关于「三段流水线:把错误变成被拦下的错误」,《双脑异构协同架构》给出的定量结论是什么?

结论:第三章说明三段流水线只要有 0.1 的公共因子相关,漏网就是独立乘积的 1.87 倍,相关 0.5 时是 8.65 倍。

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: 0.1, 0.5, 1.87, 8.65

A. s 0.10, t 0.30, g 0.10; independent leak 0.0030
rho leak ratio to product per 10,000 tasks
0.00.00300 1.00 30.0
0.10.00560 1.87 56.0
0.20.00915 3.05 91.5
0.30.01369 4.56 136.9
0.50.02595 8.65 259.5
0.70.04327 14.42 432.7
0.90.06856 22.85 685.6
B. Monte Carlo check at rho 0.5, n200,000: 0.02571
C. halving one stage's miss rate at rho 0.5:
halve s: leak 0.01619
halve t: leak 0.01928
halve g: leak 0.01619
查看完整验算程序
# Chapter 3. p31: three-stage leak with a common factor.
# A bad outcome escapes only if the strategy layer errs (rate s), the
# translation check misses it (rate t) and the deterministic gate
# misses it (rate g). One-factor Gaussian model: each stage's latent
# z = sqrt(rho) F + sqrt(1 - rho) e. rho = 0 gives the plain product
# s * t * g. The integral over F is exact (Simpson); a Monte Carlo
# run with a fixed seed checks one row. Illustrative rates.
import math
import random
from statistics import NormalDist

ND = NormalDist()
S, T, G = 0.10, 0.30, 0.10


def all_fail(rates, rho, steps=2000):
    cuts = [ND.inv_cdf(p) for p in rates]
    if rho <= 0.0:
        return math.prod(rates)
    a, b = math.sqrt(rho), math.sqrt(1.0 - rho)
    h = 16.0 / steps
    terms = []
    for i in range(steps + 1):
        f = -8.0 + i * h
        w = 1 if i in (0, steps) else (4 if i % 2 else 2)
        v = ND.pdf(f)
        for c in cuts:
            v *= ND.cdf((c - a * f) / b)
        terms.append(w * v)
    return math.fsum(terms) * h / 3.0


def mc(rates, rho, n, rng):
    cuts = [ND.inv_cdf(p) for p in rates]
    a, b = math.sqrt(rho), math.sqrt(1.0 - rho)
    hit = 0
    for _ in range(n):
        f = rng.gauss(0.0, 1.0)
        if all(a * f + b * rng.gauss(0.0, 1.0) < c for c in cuts):
            hit += 1
    return hit / n


if __name__ == "__main__":
    base = S * T * G
    print("A. s {:.2f}, t {:.2f}, g {:.2f}; independent leak {:.4f}"
          .format(S, T, G, base))
    print("   rho   leak      ratio to product   per 10,000 tasks")
    for rho in (0.0, 0.1, 0.2, 0.3, 0.5, 0.7, 0.9):
        lk = all_fail([S, T, G], rho)
        print("   {:.1f}   {:.5f}   {:>6.2f}             {:>6.1f}"
              .format(rho, lk, lk / base, lk * 10000))
    rng = random.Random(41004)
    n = 200000
    print("B. Monte Carlo check at rho 0.5, n = {:,}: {:.5f}".format(
        n, mc([S, T, G], 0.5, n, rng)))
    print("C. halving one stage's miss rate at rho 0.5:")
    for name, r in (("s", [S / 2, T, G]), ("t", [S, T / 2, G]),
                    ("g", [S, T, G / 2])):
        print("   halve {}: leak {:.5f}".format(name, all_fail(r, 0.5)))

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 3 章「三段流水线:把错误变成被拦下的错误」(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