关于「把「致命」与「剥离」换成区间:决策损失带、小效应与推翻条件」,《解释深度剃刀与校准》给出的定量结论是什么?

结论:第六章在宽示意参数下抽样,剃刀与检索分流的净损失减少中位 22.7 %,五分位 8.2 %,九十五分位 41.2 %;全用剃刀中位只有 5.1 %。

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

net loss reduction as share of baseline, 4,000 draws
policy p05 p50 p95
razor + lookup0.082 0.227 0.412
razor on all-0.097 0.051 0.234
share of draws where routing loses money0.0020
share of draws where routing beats razor-on-all1.0000
查看完整验算程序
# Chapter 6. p61: "fatal" and "stripping" as a band of loss reduction.
# Per assumption on a decision list: share m is mechanism-type, the rest
# fact-type. False-knowledge rates f_m, f_f. The razor exposes a false
# mechanism claim with d_m; a lookup exposes a false fact with d_f; an
# exposed error is fixed in time with probability a. Loss per decision
# built on a false claim L = 100; razor costs 4, lookup 1.5 per claim.
# We draw every unknown from a wide illustrative range 4,000 times and
# report the 5th, 50th and 95th percentiles of the net loss reduction as
# a share of the baseline loss. No parameter here is an estimate.
import math
import random

rng = random.Random(39661)
DRAWS = 4000
L, C_RAZOR, C_LOOK = 100.0, 4.0, 1.5


def pct(xs, q):
    s = sorted(xs)
    return s[min(len(s) - 1, int(q * len(s)))]


routed, razor_all, nothing_left = [], [], []
for _ in range(DRAWS):
    m = rng.uniform(0.4, 0.7)
    f_m, f_f = rng.uniform(0.15, 0.35), rng.uniform(0.05, 0.15)
    d_m, d_f = rng.uniform(0.3, 0.7), rng.uniform(0.6, 0.95)
    d_rf = rng.uniform(0.05, 0.25)          # razor used on a fact claim
    a = rng.uniform(0.5, 0.9)
    base = math.fsum([m * f_m * L, (1.0 - m) * f_f * L])
    save_r = math.fsum([m * f_m * L * d_m * a,
                        (1.0 - m) * f_f * L * d_f * a,
                        -m * C_RAZOR, -(1.0 - m) * C_LOOK])
    save_a = math.fsum([m * f_m * L * d_m * a,
                        (1.0 - m) * f_f * L * d_rf * a, -C_RAZOR])
    routed.append(save_r / base)
    razor_all.append(save_a / base)
    nothing_left.append(1.0 if save_r <= 0.0 else 0.0)
print("net loss reduction as share of baseline, {:,} draws".format(DRAWS))
print("policy                p05     p50     p95")
for name, xs in (("razor + lookup", routed), ("razor on all", razor_all)):
    print("{:<20} {:6.3f}  {:6.3f}  {:6.3f}".format(
        name, pct(xs, 0.05), pct(xs, 0.50), pct(xs, 0.95)))
print("share of draws where routing loses money: {:.4f}".format(
    math.fsum(nothing_left) / DRAWS))
better = len([1 for r, z in zip(routed, razor_all) if r > z]) / DRAWS
print("share of draws where routing beats razor-on-all: {:.4f}".format(
    better))

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 6 章「把「致命」与「剥离」换成区间:决策损失带、小效应与推翻条件」(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