关于「假设清单评分与信息价值:先验证哪一条」,《解释深度剃刀与校准》给出的定量结论是什么?
结论:再看信息价值。先验为 0.400 时,当下不上线,EVPI 是 24.00,试投的 EVSI 是 8.40,约为完全信息价值的 0.350。先验升到 0.550,EVSI 升到 19.05。恰好在打平点 0.625,EVSI 达到最大的 24.38,占完全信息价值的 0.650。过了打平点,当下已经会上线,EVSI 开始下降:先验 0.700 时是 17.70,0.800 时是 8.80,0.900 时降到 0.00。
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。 与结论吻合的打印值: 0.00, 0.350, 0.400, 0.550, 0.625, 0.650
| prior p | act now | value now | EVPI | EVSI | EVSI/EVPI |
|---|---|---|---|---|---|
| 0.400 | hold | 0.00 | 24.00 | 8.40 | 0.350 |
| 0.550 | hold | 0.00 | 33.00 | 19.05 | 0.577 |
| 0.625 | hold | 0.00 | 37.50 | 24.38 | 0.650 |
| 0.700 | launch | 12.00 | 30.00 | 17.70 | 0.590 |
| 0.800 | launch | 28.00 | 20.00 | 8.80 | 0.440 |
| 0.900 | launch | 44.00 | 10.00 | 0.00 | 0.000 |
| break-even prior where launch and hold tie: 0.625 |
查看完整验算程序
# Chapter 4. p42: value of information for one key assumption.
# Launch decision: if the assumption is true the launch earns G = 60;
# if false it loses L = 100; not launching is 0. Prior P(true) = p.
# A pilot test reports "true" with sensitivity se and specificity sp.
# EVPI = value with perfect knowledge - value now; EVSI uses the imperfect
# pilot (Howard 1966). The pilot is worth running only if EVSI > its cost.
# Units are illustrative.
G, L = 60.0, 100.0
SE, SP = 0.85, 0.80
def best(p):
return max(0.0, p * G - (1.0 - p) * L)
def evsi(p):
pos = p * SE + (1.0 - p) * (1.0 - SP)
neg = 1.0 - pos
p_pos = p * SE / pos
p_neg = p * (1.0 - SE) / neg
return pos * best(p_pos) + neg * best(p_neg) - best(p)
print("prior p act now value now EVPI EVSI EVSI/EVPI")
for p in (0.40, 0.55, 0.625, 0.70, 0.80, 0.90):
now = best(p)
evpi = p * G - now
act = "launch" if now > 0 else "hold"
e = evsi(p)
print(" {:5.3f} {:<6} {:6.2f} {:6.2f} {:6.2f} {:6.3f}"
.format(p, act, now, evpi, e, e / evpi if evpi else 0.0))
print("break-even prior where launch and hold tie: {:.3f}".format(
L / (G + L)))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 4 章「假设清单评分与信息价值:先验证哪一条」(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}
}