关于「实战:夜珀单品页从诊断、重排到预先登记的 A/B 测试」,《拇指热区与首屏留存工程》给出的定量结论是什么?
结论:第八章让夜珀走完全程:重排后视口税降到 0.258 与 0.267,第一屏的商品信息增加到 164 像素,价格与按钮始终可见,购买路径的操作在左右手模型下都落在舒适区;预先登记的四周测试每组 19,200 次会话,能检测 0.149 以上的相对提升,0.10 的提升只有 0.477 的把握看出来。书中以标注框区分证据、计算、原典与假说,每条假说都写明什么结果会推翻它;原典考辨取《考工记》「天有时,地有气,材有美,工有巧」的良工四条件。
数据口径:全部品牌与数字都是虚构样本,由书中程序算出;涉及隐私同意与无障碍的法规只写一般框架,不是法律意见。 书中以标注框区分证据、计算、原典与假说,每条假说都写明什么结果会推翻它;原典考辨取《考工记》「天有时,地有气,材有美,工有巧」的良工四条件。
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, 0.149, 0.477, 19200
| sessions per arm 19,200 over 4 weeks; minimum detectable lift at power 0.80 | 0.149 |
| analytic power at lift 0.10 | 0.477 |
| analytic power at lift 0.15 | 0.806 |
| analytic power at lift 0.20 | 0.961 |
| sample-ratio check: stop and debug if one arm differs from 19,200 by more than 322 sessions | |
| simulated A/A (no true change) share declared a win 0.021 | |
| simulated true lift 0.10 share declared a win 0.466 | |
| simulated true lift 0.20 share declared a win 0.957 |
查看完整验算程序
# Chapter 8. p83: The pre-registered A/B test of the redesigned Nightamber product template.
# Traffic and baseline are the Chapter 1 p12 sample assumptions. Simulated tests draw each arm's count from a
# normal approximation to the binomial (seeded); they check the design, they are not results of any test.
import math
import random
from statistics import NormalDist
WEEKLY, WEEKS = 9600, 4
BASE = 0.038 # add-to-bag per session, control arm (assumed)
ALPHA, POWER = 0.05, 0.80
N = WEEKLY * WEEKS // 2 # sessions per arm
nd = NormalDist()
za, zb = nd.inv_cdf(1 - ALPHA / 2), nd.inv_cdf(POWER)
def ztest(k1, k2, n):
p = (k1 + k2) / (2 * n)
se = math.sqrt(p * (1 - p) * 2 / n)
return (k2 - k1) / n / se if se > 0 else 0.0
def power_at(lift):
p2 = BASE * (1 + lift)
se = math.sqrt((BASE * (1 - BASE) + p2 * (1 - p2)) / N)
return nd.cdf((p2 - BASE) / se - za)
lo, hi = 0.0, 1.0
for _ in range(60):
mid = (lo + hi) / 2
lo, hi = (mid, hi) if power_at(mid) < POWER else (lo, mid)
print("sessions per arm {:,} over {} weeks; minimum detectable lift at power 0.80: {:.3f}".format(
N, WEEKS, hi))
for lift in (0.10, 0.15, 0.20):
print(" analytic power at lift {:.2f}: {:.3f}".format(lift, power_at(lift)))
srm = nd.inv_cdf(1 - 0.001 / 2) * math.sqrt(2 * N * 0.25)
print("sample-ratio check: stop and debug if one arm differs from {:,} by more than {:.0f} sessions".format(
N, srm))
rng = random.Random(8137)
def draw(p):
return max(0, round(rng.gauss(N * p, math.sqrt(N * p * (1 - p)))))
SIMS = 4000 # a win = z above the two-sided critical value: nominal A/A rate 0.025
for label, lift in (("A/A (no true change)", 0.0), ("true lift 0.10", 0.10), ("true lift 0.20", 0.20)):
hits = sum(1 for _ in range(SIMS) if ztest(draw(BASE), draw(BASE * (1 + lift)), N) > za)
print("simulated {:<22} share declared a win {:.3f}".format(label, hits / SIMS))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《移动端拇指热区动力学、暗黑奢品微交互与首屏视觉留存工程》第 8 章「实战:夜珀单品页从诊断、重排到预先登记的 A/B 测试」(版本 0.1.0). 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-137 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026137,
author = {Fan, Yafu},
title = {Mobile Thumb-Zone Kinetics, Dark-Luxury Micro-Interactions and Above-the-Fold Visual Retention Engineering},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}