关于「鉴定师一致性:Cohen κ、高一致低 κ 的两个悖论与多人投票的天花板」,《二手精品鉴定与柠檬折价》给出的定量结论是什么?

结论:第三章讨论鉴定师一致性的两个悖论,并说明当五分之一的货是共同受骗的高仿时,九人投票的准确率只有 0.8757。

数据口径:书中参数为示意;全部计算由书中程序生成;本书不是对任何具体交易的定价意见,也不构成个人化的法律或理财建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(20 段程序)、原典与作者的假说。

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: 0.8757

easy-bag accuracy in case B: 0.9500
panel sizecase A independentcase B shared difficulty
10.85000.8500
30.93920.8793
50.97340.8804
70.98790.8782
90.99440.8757
case B with 99 appraisers: 0.8317
Fleiss kappa, 3 appraisers, 300 bags: case A 0.342, case B 0.518
查看完整验算程序
# Chapter 3. p32: a panel of appraisers voting by majority.  Each one
# is right with probability 0.85 on average.  Case A: errors are
# independent (Condorcet).  Case B: a fifth of the bags are hard
# copies that fool most appraisers (each right with 0.45) and the rest
# are easy (right with 0.95), the same average.  Exact binomial sums
# for the majority; Fleiss' kappa from 300 simulated bags (seed
# fixed).  All rates are illustrative.
import math
import random

rng = random.Random(20261112)
HARD, P_HARD = 0.20, 0.45
EASY = (0.85 - HARD * P_HARD) / (1 - HARD)


def majority(p, n):
    return math.fsum(math.comb(n, k) * p ** k * (1 - p) ** (n - k)
                     for k in range(n // 2 + 1, n + 1))


def fleiss(ratings, m):
    n = len(ratings)
    pj = [math.fsum(r[j] for r in ratings) / (n * m) for j in (0, 1)]
    pi = [math.fsum(c * (c - 1) for c in r) / (m * (m - 1))
          for r in ratings]
    pbar, pe = math.fsum(pi) / n, math.fsum(x * x for x in pj)
    return (pbar - pe) / (1 - pe)


def simulate(hard_share, p_hard, p_easy, m=3, n=300, prev=0.20):
    ratings = []
    for _ in range(n):
        fake = rng.random() < prev
        p = p_hard if rng.random() < hard_share else p_easy
        says_fake = 0
        for _ in range(m):
            right = rng.random() < p
            says_fake += int(right == fake)
        ratings.append((says_fake, m - says_fake))
    return fleiss(ratings, m)


print("easy-bag accuracy in case B: {:.4f}".format(EASY))
print("panel size   case A independent   case B shared difficulty")
for n in (1, 3, 5, 7, 9):
    b = HARD * majority(P_HARD, n) + (1 - HARD) * majority(EASY, n)
    print("    {}            {:.4f}               {:.4f}".format(
        n, majority(0.85, n), b))
print("case B with 99 appraisers: {:.4f}".format(
    HARD * majority(P_HARD, 99) + (1 - HARD) * majority(EASY, 99)))
print("Fleiss kappa, 3 appraisers, 300 bags: case A {:.3f}, case B {:.3f}"
      .format(simulate(0.0, 0.85, 0.85), simulate(HARD, P_HARD, EASY)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《二手精品多特征贝叶斯鉴定、阿克洛夫柠檬折价均衡与限量包款二级市场折旧半衰期》第 3 章「鉴定师一致性:Cohen κ、高一致低 κ 的两个悖论与多人投票的天花板」(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-171 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026171,
  author    = {Fan, Yafu},
  title     = {Multi-Feature Bayesian Authentication of Pre-Owned Luxury, Akerlof Lemon-Discount Equilibria and the Resale Half-Life of Limited Bags},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论