关于「实证闭环:三站买量台账、蒸馏小工具与预注册证伪」,《收费站买量精算》给出的定量结论是什么?

结论:第八章给出三站买量台账、蒸馏小工具「买量止损与扩量阶梯试算器」与预注册证伪方案。书中以标注框区分已发表的证据、可运行的 Python 计算(20 段程序)、原典与作者的假说。

数据口径:按示意参数,tarotmalaysia 一个破冰买家的期望净值是破冰价的 2.68 倍,按首单回本线止损,在四十条合成素材上只拿到可得利润的 0.222。 书中参数为示意,全部计算由书中程序生成。

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

rowspendbuyVincCPAllrstatenext/day
D5D d-076109.80n/a3.11STOP0
D5D d-111402210.199.09-3.80KEEP86
TMY m-034203128.2220.84-6.53KEEP157
TMY m-09160225.85123.082.19TEST100
TMY m-122,60011822.7733.903.93STOP0
TTW t-023,40030 193.13188.89-3.67KEEP900
TTW t-055200 186.64n/a1.39TEST600
查看完整验算程序
# Chapter 8. p81: one day of the ad ledger for the three live paywalls.
# Each row is one creative: spend, icebreaker buyers from the site's own
# order table, upsells, refunds, and (for kept creatives) the marginal
# ROAS from the last dither pair (chapter 6). Steps:
#  1. funnel value V with upsell and refund rates shrunk toward the
#     site prior (prior weight 50 buyers; p11 funnel)
#  2. incremental buyers = buyers * kappa (weekly holdout, chapter 5)
#  3. sequential state, H0 CPA = V, H1 CPA = 2 V (p32): STOP, KEEP or
#     TEST; zero buyers at 2.996 V also stops (p21)
#  4. next budget: STOP -> 0; TEST -> hold; KEEP -> x exp(clip(0.25
#     (mROAS - 1), -0.2, 0.2)) (p62). All rows are synthetic.
import math

SITE = {  # P1, rate, fix, tax, cost1, u0, P2, cost2, refund0, rep, kappa
    "D5D": (4.90, 0.044, 0.30, 0.00, 0.20, 0.10, 39.0, 0.60, 0.05, 2.5,
            0.70),
    "TMY": (9.90, 0.034, 0.50, 0.06, 0.40, 0.12, 129.0, 1.50, 0.06, 6.0,
            0.65),
    "TTW": (99.0, 0.028, 5.0, 0.05, 4.0, 0.08, 990.0, 15.0, 0.07, 45.0,
            0.60)}
CB, W = 0.005, 50.0
A_UP, B_LO = math.log(0.90 / 0.05), math.log(0.10 / 0.95)
ROWS = [  # site, id, budget/day, spend to date, buyers, upsells,
    #       refunds, mROAS or None
    ("D5D", "d-07", 60.0, 61.0, 0, 0, 0, None),
    ("D5D", "d-11", 80.0, 140.0, 22, 3, 1, 1.31),
    ("TMY", "m-03", 150.0, 420.0, 31, 5, 2, 1.18),
    ("TMY", "m-09", 100.0, 160.0, 2, 0, 1, None),
    ("TMY", "m-12", 300.0, 2600.0, 118, 9, 14, 0.74),
    ("TTW", "t-02", 900.0, 3400.0, 30, 3, 2, None),
    ("TTW", "t-05", 600.0, 520.0, 0, 0, 0, None)]


def funnel(site, buyers, ups, refunds):
    p1, rate, fix, tax, c1, u0, p2, c2, rf0, rep, kap = SITE[site]
    cbf = 300.0 if site == "TTW" else 15.0
    u = (ups + u0 * W) / (buyers + W)
    rf = (refunds + rf0 * W) / (buyers + W)

    def net(price, refund, cost):
        kept = price * (1 - refund - CB) * (1 - tax)
        return kept - (price * rate + fix) - CB * cbf - cost
    return math.fsum([net(p1, rf, c1), u * net(p2, 0.04, c2), rep]), kap


if __name__ == "__main__":
    print("  row       spend  buy     V  incCPA    llr  state   next/day")
    for site, cid, bud, spend, buy, ups, rfd, mroas in ROWS:
        v, kap = funnel(site, buy, ups, rfd)
        inc = buy * kap
        llr = inc * math.log(0.5) + 0.5 * spend / v
        if llr >= A_UP or (inc == 0 and spend >= 2.996 * v):
            state, nxt = "STOP", 0.0
        elif llr <= B_LO:
            state = "KEEP"
            step = 0.0 if mroas is None else 0.25 * (mroas - 1.0)
            nxt = bud * math.exp(max(-0.2, min(0.2, step)))
        else:
            state, nxt = "TEST", bud
        cpa = spend / inc if inc > 0 else float("inf")
        cpa_s = "{:>7.2f}".format(cpa) if inc > 0 else "    n/a"
        print("  {} {} {:>8,.0f} {:>4d} {:>6.2f} {} {:>6.2f}  {:<5} {:>9,.0f}"
              .format(site, cid, spend, buy, v, cpa_s, llr, state, nxt))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《测评与命理收费站买量精算、服务端转化回传校准与日入百单扩量母典》第 8 章「实证闭环:三站买量台账、蒸馏小工具与预注册证伪」(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-345 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026345,
  author    = {Fan, Yafu},
  title     = {Paid-Ad Scaling for Diagnostic Paywalls: Meta and TikTok Stopping Rules, Server-Side CAPI Calibration and Hundred-Order Daily Unit Economics},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论