5D Library › 股东搭便车博弈、动态股权成熟与背刺早期预警模型 · English

第八章 实践:股权表、关键人物测试,以及警报不是证据

经历一次离开与一轮融资的股权表

p14 跟踪三位按四年成熟的创始人,经历其中一位离开与一轮种子融资。

推导

# Chapter 8. p14: A cap table through one departure and one round.
# Founders A, B, C start with 40%, 35% and 25%, all on four-year
# monthly vesting with a one-year cliff. C leaves in month 15 as a good
# leaver: C keeps the vested part, and the firm buys back the unvested
# part at cost and cancels it. In month 18 a 10% option pool is
# created and a seed investor buys 20%, both after the buy-back
# (the pool and the investor's shares are counted in the new total).
def run(c_month=15):
    shares = {"A": 40.0, "B": 35.0, "C": 25.0}
    v = 0.0 if c_month < 12 else c_month / 48
    shares["C"] = 25.0 * v
    old = sum(shares.values())
    new_total = old / (1 - 0.10 - 0.20)
    shares["pool"] = 0.10 * new_total
    shares["seed"] = 0.20 * new_total
    return shares, new_total

if __name__ == "__main__":
    sh, tot = run()
    print("   holder   shares    stake")
    for k, x in sh.items():
        print("   {:<6} {:8.2f}   {:6.3f}".format(k, x, x / tot))
    print("   C's stake if C had held a static 25%: {:.3f}".format(
        25.0 / (100 / 0.7)))
    sh2, tot2 = run(c_month=11)
    print("   if C had left in month 11: C {:.3f}, A {:.3f}".format(
        sh2["C"] / tot2, sh2["A"] / tot2))
#    holder   shares    stake
#    A         40.00    0.338
#    B         35.00    0.296
#    C          7.81    0.066
#    pool      11.83    0.100
#    seed      23.66    0.200
#    C's stake if C had held a static 25%: 0.175
#    if C had left in month 11: C 0.000, A 0.373

C 在第 15 个月作为好离职者离开、未成熟股份被回购之后,再设立 10% 的期权池、引入 20% 的种子投资,A 持有公司的 0.338,B 0.296,C 0.066,期权池 0.100,投资人 0.200。如果 C 持有静态的 25%,C 仍会拥有一家自己已经不再为之工作的公司的 0.175。如果 C 在悬崖期之前的第 11 个月离开,C 一股也没有,A 持有 0.373。

一个人能带走多少收入?

推导

# Chapter 8. p15: How much revenue walks out with one person? Twenty
# client accounts of unequal size (shares proportional to 1/k, k = 1
# to 20) are each held by one of four executives, assigned in turn.
# Each executive leaves to a competitor in a year with chance 0.1; a
# client follows the leaver with chance 0.6 if only the leaver knows
# them, and 0.15 if a second executive also covers the account. Expected
# revenue lost, and the chance of losing a fifth or more, with and
# without dual coverage. 50,000 years simulated, seed 93.
import random

W = [1 / k for k in range(1, 21)]
S = [w / sum(W) for w in W]
OWNER = [k % 4 for k in range(20)]

def year(rng, follow):
    gone = [rng.random() < 0.1 for _ in range(4)]
    lost = 0.0
    for k in range(20):
        if gone[OWNER[k]] and rng.random() < follow:
            lost += S[k]
    return lost

if __name__ == "__main__":
    rng = random.Random(93)
    print("   coverage        expected loss   P(loss >= 0.2)")
    for lab, f in (("single owner", 0.6), ("dual coverage", 0.15)):
        ys = [year(rng, f) for _ in range(50000)]
        print("   {:<15} {:13.4f}   {:14.4f}".format(
            lab, sum(ys) / len(ys), sum(1 for y in ys if y >= 0.2) /
            len(ys)))
    print("   largest client's share of revenue: {:.3f}".format(S[0]))
    print("   executive 0's book: {:.3f}".format(
        sum(S[k] for k in range(20) if OWNER[k] == 0)))
#    coverage        expected loss   P(loss >= 0.2)
#    single owner           0.0594           0.1099
#    dual coverage          0.0152           0.0172
#    largest client's share of revenue: 0.278
#    executive 0's book: 0.402

二十个大小不一的客户分给四位高管,每位一年内离开的机会是 0.1;每个客户只有一位高管认识时,预期收入损失 0.0594,损失五分之一或以上的机会是 0.1099。当每个客户都有第二位高管覆盖、客户跟着离开者走的机会只有 15% 时,预期损失降到 0.0152,损失五分之一的机会降到 0.0172。最大的客户单独占收入的 0.278,一位高管的客户合计占 0.402。关键客户的双人覆盖,是本书里最便宜的装甲。

警报不是证据

推导

# Chapter 8. p16: An alarm is not proof. Start from a 2% base rate that
# a given partner is preparing to defect. An SPRT alarm (p09) has
# likelihood ratio (1 - beta) / alpha = 95. Each further independent
# check that comes back adverse (an audit of client records, a
# repository log, a supplier's statement) has likelihood ratio 5; one
# that comes back clean has 0.4. Belief after the alarm and after each
# result, and the number of adverse checks needed to reach 0.95.
def update(p, lr):
    o = p / (1 - p) * lr
    return o / (1 + o)

if __name__ == "__main__":
    p = update(0.02, 95)
    print("   after the alarm alone: {:.3f}".format(p))
    q = p
    for i in range(1, 4):
        q = update(q, 5)
        print("   after {} adverse check(s): {:.3f}".format(i, q))
    print("   after the alarm and one clean check: {:.3f}".format(
        update(p, 0.4)))
    n, q = 0, p
    while q < 0.95:
        q = update(q, 5)
        n += 1
    print("   adverse checks needed to reach 0.95: {}".format(n))
#    after the alarm alone: 0.660
#    after 1 adverse check(s): 0.906
#    after 2 adverse check(s): 0.980
#    after 3 adverse check(s): 0.996
#    after the alarm and one clean check: 0.437
#    adverse checks needed to reach 0.95: 2

从 2% 的基础比率出发,第五章检验的一次警报,把「这位合伙人在准备背叛」的信念提高到 0.660。再有一次不利的核查结果,提高到 0.906;两次,0.980。警报之后有一次干净的核查结果,则降到 0.437。需要两项独立的不利发现,信念才超过 0.95。实务规则由此而来:警报触发合法的核查,以及一次让当事人可以解释的谈话;行动要等独立的证据。

利益披露

声明: 本书讨论合伙人之间激励、契约与监测的数学,不是法律、税务或投资建议。股东协议、股权成熟、离职条款与限制性条款,必须由合格律师依照适用法律起草与审阅;对员工或合伙人的任何监测,都必须遵守隐私、数据保护与劳动法律。

注

预先登记的检验: 第二章关于成熟机制与留任的假说,与第五章关于离职前资源转移率的假说,都应在收集数据之前登记数据来源、定义与分析方法,结果不论正负都发表。使用事件日志的研究必须使用依法收集、匿名化的数据,并经过伦理审查。

小结

成熟机制与回购让股权留在建设公司的人手里;关键关系的双人覆盖限制了一次离开能带走的东西;警报是核查的理由,不是指控的理由。

Fan, Y. (2026). Shareholder Free-Riding Games, Dynamic Vesting Mechanisms and Defection Early-Warning Models: Moral-Hazard Equity Dilution, Shotgun Buy-Sell Clauses and Behavioral Divergence Signals [股东搭便车博弈、动态股权成熟与背刺早期预警模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090243 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).