关于「泄题如何伪造掌握:猜对率、假证书与拟合的盲区」,《零泄题苏格拉底考官》给出的定量结论是什么?

结论:程序 p41 说明,泄题让拿到答案的人以百分之百的概率通过考官,百分之二十的泄题率就足以让四张证书里有一张是假的。程序 p42 说明,事后拟合看不出泄题:它把泄题解释成生源更好、失手更少,曲线依旧吻合。检测因此要靠平行变体、延迟迁移题与固定猜对率这些出题时就设计好的结构,而它们都只是预防之后的补救。考官的本职,是在不泄题的前提下尽快、尽准地判断掌握,那么在一串对错比特之间,下一道题该问什么才最有价值?下一章搭建追问树。

数据口径: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,不含任何密钥或服务器地址,也不收录具体的套话提示词;本书是非个人化经营建议。 第一、二章量化叮嘱为何无效,在每题 3 次尝试、40 道题的示意课程里,p 等于 0.001 时一个 30 人的小班至少漏掉一道题的概率是 0.9727;再画出四种考官架构的信息流图,计算判分预言机的比特上限与尝试上限的代价。

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

P(certified): knower 0.9258, guesser 0.0118, key holder 1.0000
fcertifiedfalse sharefalse per 1000 students
0.000.37740.01887.1
0.050.40710.090336.7
0.100.43670.152066.4
0.200.49600.2534125.7
0.400.61460.3974244.3
查看完整验算程序
# Chapter 4. p41: a leaked key turns into false mastery certificates.
# Examination mode: BKT with L0 = 0.30, T = 0, G = 0.20, S = 0.10 and
# a mastery line of 0.95, at most 6 items per session. Three kinds of
# student (illustrative cohort): knowers (40 per cent, correct with
# probability 0.90), non-knowers without the key (correct 0.20) and
# non-knowers holding a leaked key (correct 1.00). A share f of the
# cohort holds the key. We enumerate all 2^6 answer strings exactly
# and report the false share among those certified as mastered.
import itertools
import math

L0, G, S, LINE, N = 0.30, 0.20, 0.10, 0.95, 6
KNOW = 0.40


def certified(seq):
    p = L0
    for r in seq:
        if r:
            p = p * (1 - S) / (p * (1 - S) + (1 - p) * G)
        else:
            p = p * S / (p * S + (1 - p) * (1 - G))
        if p >= LINE:
            return True
    return False


def p_cert(q):
    terms = []
    for seq in itertools.product((0, 1), repeat=N):
        if certified(seq):
            k = len([r for r in seq if r])
            terms.append(q ** k * (1 - q) ** (N - k))
    return math.fsum(terms)


c_know, c_guess, c_key = p_cert(1 - S), p_cert(G), p_cert(1.0)
print("P(certified): knower %.4f, guesser %.4f, key holder %.4f"
      % (c_know, c_guess, c_key))
print("f       certified  false share  false per 1000 students")
for f in (0.0, 0.05, 0.10, 0.20, 0.40):
    true_c = KNOW * c_know
    # key holders who already know are counted as knowers
    false_c = (1 - KNOW) * ((1 - f) * c_guess + f * c_key)
    total = true_c + false_c
    print("%-7.2f %-10.4f %-12.4f %.1f"
          % (f, total, false_c / total, 1000 * false_c))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《零泄题苏格拉底闯关考官、服务端答案物理隔离与掌握度贝叶斯追问树母典》第 4 章「泄题如何伪造掌握:猜对率、假证书与拟合的盲区」(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-262 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026262,
  author    = {Fan, Yafu},
  title     = {A Zero-Leak Socratic Examiner: Server-Side Isolation of Answers and Bayesian Mastery Probing Trees},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论