关于「反方与边界:两两覆盖设计、过于整齐的合成人格与平均买家谬误」,《红蓝军合成客群沙盘》给出的定量结论是什么?

结论:程序 p21 显示,覆盖全部 392 个两两组合,贪心覆盖数组只要 51 人,随机客群要两百多人。程序 p22 显示,生成器把标准差缩到 0.6,单维尾部买家就少了 16.3 倍,两维联合尾部少了 267.2 倍。程序 p23 显示,平均买家过线时,整群里仍可能有 0.4505 的买家低于门槛。人格覆盖的问题讲完了,下一个问题是:一位挑剔的买家坐下来砍价,双方各自能拿到多少?

数据口径:本书检验相反的立项命题:只用平均买家会漏掉尾部破绽,拉丁超立方合成客群在同样测试量下覆盖更多人格组合(McKay、Beckman 与 Conover 1979 年刊于《技术计量学》),多轮砍价有 Rubinstein 1982 年给出的贴现均衡。 书中参数为示意;全部计算由书中程序生成,不调用任何模型接口;本书是非个人化经营建议。

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: 0.6, 16.3, 267.2

real 95th percentile z = 1.645
sd stail sharebuyers for 95%joint tailbuyers
1.00.05000592.50e-031197
0.90.03380881.14e-032621
0.80.019891503.96e-047573
0.70.009393188.82e-0533961
0.60.003069789.36e-06320207
0.50.0005059732.51e-0711913332
tail share falls by a factor of 16.3 at s = 0.6
joint tail falls by a factor of 267.2 at s = 0.6
查看完整验算程序
# Chapter 2. p22: a persona generator that is too tidy loses the tail.
# Bisbee et al. (2024, Political Analysis) found that answers from
# language-model personas were less varied than real survey answers.
# Model (illustrative): a buyer trait is standard normal among real
# buyers; the generator draws it with the right mean but a standard
# deviation shrunk to s. A "tail buyer" is beyond the real 95th
# percentile (z > 1.645). We print the share of tail buyers the
# generator produces, buyers needed to meet one with 95 per cent
# chance, and the same for a joint tail on two independent traits.
import math
from statistics import NormalDist

Z = NormalDist().inv_cdf(0.95)


def need(p):
    return math.ceil(math.log(0.05) / math.log(1 - p))


print("real 95th percentile z = %.3f" % Z)
print("  sd s   tail share   buyers for 95%   joint tail   buyers")
for s in (1.0, 0.9, 0.8, 0.7, 0.6, 0.5):
    p = 1 - NormalDist(0, s).cdf(Z)
    q = p * p
    print("  %.1f    %.5f      %6d          %.2e   %8d"
          % (s, p, need(p), q, need(q)))
p1 = 1 - NormalDist(0, 1.0).cdf(Z)
p6 = 1 - NormalDist(0, 0.6).cdf(Z)
print("tail share falls by a factor of %.1f at s = 0.6" % (p1 / p6))
print("joint tail falls by a factor of %.1f at s = 0.6"
      % ((p1 / p6) ** 2))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《红蓝军合成客群对抗沙盘、参数化挑剔买家多轮砍价博弈与漏单破绽预检母典》第 2 章「反方与边界:两两覆盖设计、过于整齐的合成人格与平均买家谬误」(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-259 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026259,
  author    = {Fan, Yafu},
  title     = {Red-Blue Synthetic-Customer Sandboxes: Multi-Round Bargaining against Parameterised Demanding Buyers and Pre-Launch Leak Detection},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论