5D Library › 分拆经济学、防套利身份锁定与多阶授权阶梯 · English
第八章 菜单审计与一项预注册的实验
给任何价格菜单的四个问题
本章把全书变成四个问题,可以向任何数字测算服务的价格菜单——包括作者自己的——提出。它们不是经营建议。
一、有没有一张比高估值客群单品价更便宜的全通卡? 若有,它会把这个客群整体拉下来(第二章)。
二、附赠名额是否超过了套利门槛? 门槛是 ;若允许自购返佣, 要减半(第三、四章)。
三、公司从每个被转售的名额还能净得多少? 若低于交付成本,渠道已经穿仓(第四章)。
四、价格台阶中有没有一步大到让某个客群无货可买? 第五章的最优菜单在 567 与 12537 之间有一步 22.11 倍的跳跃,这是甄别的结果,但需要人工审查:中间是否存在一个被忽略的客群(例如个人兼职从业者)。
用程序做审计
程序把这四个问题写成四个自动标记,用在两张菜单上:一张是常见的「廉价全包」菜单(单次 99、年卡 199 附赠 6 个名额且允许自购返佣、全通卡 399、牌照 2999),一张是第五章的锁定阶梯。
推导
# Chapter 8. [p15] A four-flag audit of a price menu.
# Flags: (1) CANNIBAL - an all-access pass priced below the top single
# price a high-value segment pays (p03); (2) RESALE - unlocked slots per
# card above the threshold k + 1 > P / ((1 - g) p2), with the card cost
# halved if self-referral is paid (p05, p07); (3) BLOW - firm net per
# resold slot below the delivery cost c = 15 (p07); (4) GAP - a step
# ratio above 20 with no tier between, so one segment has no product
# (p09 has one such step, 22.11, by design: flagged for review only).
G, C = 0.5, 15.0
MENUS = {
"cheap all-in": dict(p2=99, card=199, k=6, self_ref=True,
pass_=399, b_single=2999,
ladder=(9, 99, 199, 399, 2999, 9999)),
"locked ladder": dict(p2=117, card=287, k=0, self_ref=False,
pass_=None, b_single=12537,
ladder=(12, 117, 287, 567, 12537, 34537)),
}
def audit(m):
flags = []
if m["pass_"] is not None and m["pass_"] < m["b_single"]:
flags.append("CANNIBAL pass {} < B2B price {}".format(
m["pass_"], m["b_single"]))
cost = m["card"] * ((1 - G) if m["self_ref"] else 1.0)
thr = cost / ((1 - G) * m["p2"])
if m["k"] + 1 > thr:
flags.append("RESALE k+1 = {} > {:.2f}".format(m["k"] + 1, thr))
net = cost / (m["k"] + 1) - C
if m["k"] > 0 and net < 0:
flags.append("BLOW net per slot {:.2f}".format(net))
lad = m["ladder"]
for a, b in zip(lad, lad[1:]):
if b / a > 20:
flags.append("GAP {} -> {} (x{:.2f})".format(a, b, b / a))
return flags
if __name__ == "__main__":
for name, m in MENUS.items():
fl = audit(m)
print(" {}: {} flag(s)".format(name, len(fl)))
for f in fl:
print(" - " + f)
print("\nReading: the cheap all-in menu trips the first three flags")
print("at once; the locked ladder keeps only the reviewed B2B step.")
# cheap all-in: 3 flag(s)
# - CANNIBAL pass 399 < B2B price 2999
# - RESALE k+1 = 7 > 2.01
# - BLOW net per slot -0.79
# locked ladder: 1 flag(s)
# - GAP 567 -> 12537 (x22.11)
#
# Reading: the cheap all-in menu trips the first three flags
# at once; the locked ladder keeps only the reviewed B2B step.廉价全包菜单同时触发前三个标记:全通卡 399 低于 B2B 单品价 2999;附赠名额使 ,远高于自购返佣下的门槛 2.01;每个被转售名额净亏 0.79。锁定阶梯只保留第四个标记,即那一步有意设计、需要人工审查的 B2B 跳跃。
一项预注册的实验
假说 · 尚未验证
研究方案。 来访者按访问会话随机分配到现行菜单与锁定阶梯两组;渠道伙伴按伙伴随机分配到「附赠名额 + 自购返佣」与「身份锁 + 不对本人购买付佣金」两种渠道规则。主要结局:渠道伙伴中,一张卡在一年内出现一个以上出生哈希(经人工核实非家庭共用)的比例;预期从 0.15 降到 0.05。次要结局:每位来访者的利润;正常用户在身份核实步骤的放弃率(第三章的证伪条件)。检验:主要结局用双侧 0.05 的两比例检验,把握度 0.90。由与作者无经济关系的独立分析者按预先登记的代码执行,结果不论正负都发表。
推导
# Chapter 8. [p16] Sample sizes for a pre-registered menu experiment.
# Visitors are randomised between the current menu and the locked
# ladder. (a) Profit per visitor is whale-driven: its standard deviation
# under the p09 model (shares, tiers, margins) is computed exactly, and
# the visitors per arm needed to detect a 20% change (two-sided 0.05,
# power 0.90) follow from the normal approximation. (b) A secondary,
# better-powered outcome: the share of channel partners whose card is
# used for more than one birth-record hash, 0.15 -> 0.05.
import math
from statistics import NormalDist
SHARE = (0.60, 0.22, 0.10, 0.05, 0.025, 0.005)
MARGIN = (11, 114, 279, 555, 11037, 28537)
def n_means(sd, delta, alpha=0.05, power=0.90):
z = NormalDist().inv_cdf
return math.ceil(2 * ((z(1 - alpha / 2) + z(power)) * sd / delta) ** 2)
def n_props(p1, p2, alpha=0.05, power=0.90):
z = NormalDist().inv_cdf
pb = (p1 + p2) / 2
a = z(1 - alpha / 2) * math.sqrt(2 * pb * (1 - pb))
b = z(power) * math.sqrt(p1 * (1 - p1) + p2 * (1 - p2))
return math.ceil((a + b) ** 2 / (p1 - p2) ** 2)
if __name__ == "__main__":
mean = sum(s * m for s, m in zip(SHARE, MARGIN))
var = sum(s * m * m for s, m in zip(SHARE, MARGIN)) - mean ** 2
sd = math.sqrt(var)
print(" profit per visitor: mean {:.2f}, sd {:.2f}, CV {:.2f}".format(
mean, sd, sd / mean))
n = n_means(sd, 0.2 * mean)
print(" (a) visitors per arm for a 20% change: {:,d}".format(n))
print(" (b) partners per arm for 0.15 -> 0.05: {:,d}".format(
n_props(0.15, 0.05)))
print("\nReading: profit per visitor is dominated by a few B2B")
print("buyers, so a fair test of it needs tens of thousands of")
print("visitors per arm; the channel outcome needs a few hundred")
print("partners. Register both, with profit as a secondary outcome.")
# profit per visitor: mean 505.94, sd 2624.37, CV 5.19
# (a) visitors per arm for a 20% change: 14,136
# (b) partners per arm for 0.15 -> 0.05: 188
#
# Reading: profit per visitor is dominated by a few B2B
# buyers, so a fair test of it needs tens of thousands of
# visitors per arm; the channel outcome needs a few hundred
# partners. Register both, with profit as a secondary outcome.每位来访者的利润由少数 B2B 买家决定,标准差是均值的 5.19 倍;要以 0.90 的把握度检出 20% 的变化,每组需要 14,136 位来访者。渠道结局则只需每组 188 位伙伴。所以本方案把渠道结局定为主要结局,把利润列为次要结局:一项以利润为主要结局、样本又不够大的实验,几乎注定得出「没有差异」。
小结
四个问题——全通卡是否低于高估值单品价、附赠名额是否越过套利门槛、转售名额是否穿仓、台阶是否留下空档——可以写成自动审计。检验身份锁效果的实验应当以渠道结局为主要结局,每组 188 位伙伴即可;以每位来访者利润为主要结局的实验,每组需要 14,136 位来访者。
Fan, Y. (2026). Unbundling Economics, Anti-Arbitrage Identity Locking and Multi-Tier Licensing Ladders: Orthogonal Demand Partitioning, Resale Cannibalization Bounds, Affiliate Margin Preservation and Five-Figure B2B Upsell Equilibria [分拆经济学、防套利身份锁定与多阶授权阶梯] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090184 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).