关于「实证闭环:实战工具与检查器规格」,《双脑异构协同架构》给出的定量结论是什么?

结论:第八章给出双脑协同台账、蒸馏小工具「双脑误差正交增益与状态同步包校验器」、按四件套分组的检查器规格表与预注册检验:漏网率从 0.030 降到 0.015 的对照,功效 0.80 时每组需要 1,534 个任务。书中以标注框区分已发表的证据、可运行的 Python 计算(23 段程序)、原典与作者的假说。

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: 0.015, 0.030, 0.80, 1534

A. leak 0.030 (now) vs 0.015 (target); two-sided 0.05
   power 0.80: 1,534 tasks per arm
   power 0.90: 2,053 tasks per arm
   halving a 0.010 leak to 0.005 at power 0.80: 4,673 per arm
B. SPRT for H0 leak 0.015 vs H1 leak 0.030, alpha 0.05, beta 0.20
   accept H1 when log-LR >= 2.7726; accept H0 when <= -1.5581
   each leak adds +0.6931; each clean task adds -0.0153
   expected tasks: about 284 if H0 holds, 323 if H1
C. register before looking: P0, P1, alpha, power, n, and the
   rule that the protocol is not changed before n is reached
查看完整验算程序
# Chapter 8. p82: pre-registered sample sizes for the leak test.
# H0: the current set-up leaks at P0; H1: the heterogeneous set-up
# leaks at P1. Fixed-n two-proportion test (normal approximation)
# and Wald's sequential probability ratio test with its boundaries
# and approximate expected sample sizes. Rates are illustrative and
# must be replaced by the user's own baseline before registering.
import math
from statistics import NormalDist

ND = NormalDist()
P0, P1 = 0.030, 0.015


def n_two_prop(p0, p1, alpha=0.05, power=0.80):
    za = ND.inv_cdf(1.0 - alpha / 2.0)
    zb = ND.inv_cdf(power)
    pb = (p0 + p1) / 2.0
    top = (za * math.sqrt(2.0 * pb * (1.0 - pb))
           + zb * math.sqrt(p0 * (1.0 - p0) + p1 * (1.0 - p1))) ** 2
    return math.ceil(top / (p0 - p1) ** 2)


def sprt(p0, p1, alpha=0.05, beta=0.20):
    a = math.log((1.0 - beta) / alpha)
    b = math.log(beta / (1.0 - alpha))
    up = math.log(p1 / p0)
    dn = math.log((1.0 - p1) / (1.0 - p0))
    e0 = p0 * up + (1.0 - p0) * dn
    e1 = p1 * up + (1.0 - p1) * dn
    n0 = ((1.0 - alpha) * b + alpha * a) / e0
    n1 = (beta * b + (1.0 - beta) * a) / e1
    return a, b, up, dn, n0, n1


if __name__ == "__main__":
    print("A. leak {:.3f} (now) vs {:.3f} (target); two-sided 0.05"
          .format(P0, P1))
    for pw in (0.80, 0.90):
        print("   power {:.2f}: {:,} tasks per arm".format(
            pw, n_two_prop(P0, P1, power=pw)))
    print("   halving a 0.010 leak to 0.005 at power 0.80: {:,} per arm"
          .format(n_two_prop(0.010, 0.005)))
    a, b, up, dn, n0, n1 = sprt(P1, P0)
    print("B. SPRT for H0 leak {:.3f} vs H1 leak {:.3f}, alpha 0.05, "
          "beta 0.20".format(P1, P0))
    print("   accept H1 when log-LR >= {:.4f}; accept H0 when <= {:.4f}"
          .format(a, b))
    print("   each leak adds {:+.4f}; each clean task adds {:+.4f}"
          .format(up, dn))
    print("   expected tasks: about {:,.0f} if H0 holds, {:,.0f} if H1"
          .format(n0, n1))
    print("C. register before looking: P0, P1, alpha, power, n, and the")
    print("   rule that the protocol is not changed before n is reached")

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 8 章「实证闭环:实战工具与检查器规格」(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