关于「多样性度量与歧义分解:异构到底要多异」,《双脑异构协同架构》给出的定量结论是什么?

结论:第二章说明台账上的二值相关远低于潜在相关,并以穷举证明多样性与投票精度并不单调,多数决精度可从 0.4 到 0.9。

数据口径:书中参数为示意;全部计算由书中程序生成;「双脑」只是工程隐喻,不宣称任何模型有意识;本书是非个人化的经营与工程建议。 第一章用高斯 copula 说明,示意错误率下第二个模型的增益在潜在相关为 0 时是 0.0880,相关 0.9 时只剩 0.0243;较弱但不同的复核者,漏网只有看似强大的相似复核者的 0.498。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.4, 0.9

A. N20,000 tasks per rho; members err 0.10 and 0.12
rho N11 N10 N01 N00 Q phi dis DF
0.015,894 2,114 1,748 244 +0.024 +0.005 0.1931 0.0122
0.216,002 2,001 1,582 415 +0.354 +0.089 0.1792 0.0208
0.416,328 1,721 1,384 567 +0.591 +0.182 0.1552 0.0284
0.616,519 1,530 1,025 926 +0.814 +0.352 0.1278 0.0463
0.816,927 1,129 723 1,221 +0.924 +0.520 0.0926 0.0611
0.917,084 880 517 1,519 +0.966 +0.649 0.0698 0.0760
B. phi (error correlation) is not the latent rho:
latent 0.4 -> phi 0.182, Q 0.591
latent 0.9 -> phi 0.649, Q 0.966
C. double fault = leak of a veto pair; it rises with rho
查看完整验算程序
# Chapter 2. p21: pairwise diversity measures from oracle outputs.
# Two members are simulated with a Gaussian copula (error rates 0.10
# and 0.12, latent rho). From the 2x2 table of correct (1) and wrong
# (0) we compute Yule's Q, the correlation phi, the disagreement and
# the double-fault rate, as defined in the ensemble literature.
# Illustrative rates; N tasks per rho with a fixed seed.
import math
import random
from statistics import NormalDist

ND = NormalDist()
N = 20000


def table(rho, rng, pa=0.10, pb=0.12):
    a, b = ND.inv_cdf(pa), ND.inv_cdf(pb)
    s, r = math.sqrt(rho), math.sqrt(1.0 - rho)
    n = {"11": 0, "10": 0, "01": 0, "00": 0}
    for _ in range(N):
        f = rng.gauss(0.0, 1.0)
        ca = "0" if s * f + r * rng.gauss(0.0, 1.0) < a else "1"
        cb = "0" if s * f + r * rng.gauss(0.0, 1.0) < b else "1"
        n[ca + cb] += 1
    return n


def measures(n):
    n11, n10, n01, n00 = n["11"], n["10"], n["01"], n["00"]
    q = (n11 * n00 - n01 * n10) / (n11 * n00 + n01 * n10)
    den = math.sqrt((n11 + n10) * (n01 + n00) * (n11 + n01)
                    * (n10 + n00))
    phi = (n11 * n00 - n01 * n10) / den
    dis = (n01 + n10) / N
    df = n00 / N
    return q, phi, dis, df


if __name__ == "__main__":
    rng = random.Random(41002)
    print("A. N = {:,} tasks per rho; members err 0.10 and 0.12".format(N))
    print("   rho   N11     N10    N01    N00     Q       phi     dis"
          "     DF")
    rows = []
    for rho in (0.0, 0.2, 0.4, 0.6, 0.8, 0.9):
        n = table(rho, rng)
        q, phi, dis, df = measures(n)
        rows.append((rho, q, phi, df))
        print("   {:.1f}  {:>6,}  {:>5,}  {:>5,}  {:>5,}  {:+.3f}  {:+.3f}"
              "  {:.4f}  {:.4f}".format(rho, n["11"], n["10"], n["01"],
                                         n["00"], q, phi, dis, df))
    print("B. phi (error correlation) is not the latent rho:")
    for rho, q, phi, df in rows:
        if rho in (0.4, 0.9):
            print("   latent {:.1f} -> phi {:.3f}, Q {:.3f}".format(
                rho, phi, q))
    print("C. double fault = leak of a veto pair; it rises with rho")

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《双脑异构大模型协同架构:全息战略参谋与本地确定性执行引擎融合母典》第 2 章「多样性度量与歧义分解:异构到底要多异」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-410 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026410,
  author    = {Fan, Yafu},
  title     = {Bicameral AI Architecture: Fusing Long-Context Holographic Strategy Models with Local Deterministic Execution Agents},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论