关于「反向理由、事前验尸与后见之明的封存」,《解释深度剃刀与校准》给出的定量结论是什么?

结论:写一条反对理由,可以缩小过度自信;可它能改善的只是 Brier 分数里很薄的一层,在本章示意模型里,最多 0.00931。事前验尸让失败原因的覆盖率随人数饱和,五到七人足够,关键在于写的人背景不同。结果揭晓后哪怕只改写一成原始确信度,账面就会出现远大于真实改善的假进步,台账必须封存。

数据口径:书中参数为示意,数据全部合成,全部计算由书中程序生成。 第一章在合成数据里量出装置类自评下降 1.283 分,事实类下降区间从 -0.263 到 0.063,跨过零,并给出剃刀与检索的分流规则。

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: 0.00931

deltasRELBSBS gainshare of REL removed
0.01.80.009310.191810.000000.000
0.11.70.007680.190180.001630.175
0.21.60.006100.188600.003220.345
0.41.40.003180.185680.006130.658
0.81.00.000000.182500.009311.000
irreducible part E[q(1-q)] = 0.18250
查看完整验算程序
# Chapter 5. p51: what writing one reason against buys in calibration.
# The judge reports logit(p) = s logit(q) with s = 1.8 (overconfident).
# Koriat, Lichtenstein & Fischhoff (1980) found that listing contradicting
# reasons improved the fit of confidence; we do not know by how much in a
# given team, so the intervention shrinks s by an unknown delta. Because
# p is monotone in q, the reliability term is E[(p - q)^2] exactly and
# BS = REL + E[q(1 - q)]. q is uniform on (0.05, 0.95); we integrate on a
# grid of 9,000 midpoints. Delta values are illustrative.
import math

M = 9000
QS = [0.05 + 0.9 * (i + 0.5) / M for i in range(M)]
UNC = math.fsum(q * (1.0 - q) for q in QS) / M


def logit(x):
    return math.log(x / (1.0 - x))


def rel(s):
    return math.fsum((1.0 / (1.0 + math.exp(-s * logit(q))) - q) ** 2
                     for q in QS) / M


base = rel(1.8)
print("delta   s     REL      BS       BS gain   share of REL removed")
for d in (0.0, 0.1, 0.2, 0.4, 0.8):
    r = rel(1.8 - d)
    print(" {:.1f}   {:.1f}   {:.5f}  {:.5f}  {:.5f}   {:6.3f}".format(
        d, 1.8 - d, r, r + UNC, base - r, (base - r) / base))
print("irreducible part E[q(1-q)] = {:.5f}".format(UNC))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《解释深度错觉破除、认知边界校准与真知假知剥离剃刀母典》第 5 章「反向理由、事前验尸与后见之明的封存」(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-396 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026396,
  author    = {Fan, Yafu},
  title     = {Shattering the Illusion of Explanatory Depth, Epistemic Calibration and the Razor Separating True Knowledge from Assumed Familiarity},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论