5D Library › Evolutionary Game Theory, Direct and Indirect Reciprocity and the Emergence of Cooperation · 中文

Chapter 8 Practice and a Pre-Registered Trial

Note

The checklist in this chapter is a thinking tool; it is not a management standard, legal opinion, business or investment advice, or moral instruction. The example is a fictional “Qiming Makers’ Co-operative”; all names, partners, rules and figures are invented solely to show the calculations and do not come from any real organisation, community or client. Drawing up real partnership agreements or community rules requires one’s own circumstances to be taken into account and, where necessary, advice from lawyers and other professionals.

Part 1 A four-step checklist for cooperation and protection against extortion

The models of the first six chapters can be organised into four questions. They help a partnership or a community check what mechanism its cooperation rests on, whether the conditions for that mechanism hold, and what will happen when things go wrong or someone free rides.

  1. The probability ww of future interaction, and identifying one-off deals. How likely is the relationship to continue for another round? Does that probability exceed the cost-to-benefit ratio c/bc/b (Chapters 2 and 3)? Which deals are in fact one-off, or have a known last round? Such deals cannot rely on reciprocity and need contracts, deposits or third-party escrow.
  2. Tolerance for errors under execution noise. Deliveries occasionally go wrong by accident. Is the agreed response “punish every error” (tit for tat), “forgive a certain share” (generous tit for tat, forgiving no more than 1−c/b1 – c/b), or “after an error both sides return to where they were” (win-stay, lose-shift, which needs b/c>2b/c > 2) (Chapter 3)? Could one side use the arrangement to extort: appearing to reciprocate while tilting the surplus towards itself by a fixed ratio?
  3. Auditing second-order reputation rules. Does the community’s public record or “flagging” rule distinguish “not helping without cause” from “justifiably not helping a rule-breaker”? Does helping a member already flagged as a rule-breaker count as a violation (Chapter 4)? Is flagging information public and consistent, or does each person judge privately?
  4. Ostrom’s graduated monitoring and exit mechanisms. Who has the right to take part, who monitors, and to whom are monitors accountable? Are sanctions graduated, do they distinguish accidental slips from deliberate violations, and how long before a record is cleared? Are there low-cost channels for appeal and mediation, and clear procedures for leaving and for expulsion (Chapter 6)?

Calculation

We apply the checklist to the fictional “Qiming Makers’ Co-operative”. Step 1: in its dealings with several kinds of partner, an honest delivery costs the deliverer c=2c = 2 and is worth b=5b = 5 to the other side, so the threshold is c/b=0.4c/b = 0.4. Compare the ww of four relationships. Step 2: deliveries go wrong by accident with probability 3 %; compare four response rules by average payoff over 40 quarters and the share of quarters in which both sides cooperate. Step 3: use the model of Chapter 4’s second program (85 % discriminators, 15 % free riders) to check three flagging rules the co-operative has considered. Step 4: compare, for three kinds of member, the probability of expulsion within three years under “three violations in total and you are out” and under “warning, then fine, then expulsion, with one strike cleared after 12 consecutive clean months”: a member in good standing who occasionally slips (2 % a month, always detected), an occasional shirker (10 % a month, detected with probability 60 %) and a regular free rider (50 % a month, detected with probability 60 %).

Calculation

# The four-step checklist of Chapter 8 applied to a FICTIONAL makers' co-operative and its
# partners. All names, numbers and rules are invented to show the calculations.
import importlib.util, io, contextlib, pathlib, random
HERE = pathlib.Path(__file__).resolve().parent
def load(name):
    spec = importlib.util.spec_from_file_location(name, HERE / f"{name}.py")
    m = importlib.util.module_from_spec(spec)
    with contextlib.redirect_stdout(io.StringIO()):
        spec.loader.exec_module(m)
    return m
p06, p09 = load("p06_markov_noise"), load("p09_norm_dynamics")
rng = random.Random(13)

# Step 1. Shadow of the future: honest delivery costs c = 2 and is worth b = 5 to the
# other side; reciprocity can carry a relationship only if w = P(another round) > c/b =
# 0.4.
print("step 1: w vs c/b = 0.40")
for name, w in (("raw-material supplier (renews 9 times in 10)", 0.9),
                ("shared market stall (half of seasons renewed)", 0.5),
                ("one-off bulk buyer", 0.1),
                ("fixed 3-year contract, last year", 0.0)):
    print(f"   {name:46s} w = {w:.1f}: " + ("reciprocity can carry it" if w > 0.4
          else "needs contract or deposit"))

# Step 2. Noise: 3 % of deliveries go wrong by accident. 40 quarters, both start
# cooperating.
def finite(p, q, eps, rounds, D):
    tot = cc = 0.0
    p, q = p06.flip(p, eps), p06.flip(q, eps); qs = (q[0], q[2], q[1], q[3])
    M = [[a * b, a * (1 - b), (1 - a) * b, (1 - a) * (1 - b)] for a, b in zip(p, qs)]
    v = [(1 - eps) ** 2, (1 - eps) * eps, eps * (1 - eps), eps * eps]
    for _ in range(rounds):
        tot += sum(x * y for x, y in zip(v, D)); cc += v[0]
        v = [sum(v[i] * M[i][j] for i in range(4)) for j in range(4)]
    return tot / rounds, cc / rounds
D = (3.0, -2.0, 5.0, 0.0)                     # R = b - c, S = -c, T = b, P = 0
print("step 2: 3 % accidental failures, 40 quarters; payoff per quarter, share CC")
for name, st in (("tit for tat", (1, 0, 1, 0)),
                 ("forgive 1 in 2 (q = 0.5)", (1, .5, 1, .5)),
                 ("forgive at the bound q = 1 - c/b = 0.6", (1, .6, 1, .6)),
                 ("win-stay, lose-shift", (1, 0, 0, 1))):
    pay, cc = finite(st, st, 0.03, 40, D)
    print(f"   {name:38s} {pay:.2f}   {cc:.2f}")

# Step 3. Second-order reputation audit of the co-op's flagging rules, with the p09 model
# (85 % discriminators, 15 % free riders, errors as in p08).
RULES = {"A: refusing anyone's request earns a flag": "image scoring",
         "B: as A, but refusing a flagged member is exempt": "simple standing",
         "C: as B, and helping a flagged member earns a flag": "stern judging"}
print("step 3: rule | justified refusal stays Good? | helping a flagged member is Bad?"
      " |\n        share of discriminators Good")
for rule, norm in RULES.items():
    nm = p09.NORMS[norm]
    gd, _, _ = p09.reputations([0.85, 0.0, 0.15], nm)
    print(f"   {rule:50s} {'yes' if nm[('D', 'B')] else 'no ':3s}  "
          f"{'yes' if not nm[('C', 'B')] else 'no ':3s}  {gd:.2f}")

# Step 4. Graduated sanctions with forgiveness vs "three strikes, ever". An honest member
# slips (misses a duty) with prob. 0.02 per month; a free rider skips with prob. 0.5 and a
# skip is detected with prob. 0.6. Rule X: the 3rd detected lapse ever means exit.
# Rule Y: warning, then fine, then exit, but 12 clean months wipe one strike.
def p_exit(lapse, detect, rule, months=36, runs=20000):
    out = 0
    for _ in range(runs):
        strikes, clean = 0, 0
        for _ in range(months):
            if rng.random() < lapse and rng.random() < detect:
                strikes += 1; clean = 0
            else:
                clean += 1
                if rule == "Y" and clean == 12 and strikes: strikes -= 1; clean = 0
            if strikes >= 3: out += 1; break
    return out / runs
print("step 4: probability of exit within 3 years")
for who, lapse, det in (("honest member, slips 2 %/month", 0.02, 1.0),
                        ("occasional shirker, 10 %/month", 0.10, 0.6),
                        ("free rider, 50 %/month", 0.50, 0.6)):
    print(f"   {who:32s} three strikes ever {p_exit(lapse, det, 'X'):.3f};"
          f" graduated {p_exit(lapse, det, 'Y'):.3f}")
# step 1: w vs c/b = 0.40
#    raw-material supplier (renews 9 times in 10)   w = 0.9: reciprocity can carry it
#    shared market stall (half of seasons renewed)  w = 0.5: reciprocity can carry it
#    one-off bulk buyer                             w = 0.1: needs contract or deposit
#    fixed 3-year contract, last year               w = 0.0: needs contract or deposit
# step 2: 3 % accidental failures, 40 quarters; payoff per quarter, share CC
#    tit for tat                            2.04   0.48
#    forgive 1 in 2 (q = 0.5)               2.83   0.89
#    forgive at the bound q = 1 - c/b = 0.6 2.86   0.91
#    win-stay, lose-shift                   2.75   0.89
# step 3: rule | justified refusal stays Good? | helping a flagged member is Bad? |
#         share of discriminators Good
#    A: refusing anyone's request earns a flag          no   no   0.06
#    B: as A, but refusing a flagged member is exempt   yes  no   0.97
#    C: as B, and helping a flagged member earns a flag yes  yes  0.97
# step 4: probability of exit within 3 years
#    honest member, slips 2 %/month   three strikes ever 0.035; graduated 0.019
#    occasional shirker, 10 %/month   three strikes ever 0.367; graduated 0.251
#    free rider, 50 %/month           three strikes ever 1.000; graduated 0.998

Step 1: the raw-material supplier renews with probability 0.9 each round and the shared stall with 0.5, both above 0.4, so reciprocity can carry both relationships; the one-off bulk buyer (0.1) and a fixed contract in its known last year (0) cannot rely on reciprocity and need other mechanisms such as contracts, deposits or escrow. A known last round deserves particular attention: in theory, in a finitely repeated prisoner’s dilemma with a known end, cooperation unravels round by round when one reasons backwards from the last round.

Step 2: with 3 % accidental failures tit for tat averages only 2.04, and both sides cooperate in only 48 % of quarters: a cycle of retaliation set off by one slip lasts until the next slip breaks it. Forgiving half of the defections (q=0.5q = 0.5) raises the payoff to 2.83 and the share of cooperative quarters to 89 %; forgiving up to the limit q=0.6q = 0.6 gives 2.86 and 91 %, only a little more, but closer to the boundary at which a one-off defector can exploit the rule. Win-stay, lose-shift gives 2.75 and 89 %; here b/c=2.5>2b/c = 2.5 > 2, so its condition holds.

Step 3: rule A (“refusing anyone’s request earns a flag”) is image scoring; with 15 % free riders only 6 % of discriminators keep a clean record and cooperation collapses. Rule B (“refusing a flagged member’s request earns no flag”) is simple standing, and rule C (adding “helping a flagged member earns a flag”) is stern judging; both keep 97 % of discriminators in good standing. B and C differ in whether unconditional help to rule-breakers is penalised; Chapter 4 showed that C prevents unconditional cooperators from increasing, but is harsher on those flagged by mistake. Which to choose depends on how much misjudgement the community will tolerate, and should be stated publicly in advance.

Step 4: under “three violations in total and you are out”, a member in good standing who occasionally slips has a 3.5 % chance of expulsion within three years, against 1.9 % under the graduated rule with periodic clearing; for the occasional shirker the figures are 36.7 % and 25.1 %; the regular free rider is almost certain to be expelled under both rules (1.000 and 0.998). In this fictional example graduated sanctions with clearing roughly halve the chance that a member in good standing is hurt by mistake, while deterrence against the regular free rider is almost unchanged; the price is that more occasional shirkers stay. This trade-off should be weighed and made public in advance. All parameters are illustrative.

Part 2 A pre-registered randomised controlled trial

Question

In a multi-round online public goods game, does showing participants their partners’ second-order reputation labels (which distinguish “not contributing towards cooperators” from “justifiably not contributing towards defectors”), compared with showing only their partners’ first-order record of past contribution rates, raise the maintenance of cooperation in the group?

Evidence

Existing research. Experiments on indirect reciprocity show that people decide whether to help others according to what those others have given in the past (Wedekind and Milinski 2000), and that combining reputation with a public goods game can sustain high contributions (Milinski et al. 2002). The models of Chapter 4 predict that when the reputation rule distinguishes justified refusals, sanctioners do not lose reputation by sanctioning, and cooperation is easier to sustain (Ohtsuki and Iwasa 2006; Panchanathan and Boyd 2003). But whether assessments are public and consistent is decisive for the models’ conclusions (Hilbe et al. 2018). This trial tests a specific version of that prediction in a human public goods game.

Design

A contribution to a public good goes to the whole group and cannot be aimed at one person, so “justifiably not contributing towards defectors” needs an operational definition. This protocol splits each community at random into groups of three every round: withholding counts as justified only when both other members of the group are labelled “bad” at the start of the round.

Element Specification
Design Two-arm parallel cluster-randomised controlled trial with online communities of 12 as the unit of randomisation, stratified by session batch, 1:1 allocation; reporting follows the CONSORT extension for cluster randomised trials (Campbell et al. 2012)
Participants Adults aged 18 or over recruited through open calls at universities and research institutions or through academic online participant pools, paid according to their earnings as is customary in experimental economics; no recruitment from the author’s community, students or customers (including clients of the author’s businesses)
Exclusions Previous participation in public goods or reputation-game experiments; participation in this study’s pilot; failing the rule-comprehension test (taken before randomisation)
Task 30 rounds. Each round the community is split at random into 4 groups of 3; each person has 10 tokens and chooses either to invest all of them or to keep all of them; the group pool is doubled and shared equally among the 3. The rules are explained in writing and with examples beforehand, identically in both arms
Arm A (second-order labels) Before each decision, the labels “good” or “bad” of the two group partners are shown: contributors are good; withholding when at least one partner is good is bad; withholding when both other members are bad is good (justified). Everyone is good in round 1; after that, a person’s label depends only on his or her own decision in the previous round and on the two partners’ labels at that time. Labels are computed by the program according to this rule, whose full text is shown to participants
Arm B (first-order record) Before each decision, the two group partners’ contribution rates so far (percentage) are shown
Both arms After each round, the group’s contributions and one’s own earnings are shown; no information on other groups; no punishment option; no deception
Blinding Participants know their own screen but not the research hypothesis or the other arm’s screen; the primary outcome is computed automatically; analysts do not know the arm codes until the analysis plan is locked
Primary outcome Community cooperation maintenance: the share of “invest” among all 120 decisions of the community in rounds 21–30. One value per community
Secondary outcomes Share invested over all 30 rounds; slope of the share invested across rounds; in arm A, the share withholding when both partners are “bad” (justified withholding); earnings; rule comprehension and perceived fairness (open answers coded by raters blind to arm)
Analysis Community-level linear regression: primary outcome ~ arm + batch stratum, reporting the difference between arms and its 95 % confidence interval; two-sided α=0.05\alpha = 0.05

There is only one primary outcome.

Arm A uses only this rule, fixed in advance: simple standing (Chapter 4) with the strictest justification. Withholding from someone in bad standing is good, from someone in good standing bad, and contributing to someone in bad standing costs nothing. It is neither image scoring, which cannot tell justified from unjustified withholding, nor stern judging.

Sample size

Calculation

# Sample size for the Chapter 8 cluster-randomised trial. Unit of randomisation: an
# online COMMUNITY of 12 people who play 30 rounds of a public goods game in re-matched
# groups of 3. Primary outcome, one number per community: the share of "contribute"
# decisions in rounds 21-30. The analysis is at community level, so the sample size
# needs the SD of community means (sigma_g), which depends on the intra-cluster
# correlation (ICC) of individual rates. All parameter values are assumptions, to be
# checked against pilot data.
# Arm A labels: one fixed rule, simple standing with strict justification (see p15).
import math, random
from statistics import NormalDist, mean, pstdev
rng = random.Random(14)
z = NormalDist().inv_cdf
ALPHA, POWER, M = 0.05, 0.90, 12

def n_per_arm(sd, delta):
    return math.ceil(2 * (z(1 - ALPHA / 2) + z(POWER)) ** 2 * sd ** 2 / delta ** 2)

print("communities per arm (alpha 0.05 two-sided, 90 % power)")
print("   SD of community means   delta 0.08   delta 0.10   delta 0.12")
for sd in (0.12, 0.15, 0.18):
    print(f"           {sd:.2f}              " + "         ".join(
        f"{n_per_arm(sd, d):3d}" for d in (0.08, 0.10, 0.12)))

# A generative model for individuals: person j in community g contributes in each of the
# 10 scored rounds with probability p = base + u_g + e_j (clipped to 0..1), where
# u_g ~ N(0, 0.13^2) is shared by the community and e_j ~ N(0, 0.25^2) is personal.
def community(base, tau=0.13, s_e=0.25):
    u = rng.gauss(0, tau); rates = []
    for _ in range(M):
        p = min(1.0, max(0.0, base + u + rng.gauss(0, s_e)))
        rates.append(sum(rng.random() < p for _ in range(10)) / 10)
    return rates
cs = [community(0.40) for _ in range(3000)]
means = [mean(c) for c in cs]
sd_g = pstdev(means)
within = mean(pstdev(c) ** 2 * M / (M - 1) for c in cs)
icc = (sd_g ** 2 - within / M) / (sd_g ** 2 - within / M + within)
print(f"model: SD of community means {sd_g:.3f}; SD of individual rates"
      f" {math.sqrt(within + sd_g ** 2 - within / M):.3f}; ICC {icc:.2f}")
print(f"   design effect if analysed per person: 1 + (m - 1) ICC ="
      f" {1 + (M - 1) * icc:.2f}")

SD, DELTA = 0.15, 0.10
n = n_per_arm(SD, DELTA); rec = math.ceil(n / 0.9)
print(f"chosen: SD 0.15, delta 0.10 (control 0.40 -> 0.50): {n} communities per arm"
      f" analysable;")
print(f"   allowing 10 % incomplete communities, recruit {rec} per arm =\n  "
      f" {rec * M} people per arm, {2 * rec * M} in all")

# What the decision rule (p15) will say if the true effect equals the smallest effect of
# interest (0.10): 95 % CI of the difference from a two-sample t-type interval.
def trial(true_delta, n):
    a = [rng.gauss(0.40, SD) for _ in range(n)]; b = [rng.gauss(0.40 + true_delta, SD)
                                                       for _ in range(n)]
    d = mean(b) - mean(a)
    se = math.sqrt((pstdev(a) ** 2 + pstdev(b) ** 2) * n / (n - 1) / n)
    return d - 1.96 * se, d + 1.96 * se
for td in (0.10, 0.15, 0.20):
    res = [trial(td, n) for _ in range(4000)]
    benefit = sum(lo > 0 for lo, hi in res) / 4000
    strong = sum(lo >= 0.10 for lo, hi in res) / 4000
    print(f"true effect {td:.2f}: some benefit shown {benefit:.3f};"
          f" 'at least 0.10' shown {strong:.3f}")
# communities per arm (alpha 0.05 two-sided, 90 % power)
#    SD of community means   delta 0.08   delta 0.10   delta 0.12
#            0.12               48          31          22
#            0.15               74          48          33
#            0.18              107          69          48
# model: SD of community means 0.138; SD of individual rates 0.291; ICC 0.15
#    design effect if analysed per person: 1 + (m - 1) ICC = 2.70
# chosen: SD 0.15, delta 0.10 (control 0.40 -> 0.50): 48 communities per arm analysable;
#    allowing 10 % incomplete communities, recruit 54 per arm =
#    648 people per arm, 1296 in all
# true effect 0.10: some benefit shown 0.905; 'at least 0.10' shown 0.027
# true effect 0.15: some benefit shown 1.000; 'at least 0.10' shown 0.383
# true effect 0.20: some benefit shown 1.000; 'at least 0.10' shown 0.902

The primary outcome is defined and analysed at community level, with each community contributing a single value, so the analysis needs no intra-cluster correlation coefficient (ICC) or design effect; what the sample-size calculation needs is the standard deviation of community means. That standard deviation is related to the ICC, and the program illustrates the relation with an illustrative model of individuals: with a shared community fluctuation of SD 0.13 and individual differences of SD 0.25, the SD of community means is 0.138 and the ICC of individual contribution rates is about 0.15; if the analysis were done per person instead, the design effect would be 1+(m−1)ICC=2.701 + (m – 1)\,\mathrm{ICC} = 2.70, that is, 2.7 times as many people would be needed as when the cluster structure is ignored.

With an SD of community means of 0.15 (slightly more conservative than the illustrative model above), a smallest difference of interest of 0.10 (control 0.40, arm A 0.50), two-sided α=0.05\alpha = 0.05 and 90 % power, 48 analysable communities are needed per arm (normal approximation; a t-distribution correction adds about one); allowing for about 10 % of communities being incomplete because someone drops out, 54 communities are recruited per arm, that is 648 people per arm and 1,296 in all. With an SD of 0.18, 69 communities per arm would be needed; with a smallest difference of interest of 0.08, 74. All these parameters are assumptions and should be checked against pilot data before the trial begins.

  • What “90 % power” means. The last three lines of the program show that when the true effect is exactly the smallest difference of interest, 0.10, the decision rule reaches some “benefit” conclusion (lower confidence limit above 0) with probability about 90 % (0.905 in the simulation), but reaches the strongest conclusion, “an increase of at least 0.10” (lower confidence limit at least 0.10), with probability only about 2.5 % (0.027 in the simulation). With a true effect of 0.15 the strongest conclusion has probability 38 %, and with 0.20 it has 90 %. This agrees with the explanation in Volume 33: 90 % is the power to reach some benefit conclusion, not the power to reach the strongest conclusion.

Decision rule

Calculation

# Decision rule for the Chapter 8 trial, fixed before any data are collected.
# diff = (share of contribute decisions in rounds 21-30, second-order label arm) minus
# (same, first-order record arm), one value per community; 95 % CI from the
# pre-specified community-level linear model (arm + batch stratum). Smallest difference
# of interest 0.10.
# Arm A labels follow ONE rule fixed in advance: simple standing, strict justification.
# Contributing is always Good; withholding is Good only if BOTH group partners were Bad
# at the start of the round, otherwise Bad. Everyone is Good in round 1. This is not
# image scoring (no justified withholding) and not stern judging (helping Bad is fine).
# Fidelity: in the label arm, labels must have been computed and shown as specified in at
# least 95 % of community-rounds; label comprehension check passed by >= 80 % of people.
SESOI = 0.10

def verdict(diff, lo, hi, fidelity, understood):
    if fidelity < 0.95 or understood < 0.80:
        return (f"diff {diff:+.2f}: NOT INTERPRETABLE - labels as specified"
                f" {fidelity:.0%}, understood by {understood:.0%}")
    if lo >= SESOI:              v = "labels raise cooperation by at least 0.10"
    elif lo > 0:                 v = "labels raise cooperation; 0.10 or more not shown"
    elif hi < 0:                 v = "labels LOWER cooperation"
    elif lo > -SESOI and hi < SESOI: v = "no difference worth acting on"
    else:                        v = "unclear: the interval is too wide to decide"
    return f"diff {diff:+.2f} ({lo:+.2f} to {hi:+.2f}): {v}"

for case in ((0.17, 0.11, 0.23, 0.98, 0.91), (0.09, 0.03, 0.15, 0.97, 0.88),
             (0.01, -0.05, 0.07, 0.99, 0.90), (-0.08, -0.14, -0.02, 0.96, 0.86),
             (0.15, 0.09, 0.21, 0.91, 0.90), (0.04, -0.07, 0.15, 0.97, 0.85),
             (0.12, 0.06, 0.18, 0.98, 0.71)):
    print(verdict(*case))
# diff +0.17 (+0.11 to +0.23): labels raise cooperation by at least 0.10
# diff +0.09 (+0.03 to +0.15): labels raise cooperation; 0.10 or more not shown
# diff +0.01 (-0.05 to +0.07): no difference worth acting on
# diff -0.08 (-0.14 to -0.02): labels LOWER cooperation
# diff +0.15: NOT INTERPRETABLE - labels as specified 91%, understood by 90%
# diff +0.04 (-0.07 to +0.15): unclear: the interval is too wide to decide
# diff +0.12: NOT INTERPRETABLE - labels as specified 98%, understood by 71%

The example results are invented. If labels in arm A were not computed and displayed as specified in more than 5 % of community-rounds, or fewer than 80 % of participants passed the label-comprehension test, the report will state “not interpretable” whatever the result. Only if the lower limit of the 95 % confidence interval for the difference is at least 0.10 will the report say “labels raised cooperation by at least the smallest amount of interest”; if the lower limit is above 0 but below 0.10, it will say “labels raised cooperation, but an increase of 0.10 was not shown”. Secondary outcomes are reported only as estimates with intervals and are not used to overturn the main conclusion.

What this trial cannot show

First, arms A and B differ in more than “first order versus second order”: arm A sees a binary label computed by a rule, arm B a continuous percentage, and the simplicity of a label may itself affect behaviour. A positive result cannot tell whether “second-order information” or “the label format” was responsible. The protocol has only two arms and no arm with “first-order binary labels” (contributing is good, withholding is bad), so that distinction is left to later trials. Second, “justified withholding” has a single rule fixed in advance (simple standing with strict justification: both other members bad); the conclusions apply only to that rule and cannot be extended to other assessment rules, such as stern judging in Chapter 4. Third, this is a 30-round online experiment with communities of 12, no punishment, no communication and labels computed impartially by a program; reputations in real communities are assessed by people privately and incompletely (the note in Chapter 4), and the conclusions cannot be carried over directly to any real community’s rules or any platform’s design. Fourth, the participants will mostly be students or members of online participant pools. Fifth, only one set of payoffs is used (the pool is doubled). Sixth, the primary outcome looks only at rounds 21–30, and another time window might give different conclusions, so the secondary outcomes will report the curve over all rounds.

Ethics and conflicts of interest

This is a low-risk study: the task is an investment game on a computer, involving no health intervention and collecting no health information; the foreseeable discomfort is only displeasure at being labelled “bad” or at earning less. Even so, the trial must be reviewed and approved by an independent ethics committee, every participant must give informed consent, and the protocol and analysis plan must be registered before the first community is enrolled (Nosek et al. 2018). The labelling rule is fully disclosed to participants and no deception is used; not disclosing the research hypothesis is withholding of information, and the purpose of the study is explained to everyone after the task. Labels are shown only within the experiment and are not linked to participants’ real identities; the payment rules are announced in advance; anyone may withdraw at any time without giving a reason and is paid for the part completed, and the community is then recorded as incomplete. Data are stored anonymously. The results will be published whether they are positive or negative.

Disclosure

The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base, and therefore has a commercial interest in the subject of this book. The author also runs communities and founded and runs Asia Influencer X and Yafu Global, whose business includes community operations and influencer and brand partnerships, and therefore has a direct commercial interest in the subjects of cooperation, reputation and community governance discussed in this book. The book discusses mathematical models of evolutionary games; it is not moral instruction, promotes no course, card deck or service, and is not investment or business advice; it uses no client or community data from these communities and businesses; the Chapter 8 trial is designed to be run and analysed by researchers independent of the author’s business and does not recruit from the author’s community, students or customers, including the clients of these businesses; all the code and the study protocol are published openly for anyone to check.

References

Campbell, Marion K., Gilda Piaggio, Diana R. Elbourne, and Douglas G. Altman. 2012. “CONSORT 2010 Statement: Extension to Cluster Randomised Trials.” BMJ 345: e5661. https://doi.org/10.1136/bmj.e5661.
Hilbe, Christian, Laura Schmid, Josef Tkadlec, Krishnendu Chatterjee, and Martin A. Nowak. 2018. “Indirect Reciprocity with Private, Noisy, and Incomplete Information.” Proceedings of the National Academy of Sciences 115 (48): 12241–46. https://doi.org/10.1073/pnas.1810565115.
Milinski, Manfred, Dirk Semmann, and Hans-Jürgen Krambeck. 2002. “Reputation Helps Solve the ‘Tragedy of the Commons’.” Nature 415 (6870): 424–26. https://doi.org/10.1038/415424a.
Nosek, Brian A., Charles R. Ebersole, Alexander C. DeHaven, and David T. Mellor. 2018. “The Preregistration Revolution.” Proceedings of the National Academy of Sciences 115 (11): 2600–2606. https://doi.org/10.1073/pnas.1708274114.
Ohtsuki, Hisashi, and Yoh Iwasa. 2006. “The Leading Eight: Social Norms That Can Maintain Cooperation by Indirect Reciprocity.” Journal of Theoretical Biology 239 (4): 435–44. https://doi.org/10.1016/j.jtbi.2005.08.008.
Panchanathan, Karthik, and Robert Boyd. 2003. “A Tale of Two Defectors: The Importance of Standing for Evolution of Indirect Reciprocity.” Journal of Theoretical Biology 224 (1): 115–26. https://doi.org/10.1016/S0022-5193(03)00154-1.
Wedekind, Claus, and Manfred Milinski. 2000. “Cooperation Through Image Scoring in Humans.” Science 288 (5467): 850–52. https://doi.org/10.1126/science.288.5467.850.

Fan, Y. (2026). Evolutionary Game Theory, Direct and Indirect Reciprocity and the Emergence of Cooperation: From Hamilton's Rule and Axelrod's Tournaments to Reputation and Costly Punishment, Read Beside Classical Confucian and Mohist Ethics [演化博弈、直接与间接互惠及合作涌现的数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23788551 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).