关于「防套话的纵深架构:分层失效的相关性、只存摘要的输出检查与题库暴露」,《零泄题苏格拉底考官》给出的定量结论是什么?

结论:程序 p61 说明,纵深防御的各层只有在独立失效时才能连乘,而现有研究揭示的失效机制恰恰会让它们一起失效;只有隔离把极限从 1 变成 0。程序 p62 演示了只存摘要的输出检查:它在隔离架构里负责防倒退和防模型自己算出答案,能拦书写变体,拦不住改写。程序 p63 估算了隔离之后唯一关不上的通道,即学员对自己答对题目的分享,并说明变体与题库更新的作用。账算完了,接下来要回答一个更古老的问题:苏格拉底和孔子在没有服务器的年代,是怎样处理「问而不告」的?下一章回到原典。

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。

layersrhoper attemptk=1000k=10000vs naive
base0.000.0200001.00001.000033.33
instruction0.000.0060000.99761.000010.00
filter0.000.0006000.45130.99751.00
correlated0.050.0013700.74611.00002.28
correlated0.100.0021400.88261.00003.57
isolated0.000.0000000.00000.00000
naive product of layers: 0.000600 per attempt
isolation is the only row whose limit is 0, not 1
查看完整验算程序
# Chapter 6. p61: layered defences multiply only if they fail apart.
# Per-attempt leak probabilities (all illustrative):
#   base         answer in context, no instruction        p0 = 0.02
#   instruction  "never reveal it" keeps 0.3 of leaks      0.3 * p0
#   filter       an output check stops 0.9 of what leaks
#   correlated   a share rho of attempts are of a kind that defeats
#                the instruction outright and slips past the filter
#                with probability 0.8 (for example, reworded output)
#   isolated     the answer never enters the context: the content
#                channel is closed, only the verdict bit remains
P0, KEEP, STOP = 0.02, 0.3, 0.9


def per_attempt(cfg, rho=0.0):
    if cfg == "base":
        return P0
    if cfg == "instruction":
        return KEEP * P0
    weak = KEEP * P0 * (1 - STOP)
    if cfg == "filter":
        return weak
    if cfg == "correlated":
        return rho * P0 * 0.8 + (1 - rho) * weak
    return 0.0


def cum(p, k):
    return 1.0 - (1.0 - p) ** k


rows = (("base", 0.0), ("instruction", 0.0), ("filter", 0.0),
        ("correlated", 0.05), ("correlated", 0.10), ("isolated", 0.0))
print("layers        rho   per attempt  k=1000   k=10000  vs naive")
naive = per_attempt("filter")
for cfg, rho in rows:
    p = per_attempt(cfg, rho)
    ratio = "%.2f" % (p / naive) if p > 0 else "0"
    print("%-13s %-5.2f %-12.6f %-8.4f %-8.4f %s"
          % (cfg, rho, p, cum(p, 1000), cum(p, 10000), ratio))
print("naive product of layers: %.6f per attempt" % naive)
print("isolation is the only row whose limit is 0, not 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 6 章「防套话的纵深架构:分层失效的相关性、只存摘要的输出检查与题库暴露」(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