关于「实证闭环:合规改写库接攻心文案检查器与非劣效检验」,《极限词合规改写》给出的定量结论是什么?

结论:第八章给出合规改写库接攻心文案检查器的每周流程、蒸馏小工具「合规改写自动机」与预注册非劣效检验方案:相对界值 0.10、效能 0.80 时每组需要 39,981 次访问。书中以标注框区分已发表的证据、可运行的 Python 计算(18 段程序)、原典与作者的假说。

数据口径:在示意参数下,一个极限词要抵消每月 0.010 的出事概率乘以出事成本,需要 0.0333 的相对转化增益;已发表的研究提示吹嘘的作用在不同顾客那里正负相抵,混合模型中增益超过盈亏平衡值的情形只占 0.042。 书中的词表、文案、事实与参数均为示意;全部计算由书中程序生成;本书是非个人化法律建议,也不提供任何绕开平台审核的手法。

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: 0.10, 0.80, 39981

base conversion 0.030, one-sided alpha 0.05, 4000 sessions a day
margin m  power   n per arm   days   (true rewrite effect 0)
  0.05    0.80      159923     80
  0.05    0.90      221519    111
  0.10    0.80       39981     20
  0.10    0.90       55380     28
  0.15    0.80       17770      9
  0.15    0.90       24614     13
if the rewrite truly loses 0.03 of conversion, margin 0.10:
  n per arm 80406, days 41
decision rule: claim non-inferiority only if the lower one-sided
95 % bound of cB - cA * (1 - m) is above 0; report it either way
查看完整验算程序
# Chapter 8. p82: sample size for a pre-registered non-inferiority test.
# Model: arm A shows the original line, arm B the compliant rewrite.
# H0: c_B <= c_A * (1 - m), H1: c_B > c_A * (1 - m), m = relative margin.
# Normal approximation, one-sided level alpha = 0.05, equal arms:
#   n per arm = (z_{1-alpha} + z_{1-beta})^2 * (vA + vB) / gap^2,
# vA = cA (1 - cA), vB = cB (1 - cB), gap = cB - cA (1 - m), assuming
# the true conversions are equal (cB = cA). Base conversion and daily
# traffic are illustrative. Traffic pooled over a library of listings.
import math
from statistics import NormalDist

Z = NormalDist().inv_cdf
CA = 0.030
DAILY = 4000            # sessions a day across the pooled listings


def n_per_arm(m, power, cb=CA, alpha=0.05):
    gap = cb - CA * (1.0 - m)
    var = CA * (1 - CA) + cb * (1 - cb)
    return math.ceil((Z(1 - alpha) + Z(power)) ** 2 * var / gap ** 2)


print("base conversion %.3f, one-sided alpha 0.05, %d sessions a day"
      % (CA, DAILY))
print("margin m  power   n per arm   days   (true rewrite effect 0)")
for m in [0.05, 0.10, 0.15]:
    for pw in [0.80, 0.90]:
        n = n_per_arm(m, pw)
        print("  %.2f    %.2f  %10d  %5d" % (m, pw, n,
                                            math.ceil(2 * n / DAILY)))
print()
print("if the rewrite truly loses 0.03 of conversion, margin 0.10:")
n = n_per_arm(0.10, 0.80, cb=CA * 0.97)
print("  n per arm %d, days %d" % (n, math.ceil(2 * n / DAILY)))
print("decision rule: claim non-inferiority only if the lower one-sided")
print("95 % bound of cB - cA * (1 - m) is above 0; report it either way")

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《平台敏感词合规改写、广告法极限词自动机与语义保真转化弹性母典》第 8 章「实证闭环:合规改写库接攻心文案检查器与非劣效检验」(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-274 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026274,
  author    = {Fan, Yafu},
  title     = {Compliant Rewriting for Platform-Sensitive Terms: Absolute-Term Automata under the Advertising Law, Meaning-Preserving Rewrites and Conversion Elasticity},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论