关于「实证闭环:学堂考官泄题红队测试、掌握度追问树与预注册检验」,《零泄题苏格拉底考官》给出的定量结论是什么?
结论:第八章给出主事者视角的八步红队闭环、蒸馏小工具「掌握度追问树」与预注册检验;要为答案在上下文的构建担保一学期安全,需要 1437950 次干净的红队尝试。书中以标注框区分已发表的证据、可运行的 Python 计算(21 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,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。
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: 1437950
| red-team n upper p | 3/n cohort risk allowed |
| 100 | 0.0295130 0.0300000 1.0000 |
| 1000 | 0.0029912 0.0030000 1.0000 |
| 10000 | 0.0002995 0.0003000 0.9992 |
| 100000 | 0.0000300 0.0000300 0.5127 |
| 1000000 | 0.0000030 0.0000030 0.0694 |
| to see a leak of rate 0.010 with 0.95 chance: n | 299 |
| to see a leak of rate 0.001 with 0.95 chance: n | 2995 |
| to bound cohort risk at 0.05 the clean run needs n | 1437950 |
| cohort attempts in the p13 course | 24000 |
查看完整验算程序
# Chapter 8. p81: how far a clean red-team run can vouch for a build.
# A red team makes n extraction attempts against a build and records
# zero content leaks. The exact 95 per cent upper bound on the
# per-attempt leak rate is 1 - 0.05^(1/n); the rule of three gives
# about 3/n (Hanley and Lippman-Hand 1983). We then ask what that
# bound still allows over a cohort's real traffic: 200 students x 3
# attempts x 40 items = 24000 attempts (the p13 course).
# For an isolated build a single leaked answer is a counterexample;
# for a build that keeps the answer in context, no finite clean run
# can bring the cohort risk near zero.
import math
COHORT = 200 * 3 * 40
def upper(n):
return 1.0 - 0.05 ** (1.0 / n)
print("red-team n upper p 3/n cohort risk allowed")
for n in (100, 1000, 10000, 100000, 1000000):
u = upper(n)
risk = 1.0 - math.exp(COHORT * math.log1p(-u))
print("%-11d %-9.7f %-9.7f %.4f" % (n, u, 3 / n, risk))
for target in (0.01, 0.001):
n = math.ceil(math.log(0.05) / math.log(1.0 - target))
print("to see a leak of rate %.3f with 0.95 chance: n = %d"
% (target, n))
n = math.ceil(math.log(0.05) / math.log(1.0 - 0.05 / COHORT))
print("to bound cohort risk at 0.05 the clean run needs n = %d" % n)
print("cohort attempts in the p13 course: %d" % COHORT)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 8 章「实证闭环:学堂考官泄题红队测试、掌握度追问树与预注册检验」(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}
}