关于「第一锚:标准库验算、确定性指纹与双锚的相关失效」,《抗模型崩溃双锚工程》给出的定量结论是什么?

结论:第六章讨论第一锚:标准库验算、math.fsum 的顺序无关与双锚的相关失效,两锚合用使一本 300 个数字的书无错的概率升到 0.6769。

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,不使用任何读者或客户的数据;本书是非个人化的经营与工程参考。 书中以标注框区分已发表的证据、可运行的 Python 计算(22 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: 0.6769, 300

anchors wrong/number per_300 P(book clean)
none0.05000 15.00 0.0000
recompute0.02030 6.09 0.0021
collate0.03100 9.30 0.0001
both0.00130 0.39 0.6769
shared-kind error, each anchor misses with m0.1
rho0.0: both miss 0.0100, gain over one anchor 10.0 times
rho0.5: both miss 0.0550, gain over one anchor 1.8 times
rho0.9: both miss 0.0910, gain over one anchor 1.1 times
rho1.0: both miss 0.1000, gain over one anchor 1.0 times
check: P(clean) with both0.6769, exp(-n r) = 0.6771
查看完整验算程序
# Chapter 6. p62: one anchor or two, and what correlation costs.
# Model: each number in a book is wrong with probability e0 = 0.05
# (illustrative). A share c = 0.6 of the errors are computational
# (a formula, a sum, a seed); the rest are source errors (a misread
# year, a wrong author, a quote altered). Anchor 1, recomputation with
# the standard library, catches a computational error with probability
# d1 = 0.99 and a source error never. Anchor 2, collation with the
# original, catches a source error with d2 = 0.95 and a computational
# error never. A third kind, an error both anchors could see, is
# missed by each with m = 0.1; the misses have correlation rho.
import math

E0, C, D1, D2 = 0.05, 0.6, 0.99, 0.95
NUMS = 300


def residual(a1, a2):
    comp = E0 * C * ((1.0 - D1) if a1 else 1.0)
    src = E0 * (1.0 - C) * ((1.0 - D2) if a2 else 1.0)
    return comp + src


print("anchors      wrong/number  per_%d  P(book clean)" % NUMS)
for name, a1, a2 in (("none", 0, 0), ("recompute", 1, 0),
                     ("collate", 0, 1), ("both", 1, 1)):
    r = residual(a1, a2)
    print("%-12s %-13.5f %-8.2f %.4f" % (name, r, r * NUMS,
                                         (1.0 - r) ** NUMS))
m = 0.1
print("shared-kind error, each anchor misses with m = %.1f" % m)
for rho in (0.0, 0.5, 0.9, 1.0):
    both = m * m + rho * m * (1.0 - m)
    print("rho = %.1f: both miss %.4f, gain over one anchor %.1f times"
          % (rho, both, m / both))
print("check: P(clean) with both = %.4f, exp(-n r) = %.4f" % (
    (1.0 - residual(1, 1)) ** NUMS, math.exp(-NUMS * residual(1, 1))))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《抗模型崩溃双锚真源工程、自产语料回喂方差坍缩速率与标准库验算校勘免疫母典》第 6 章「第一锚:标准库验算、确定性指纹与双锚的相关失效」(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-265 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026265,
  author    = {Fan, Yafu},
  title     = {Anti-Collapse Engineering with Two Anchors of Truth: the Variance-Collapse Rate of Self-Generated Training Data and Immunity through Standard-Library Recomputation and Textual Collation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论