关于「回应谁、怎么回、多快回:管理者回应的反转条件与文案红线」,《声誉反脆弱与受众主权》给出的定量结论是什么?
结论:B 段才是重点。真实效果是 +0.12 星,可是一家店前后各只有 30 条评价时,观察到的均值差标准差是 0.297,有 0.316 的机会看到的是负数,也就是说,近三分之一的店会以为回应让评分变差了。前后各 100 条时,看错方向的机会仍有 0.225;300 条时 0.104;到 1,000 条才降到 0.013。
数据口径:书中参数为示意;全部计算由书中程序生成;本书是非个人化经营建议,不是法律意见。 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。
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: 0.013, 0.104, 0.12, 0.225, 0.297, 0.316
| A. policy | 1-star 5-star mean |
| tailored to negatives | 0.042 0.507 4.102 |
| template to all | 0.051 0.476 4.012 |
| reply to positives | 0.060 0.450 3.932 |
| silence | 0.054 0.466 3.982 |
| B. true shift +0.12: what one shop observes | |
| n per side mean diff sd P(diff > 0) P(diff < 0) | |
| 30 +0.121 | 0.297 0.643 0.316 |
| 100 +0.122 | 0.166 0.761 0.225 |
| 300 +0.122 | 0.098 0.889 0.104 |
| 1,000 +0.120 | 0.053 0.986 0.013 |
查看完整验算程序
# Chapter 3. p31: can one shop see a reply policy work in its own data?
# Posted star shares come from p11. A policy tilts the shares
# p_i ~ p_i exp(theta i) until the mean moves by a target: +0.12 for
# tailored replies to negative reviews (the order of magnitude that
# Proserpio and Zervas 2017 report after hotels began responding),
# +0.03 for one template reply to everything, -0.05 for replying to
# positive reviews. The shifts are illustrative, not estimates for
# any platform. Monte Carlo with a fixed seed: n reviews before and
# n after, 2,000 simulated shops per n.
import math
import random
import statistics
BASE = (0.054, 0.095, 0.132, 0.252, 0.466)
POLICY = {"tailored to negatives": 0.12, "template to all": 0.03,
"reply to positives": -0.05, "silence": 0.0}
SHOPS = 2000
def mean(p):
return math.fsum((i + 1) * x for i, x in enumerate(p))
def tilt(p, theta):
w = [x * math.exp(theta * (i + 1)) for i, x in enumerate(p)]
t = math.fsum(w)
return [x / t for x in w]
def shift_to(p, delta):
target = mean(p) + delta
lo, hi = -3.0, 3.0
for _ in range(80):
mid = (lo + hi) / 2.0
if mean(tilt(p, mid)) < target:
lo = mid
else:
hi = mid
return tilt(p, (lo + hi) / 2.0)
def cum(p):
return [math.fsum(p[:i + 1]) for i in range(len(p))]
if __name__ == "__main__":
base = [x / math.fsum(BASE) for x in BASE]
print("A. policy 1-star 5-star mean")
for name, d in POLICY.items():
q = shift_to(base, d)
print(" {:<23} {:.3f} {:.3f} {:.3f}".format(
name, q[0], q[4], mean(q)))
rng = random.Random(40301)
stars = [1, 2, 3, 4, 5]
cb = cum(base)
cq = cum(shift_to(base, POLICY["tailored to negatives"]))
print("B. true shift +0.12: what one shop observes")
print(" n per side mean diff sd P(diff > 0) P(diff < 0)")
for n in (30, 100, 300, 1000):
diffs = []
for _ in range(SHOPS):
a = rng.choices(stars, cum_weights=cb, k=n)
b = rng.choices(stars, cum_weights=cq, k=n)
diffs.append(statistics.fmean(b) - statistics.fmean(a))
pos = len([d for d in diffs if d > 0]) / SHOPS
neg = len([d for d in diffs if d < 0]) / SHOPS
print(" {:>10,} {:+.3f} {:.3f} {:.3f} {:.3f}".format(
n, statistics.fmean(diffs), statistics.pstdev(diffs), pos, neg))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《商业声誉反脆弱柔道、恶意差评证据链反制与去平台依赖私域主权母典》第 3 章「回应谁、怎么回、多快回:管理者回应的反转条件与文案红线」(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-403 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026403,
author = {Fan, Yafu},
title = {Antifragile Reputation Judo, Evidence-Backed Smear Neutralisation and Platform-Independent Audience Sovereignty},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}