定性建议 C 盲区：BLINDSPOT 邻域聚合 + unresolved 清单 D 信任：P_gap / P_trust / 情感（d²T/dt²） E 报告：四观测面 + 建议带证据 + 独立留痕 F 闸门：相似历史 → 建议分流 / 无历史 → no_prior G 独立性：不改 frontmatter / 不参与裁决 H 留痕 / 自描述 / health 面 运行：python -m md_cg.test_p12_metacognition """ from __future__ import annotations import json import tempfile from .mdcos import MdCGOS from .mdcg import STATE_ACCEPT, STATE_REJECT, STATE_BLINDSPOT from . import metacognition as mc PASS = FAIL = 0 FAILS = [] def check(name, cond, detail=""): global PASS, FAIL if cond: PASS += 1 print(f" [PASS] {name}" + (f" · {detail}" if detail else "")) else: FAIL += 1 FAILS.append(name) print(f" [FAIL] {name} · {detail}") def _r(state): """构造一次检索结果（精确控制信息差 D：ACCEPT→0，其它→1）。""" return [({"id": "x", "frontmatter": {}, "content": "", "path": "x"}, 1.0, {"state": state})] def main(): root = tempfile.mkdtemp(prefix="mdcg_p12_") cg = MdCGOS(root) # ---------- A. 轨迹 ---------- print("\n[A] 信息差轨迹（D / dD / d²D → 情绪，§十一）") tr0 = cg.metacognition_trace() check("A1 无留痕 → insufficient_data（不编造数值）", tr0["ok"] is False and tr0["reason"] == "insufficient_data", str(tr0)) cg.reflect("缓存失效策略", _r(STATE_ACCEPT)) cg.reflect("缓存失效策略", _r(STATE_ACCEPT)) tr = cg.metacognition_trace() check("A2 命中 → D 收敛到 0", tr["d_current"] == 0.0, str(tr["d_current"])) check("A3 四态分布可读", "ACCEPT" in tr["states"], str(tr["states"])) cg.reflect("缓存失效策略", _r(STATE_REJECT)) cg.reflect("缓存失效策略", _r(STATE_ACCEPT)) tr = cg.metacognition_trace() check("A4 二阶差分识别收敛 → approaching", tr["emotion"] == "approaching", f'emotion={tr["emotion"]} d2={tr["d2"]}') check("A5 一阶方向 dD/dt 可算", isinstance(tr["d1"], float), str(tr["d1"])) # ---------- B. 校准 ---------- print("\n[B] 自信校准（期望正确率 vs 实际验证通过率）") cal0 = cg.metacognition_calibration() check("B1 无证据 → insufficient_data", cal0["ok"] is False, str(cal0)) for i in range(6): cg.add(f"v{i}", f"# 功能：验证样本 {i}\n", layer="knowledge", importance=0.5) cg.verify(f"v{i}", "实测通过", "confirmed") for i in range(3): cg.verify(f"v{i}", "反例", "weakened") cal = cg.metacognition_calibration() check("B2 期望/实际/gap 自洽", cal["ok"] and abs(cal["gap"] - (cal["expected_accuracy"] - cal["actual_accuracy"])) < 1e-6, str({k: cal[k] for k in ("expected_accuracy", "actual_accuracy", "gap")})) check("B3 ECE 非负", cal["ece"] >= 0, str(cal["ece"])) check("B4 分箱覆盖已观测区间", len(cal["bins"]) > 0, str(cal["bins"])) check("B5 判定 ∈ 四类", cal["verdict"] in ("calibrated", "overconfident", "underconfident", "insufficient_data"), cal["verdict"]) adv = mc._advise( {"ok": True, "d2": 0.0, "emotion": "stable", "state_rates": {}}, {"verdict": "overconfident", "expected_accuracy": 0.8, "actual_accuracy": 0.5, "gap": 0.3}, {"unresolved_count": 0}, {}) check("B6 过度自信 → 建议抽查复核", any(a["code"] == "overconfident" for a in adv), str(adv)) adv2 = mc._advise({"ok": False}, {"verdict": "calibrated"}, {"unresolved_count": 0}, {}) check("B7 无轨迹 → 提示先积累留痕", any(a["code"] == "no_trace" for a in adv2), str(adv2)) # ---------- C. 盲区 ---------- print("\n[C] 盲区地图（推论三：局部不可知）") cg.reflect("量子纠缠退相干时间", _r(STATE_BLINDSPOT)) cg.reflect("量子纠缠退相干时间", _r(STATE_BLINDSPOT)) cg.reflect("另一个模糊问题", _r(STATE_BLINDSPOT)) bs = cg.metacognition_blindspots() top = bs["items"][0] if bs["items"] else {} check("C1 盲区按邻域聚合", top.get("blindspot", 0) >= 2, str(top)) check("C2 返回 unresolved 计数", "unresolved_count" in bs, str(bs["unresolved_count"])) # ---------- D. 信任 ---------- print("\n[D] 信任（P_gap / P_trust / 情感 d²T/dt²）") tt = cg.metacognition_trust() check("D1 P_gap 可算", tt["p_gap"] is not None, str(tt["p_gap"])) check("D2 P_trust 来自验证留痕", tt["p_trust"] is not None, str(tt["p_trust"])) check("D3 情感取自二阶差分", tt["emotion"] in ("approaching", "avoiding", "stable"), str(tt["emotion"])) # ---------- E. 报告 ---------- print("\n[E] 元认知报告（四观测面 + 建议）") rep = cg.metacognition_report() check("E1 四观测面齐全", all(k in rep for k in ("trace", "calibration", "blindspots", "trust")), str(list(rep))) check("E2 建议带 why/action（可复核）", bool(rep["advice"]) and all("why" in a and "action" in a for a in rep["advice"]), str([a["code"] for a in rep["advice"]])) check("E3 独立性声明", rep["independent"] is True and "不" in rep["note"], rep["note"][:40]) # ---------- F. 闸门 ---------- print("\n[F] 元认知闸门 self_check") sc = cg.self_check("量子纠缠退相干时间") check("F1 相似历史命中 → 建议声明盲区", sc["recommendation"] == "declare_blindspot" and sc["prior_attempts"] >= 2, str({k: sc[k] for k in ("prior_attempts", "recommendation", "warning")})) sc2 = cg.self_check("完全无关的全新问题 xyz") check("F2 无相似历史 → no_prior", sc2["warning"] == "no_prior" and sc2["recommendation"] == "search_then_decide", str(sc2["recommendation"])) check("F3 空查询被拒", cg.self_check("")["ok"] is False) # ---------- G. 独立性 ---------- print("\n[G] 独立性（不参与裁决）") before = json.dumps(cg.get("v0")["frontmatter"], sort_keys=True, ensure_ascii=False) cg.metacognition_report() cg.self_check("量子纠缠退相干时间") after = json.dumps(cg.get("v0")["frontmatter"], sort_keys=True, ensure_ascii=False) check("G1 元认知不改节点 frontmatter", before == after) check("G2 独立性在自描述中声明", any(("不写" in c) or ("不参与" in c) for c in mc.catalog()["constraints"]), str(mc.catalog()["constraints"])) # ---------- H. 留痕 / 自描述 ---------- print("\n[H] 留痕 / 自描述 / health 面") h = cg.metacognition_history() check("H1 留痕写入 _metacognition.jsonl", h["n"] > 0, str(h["n"])) check("H2 留痕含 report 与 self_check", {"report", "self_check"} <= {r.get("op") for r in h["records"]}, str({r.get("op") for r in h["records"]})) cat = cg.metacognition_catalog() check("H3 自描述含四观测面 + 闸门", {"trace", "calibration", "blindspots", "trust", "self_check"} <= set(cat["faces"]), str(cat["faces"])) hh = cg.health_os() check("H4 health 报告 metacognition 面", "metacognition" in hh.get("os", {}), str(hh.get("os", {}).get("metacognition"))) # MCP 入口冒烟（注册 + 两条 dispatch 路径） from . import mcp_server as ms check("H5 MCP 注册 mdcg_metacognition", any(t["name"] == "mdcg_metacognition" for t in ms.TOOLS)) check("H6 cg op=metacognition 路由", ms._cg_call(cg, {"op": "metacognition", "action": "catalog"}).get("module") == "metacognition") check("H7 mdcg_metacognition dispatch", ms._metacognition_call(cg, {"action": "trust"})["ok"] is True) print(f"\n==== P12 结果：{PASS} 通过 / {FAIL} 失败 ====") if FAILS: print("失败项：" + "、".join(FAILS)) return FAIL == 0 if __name__ == "__main__": import sys # 用 reconfigure 而非「包一层 TextIOWrapper」：后者在 stdout 被重定向到文件时 # 会在解释器退出阶段丢缓冲，CI 里会看不到失败原因。 # 本用例原先缺这一步：Windows 默认 gbk 控制台下打印「d²D/dt²」的 ²(U+00B2) # 直接 UnicodeEncodeError，崩在**第一条断言之前**，输出只剩 725 字节—— # 于是「P12 通过」这件事从未被真正执行过，也从未被任何人看见。 if hasattr(sys.stdout, "reconfigure"): sys.stdout.reconfigure(encoding="utf-8", errors="replace") sys.exit(0 if main() else 1)