5D Library › Diagnostic-to-Pedagogy Conversion Funnels, Mastery-Gated Curriculum Architecture and Certification Network Economics · 中文

Chapter 1. The Self-Cannibalization Bound: When Does a Course Pay?

The question: does teaching readers to read for themselves break your own rice bowl?

A reader who buys three readings a year stops buying once she can do them herself. Course revenue comes once; reading revenue keeps coming; the better the course sells, the less retail revenue is left. This is a textbook case of self-cannibalization. In product economics, a seller who launches a product that replaces its own older one gains only if the new product’s revenue, plus whatever revenue it leads to later, exceeds the revenue it replaces.

Evidence

Related published results. (1) When quality cannot be observed, low-quality supply erodes the market – Akerlof’s “market for lemons” (Akerlof 1970); a classical reason for practice licences is precisely to set a minimum quality standard that stops this erosion (Leland 1979). (2) Education and credentials can work as signals of ability even when they do not change it (Spence 1973). This chapter uses the two only as background: part of a course’s value comes from the licence it leads to, and the licence is worth something only if it really filters quality (Chapter 5).

One reader’s accounts

Let a reader buy r=3r = 3 readings a year at 39 each, with a gross margin of 0.85; she remains a customer each year with probability 0.70, cash is discounted at 10%, and the horizon is ten years. The course costs 480, with a margin of 0.90; after it she buys no more readings. The licence funnel: a share ee of graduates sit the exam, a share qq of them pass, and a licensee pays a licence fee of 1,200 a year (margin 0.70), renewing with probability 0.80.

Calculation

# Chapter 1. [p01] Self-cannibalization bound and the licence funnel.
# A reader buys r retail readings a year at price p with gross margin g,
# stays a customer each year with probability s, and cash is discounted
# at rate d over a 10-year horizon:
#     PV_retail = sum_{t<10} r p g (s / (1 + d))^t.
# A course at price P (margin gc) ends retail buying: the graduate now
# reads for himself. Without anything after the course, it pays only if
#     gc P > PV_retail          (the self-cannibalization bound).
# The licence funnel: a share e of graduates sit the exam, a share q of
# them pass, and a licensee pays L a year (margin gl), renewing with
# probability u. PV_after = e q PV_licence.

P_READ, R, G, S, D = 39.0, 3.0, 0.85, 0.70, 0.10
P_COURSE, GC = 480.0, 0.90
L, GL, U = 1200, 0.70, 0.80
YEARS = 10

def annuity(cash, keep):
    k = keep / (1 + D)
    return sum(cash * k ** t for t in range(YEARS))

if __name__ == "__main__":
    pv_retail = annuity(R * P_READ * G, S)
    pv_lic = annuity(L * GL, U)
    print("   PV of retail readings per reader : {:8.2f}".format(pv_retail))
    print("   course price that only breaks even: {:8.2f}".format(
        pv_retail / GC))
    print("   PV of one licence ({:,d} a year)  : {:8.2f}".format(L, pv_lic))
    print("   course LTV / retail LTV, P = {:.0f}:".format(P_COURSE))
    print("     exam take-up e   q=0.10   q=0.25   q=0.50")
    for e in (0.0, 0.2, 0.4, 0.6):
        row = []
        for q in (0.10, 0.25, 0.50):
            ltv = GC * P_COURSE + e * q * pv_lic
            row.append(ltv / pv_retail)
        print("     {:14.1f} {:8.2f} {:8.2f} {:8.2f}".format(e, *row))
    for e, q in ((0.1, 0.10), (0.4, 0.25)):
        need = (pv_retail - e * q * pv_lic) / GC
        print("   e={:.1f}, q={:.2f}: break-even course price {:8.2f}".format(
            e, q, need))
    print("\nReading: a 480 course is already worth 1.60 times the")
    print("retail stream it ends; the licence funnel adds a long tail,")
    print("and a strong funnel alone covers the readings it replaces")
    print("(a negative break-even price means the course could be free).")
    print("'Several times' needs both exam take-up and pass rate to be")
    print("high: at e = 0.4 and q = 0.25 the ratio is 2.69, not 5.")
#    PV of retail readings per reader :   270.51
#    course price that only breaks even:   300.57
#    PV of one licence (1,200 a year)  :  2952.50
#    course LTV / retail LTV, P = 480:
#      exam take-up e   q=0.10   q=0.25   q=0.50
#                 0.0     1.60     1.60     1.60
#                 0.2     1.82     2.14     2.69
#                 0.4     2.03     2.69     3.78
#                 0.6     2.25     3.23     4.87
#    e=0.1, q=0.10: break-even course price   267.76
#    e=0.4, q=0.25: break-even course price   -27.49
#
# Reading: a 480 course is already worth 1.60 times the
# retail stream it ends; the licence funnel adds a long tail,
# and a strong funnel alone covers the readings it replaces
# (a negative break-even price means the course could be free).
# 'Several times' needs both exam take-up and pass rate to be
# high: at e = 0.4 and q = 0.25 the ratio is 2.69, not 5.
Course LTV over retail LTV as pass rate and exam take-up vary.

A reader’s retail stream is worth 270.51, so, leaving the licence aside, a course breaks even at any price above 300.57. A course at 480 is by itself worth 1.60 times that stream. A licence is worth 2952.50, far more than retail readings; once a share of graduates go through the funnel, the course’s lifetime value rises quickly: 2.69 times at take-up 0.4 and pass rate 0.25, 4.87 times at take-up 0.6 and pass rate 0.50. With a strong funnel (e=0.4e = 0.4, q=0.25q = 0.25) the break-even course price becomes -27.49: licence income alone covers the retail revenue replaced, and the course could even be free. With a weak funnel (e=0.1e = 0.1, q=0.10q = 0.10) the break-even price is 267.76, close to the figure without the licence.

So “LTV multiplied several times” is not a property of the course but of the licence funnel, and it needs both take-up and pass rate to be high enough.

Who buys the course?

The accounts above treat every reader as the average reader. In reality those keenest to learn are those who buy the most readings. The program tests this with 20,000 simulated readers: readings per year are lognormal (median 1.5), and a reader’s value for the course rises with how many readings she buys.

Calculation

# Chapter 1. [p02] Who buys the course? Self-selection by heavy readers.
# 20,000 simulated readers differ in how many readings a year they buy:
# r is lognormal with median 1.5 and log-sd 0.9. A reader values the
# course at v = 60 + 110 r times a lognormal taste (log-sd 0.4) and buys
# it when v > P. Heavy readers value self-reading most, so the course
# selects exactly the readers whose retail stream it ends. Net gain of
# offering the course at P (retail margins and horizon as in p01; the
# funnel at e = 0.4, q = 0.25):
#     sum over buyers of (gc P + e q PV_licence - PV_retail(r)).
import math
import random

G, S, D, P_READ, GC = 0.85, 0.70, 0.10, 39.0, 0.90
YEARS, PV_LIC, E, Q = 10, 2952.50, 0.4, 0.25

def pv_retail(r):
    k = S / (1 + D)
    return sum(r * P_READ * G * k ** t for t in range(YEARS))

if __name__ == "__main__":
    rng = random.Random(74002)
    pop = []
    for _ in range(20000):
        r = 1.5 * math.exp(rng.gauss(0, 0.9))
        v = (60 + 110 * r) * math.exp(rng.gauss(0, 0.4))
        pop.append((r, v))
    mean_all = sum(pv_retail(r) for r, _ in pop) / len(pop)
    print("   mean retail PV, all readers: {:8.2f}".format(mean_all))
    print("      P buyers mean r buyers' PV  net/buyer  net/buyer  per 1000")
    print("                                 no funnel  w/ funnel   readers")
    for P in (180, 300, 480, 720, 960):
        b = [r for r, v in pop if v > P]
        pv_b = sum(pv_retail(r) for r in b) / len(b)
        n0 = GC * P - pv_b
        n1 = n0 + E * Q * PV_LIC
        sh = len(b) / len(pop)
        print("   {:4d} {:6.3f} {:6.2f} {:10.2f} {:10.2f} {:10.2f} {:9.0f}"
              .format(P, sh, sum(b) / len(b), pv_b, n0, n1, 1000 * sh * n1))
    print("\nReading: the readers who buy the course are the heaviest")
    print("retail buyers, so the retail stream it ends is well above")
    print("the average reader's. Priced on the average, the bound")
    print("of p01 is too lax; without the funnel the course loses money")
    print("on its buyers at every price up to 720. With the funnel, the")
    print("total per 1,000 readers is highest at the lowest price.")
#    mean retail PV, all readers:   203.38
#       P buyers mean r buyers' PV  net/buyer  net/buyer  per 1000
#                                  no funnel  w/ funnel   readers
#     180  0.636   3.06     275.82    -113.82     181.43    115370
#     300  0.370   4.08     368.32     -98.32     196.93     72807
#     480  0.184   5.54     499.25     -67.25     228.00     42010
#     720  0.085   7.45     671.83     -23.83     271.42     23017
#     960  0.045   9.25     833.88      30.12     325.37     14755
#
# Reading: the readers who buy the course are the heaviest
# retail buyers, so the retail stream it ends is well above
# the average reader's. Priced on the average, the bound
# of p01 is too lax; without the funnel the course loses money
# on its buyers at every price up to 720. With the funnel, the
# total per 1,000 readers is highest at the lowest price.

The average reader’s retail stream is worth 203.38, but course buyers sit far above the average: at a course price of 480 they buy 5.54 readings a year on average and their retail stream is worth 499.25; at a price of 180 it is worth 275.82. Without a licence funnel, the course loses money on its buyers at every price of 180, 300, 480 and 720 (at 480 it loses 67.25 per buyer); only at 960, where just 0.045 of readers still buy, does it make a small surplus of 30.12 on its buyers.

This is adverse self-selection: the course attracts precisely those whose retail revenue it will replace the most. Computed for the average reader, the bound of the first section is too lax. With the licence funnel, net gain per buyer turns positive everywhere, and the total per thousand readers is highest at the lowest price of 180 (115370), because in this model every graduate has the same chance of entering the funnel. That is an assumption of the model, not a finding: if heavy readers are also more likely to sit the exam and take the licence, the lowest price need not be best.

Hypothesis · not established

The author’s interpretation (not yet tested): in the author’s business, course buyers are mainly long-standing readers who have already bought many readings, so a course should be judged not as “new revenue” but as “retail revenue replaced plus the licence funnel”; a course without a licence funnel cannibalizes its seller. What would show this wrong: if real data (not this book’s parameters) showed that course buyers bought no more readings in the year before the course than other readers, or that their reading purchases did not fall afterwards, then adverse self-selection would be absent and the average bound of the first section would suffice.

Summary

A course by itself is worth only 1.60 times the retail revenue it ends; whether lifetime value rises several times depends on the take-up and pass rate of the licence funnel. Course buyers are heavy readers, whose retail stream is worth more than twice the average reader’s, so a course without a licence funnel loses money at most prices. And the funnel is worth as much as the licence is credible – the subject of Chapter 5.

References

Akerlof, George A. 1970. “The Market for ‘Lemons’: Quality Uncertainty and the Market Mechanism.” Quarterly Journal of Economics 84 (3): 488–500.
Leland, Hayne E. 1979. “Quacks, Lemons, and Licensing: A Theory of Minimum Quality Standards.” Journal of Political Economy 87 (6): 1328–46.
Spence, Michael. 1973. “Job Market Signaling.” Quarterly Journal of Economics 87 (3): 355–74.

Fan, Y. (2026). Diagnostic-to-Pedagogy Conversion Funnels, Mastery-Gated Curriculum Architecture and Certification Network Economics: Self-Cannibalization Bounds, Spaced-Retrieval Curves and Two-Sided Practitioner Licensing [诊学一体转化漏斗、掌握度门控课程与持牌人网络经济学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090197 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).