关于「叮嘱无效:上下文里的答案与累计泄露概率」,《零泄题苏格拉底考官》给出的定量结论是什么?
结论: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。
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: 0.9727, 30, 40
| Q | 40 items, k = 3 attempts per student per item |
| p N P(item) E[leaked] P(any) | |
| 0.0010 | 30 0.0861 3.44 0.9727 |
| 0.0010 | 200 0.4514 18.05 1.0000 |
| 0.0010 | 1000 0.9503 38.01 1.0000 |
| 0.0001 | 30 0.0090 0.36 0.3023 |
| 0.0001 | 200 0.0582 2.33 0.9093 |
| 0.0001 | 1000 0.2592 10.37 1.0000 |
| p | 0.00100: an item is at even odds after 693 attempts on it |
| p | 0.00010: an item is at even odds after 6932 attempts on it |
| p | 0.00001: an item is at even odds after 69315 attempts on it |
| tenfold smaller p, tenfold more attempts; the limit stays 1 |
查看完整验算程序
# Chapter 1. p13: one cohort, many items: the leak is a cohort event.
# A course has Q items whose answers sit in the examiner's context.
# Each of N students makes k probing attempts per item (illustrative).
# A leaked answer can be shared, so the owner's question is not "will
# this student extract it" but "will anyone extract any of them".
# P(item leaks) = 1 - (1 - p)^(N k)
# expected leaked items = Q * P(item leaks)
# P(any item leaks) = 1 - (1 - p)^(N k Q)
# An instruction that cuts p tenfold only multiplies the attempts
# needed by about ten: it rescales the clock, not the limit.
import math
Q, K = 40, 3
def leak(p, n):
return 1.0 - math.exp(n * math.log1p(-p))
print("Q = %d items, k = %d attempts per student per item" % (Q, K))
print("p N P(item) E[leaked] P(any)")
for p in (0.001, 0.0001):
for n in (30, 200, 1000):
att = n * K
per_item = leak(p, att)
print("%-8.4f %-6d %-9.4f %-10.2f %.4f"
% (p, n, per_item, Q * per_item, leak(p, att * Q)))
for p in (0.001, 0.0001, 0.00001):
half = math.log(0.5) / math.log1p(-p)
print("p = %.5f: an item is at even odds after %d attempts on it"
% (p, math.ceil(half)))
print("tenfold smaller p, tenfold more attempts; the limit stays 1")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 1 章「叮嘱无效:上下文里的答案与累计泄露概率」(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}
}