关于「解释深度错觉的边界:机制类最强,事实类最弱」,《解释深度剃刀与校准》给出的定量结论是什么?
结论: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。
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。
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: -0.263, 0.063, 1.283
| domain | T1 | T2 | drop | sd | d_z | 95% CI of drop | >=1 | |
| device | 4.53 | 3.25 | 1.283 | 1.04 | 1.24 | [ 1.098, | 1.469] | 0.78 |
| natural | 4.47 | 3.63 | 0.833 | 1.00 | 0.83 | [ 0.655, | 1.012] | 0.59 |
| procedure | 4.41 | 4.08 | 0.325 | 1.06 | 0.31 | [ 0.135, | 0.515] | 0.45 |
| narrative | 4.55 | 4.30 | 0.250 | 0.99 | 0.25 | [ 0.073, | 0.427] | 0.38 |
| fact | 4.24 | 4.34 -0.100 | 0.91 -0.11 | [-0.263, | 0.063] | 0.24 | ||
| raters per domain: 120; drop = T1 - T2 on a 1-7 scale |
查看完整验算程序
# Chapter 1. p11: rating drop after writing the mechanism, by domain.
# Model: each synthetic rater gives a 1-7 self-rating of understanding (T1),
# writes an explanation, then rates again (T2). The latent drop has a mean
# that depends on the knowledge domain. The means are illustrative and only
# follow the direction reported by Rozenblit & Keil (2002): largest for
# devices and natural phenomena, small for procedures, narratives and facts.
# Ratings are rounded and clipped to 1..7, as on a real seven-point scale.
import math
import random
import statistics
rng = random.Random(39611)
N = 120 # raters per domain (illustrative)
DOMAINS = [ # (label, mean T1, mean latent drop)
("device", 4.6, 1.10),
("natural", 4.3, 0.90),
("procedure", 4.4, 0.40),
("narrative", 4.5, 0.20),
("fact", 4.2, 0.10),
]
SD_T1, SD_DROP = 1.0, 0.9
def clip7(x):
return min(7, max(1, int(round(x))))
print("domain T1 T2 drop sd d_z 95% CI of drop >=1")
for label, m1, md in DOMAINS:
drops, t1s, t2s = [], [], []
for _ in range(N):
a = rng.gauss(m1, SD_T1)
t1 = clip7(a)
t2 = clip7(a - rng.gauss(md, SD_DROP))
t1s.append(t1)
t2s.append(t2)
drops.append(t1 - t2)
mean = statistics.fmean(drops)
sd = statistics.stdev(drops)
se = sd / math.sqrt(N)
lo, hi = mean - 1.96 * se, mean + 1.96 * se
share = len([d for d in drops if d >= 1]) / N
print("{:<9} {:5.2f} {:5.2f} {:6.3f} {:5.2f} {:5.2f} [{:6.3f}, {:6.3f}]"
" {:4.2f}".format(label, statistics.fmean(t1s),
statistics.fmean(t2s), mean, sd, mean / sd,
lo, hi, share))
print("raters per domain: {}; drop = T1 - T2 on a 1-7 scale".format(N))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 1 章「解释深度错觉的边界:机制类最强,事实类最弱」(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}
}