What does Star Ratings, Self-Selection and Bayesian Ranking conclude, in numbers, about “Classic Texts East and West: Liu Yi on the Nine-Rank System and Galton's "Vox Populi"”?
Answer: Chapter 7 examines Liu Yi's critique of the Nine-Rank system in the Book of Jin, "no humble families above, no powerful clans below" (with family weighted about three times talent, humble men hold 0.072 of the upper grades in the model), and Galton's "Vox Populi" of 1907 (the crowd mean rebuilt as a two-piece normal is 1197.5 lb).
Scope of the numbers: Chapter 1 explains the J-shaped star distribution with two filters, a purchasing bias and an under-reporting bias; with illustrative parameters the purchasing bias lifts the mean by +0.532 stars. Marked boxes separate published evidence, runnable Python calculations (21 programs), classical sources and the author's hypotheses.
Data table
The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.072
| A. fixed quotas, upper grades = 0.10 of candidates | |||
| w | commoner up | clan in low | rho(T) |
| 0.0 | 0.843 | 0.4532 | 0.861 |
| 1.0 | 0.589 | 0.1749 | 0.817 |
| 2.0 | 0.281 | 0.0425 | 0.730 |
| 3.0 | 0.072 | 0.0030 | 0.668 |
| B. w = 2, the upper quota grows (grade inflation) | |||
| upper | commoner up | rho(T) talent of up | |
| 0.10 | 0.281 | 0.730 | 1.088 |
| 0.25 | 0.526 | 0.736 | 0.864 |
| 0.40 | 0.662 | 0.736 | 0.717 |
| 0.55 | 0.740 | 0.720 | 0.560 |
Show the full program
# Chapter 7. p71: when the grader weighs family, the top grades fill up.
# Toy Nine-Rank model (illustrative, not a reconstruction of Jin data):
# 20000 candidates, talent T ~ N(0, 1); 15 percent belong to powerful
# clans (S = 1). The rectifier's score is T + w S + e, e ~ N(0, 0.5).
# Candidates are cut into nine grades by fixed quotas, grade 1 highest.
# "Upper" = grades 1-3, "lower" = grades 7-9. Reported: share of commoner
# (S = 0) candidates in the upper grades, share of clan candidates in the
# lower grades, and the rank correlation of grade with talent; in part
# two also the mean talent (in sd units) of those graded upper.
# Part two keeps w = 2 and lets the upper quota grow (grade inflation).
import math
import random
N, CLAN = 20000, 0.15
rng = random.Random(20261112)
PEOPLE = [(rng.gauss(0, 1), 1 if rng.random() < CLAN else 0,
rng.gauss(0, 0.5)) for _ in range(N)]
def grades(w, quotas):
order = sorted(range(N), key=lambda i: -(PEOPLE[i][0]
+ w * PEOPLE[i][1]
+ PEOPLE[i][2]))
g = [0] * N
cuts, acc = [], 0.0
for q in quotas:
acc += q # running float total of nine quota shares
cuts.append(int(round(acc * N)))
k = 0
for pos, i in enumerate(order):
while k < 8 and pos >= cuts[k]:
k += 1
g[i] = k + 1
return g
def spearman(xs, ys):
def ranks(v):
o = sorted(range(len(v)), key=lambda i: (v[i], i))
r = [0.0] * len(v)
for pos, i in enumerate(o):
r[i] = float(pos)
return r
a, b = ranks(xs), ranks(ys)
ma, mb = math.fsum(a) / N, math.fsum(b) / N
cov = math.fsum((x - ma) * (y - mb) for x, y in zip(a, b))
va = math.fsum((x - ma) ** 2 for x in a)
vb = math.fsum((y - mb) ** 2 for y in b)
return cov / math.sqrt(va * vb)
def report(w, quotas):
g = grades(w, quotas)
up = [i for i in range(N) if g[i] <= 3]
low = [i for i in range(N) if g[i] >= 7]
com_up = len([i for i in up if PEOPLE[i][1] == 0]) / len(up)
clan = [i for i in range(N) if PEOPLE[i][1] == 1]
clan_low = len([i for i in clan if g[i] >= 7]) / len(clan)
rho = -spearman([float(x) for x in g], [p[0] for p in PEOPLE])
t_up = math.fsum(PEOPLE[i][0] for i in up) / len(up)
return len(up) / N, com_up, clan_low, rho, t_up
BASE = (0.01, 0.03, 0.06, 0.10, 0.15, 0.20, 0.20, 0.15, 0.10)
print("A. fixed quotas, upper grades = %.2f of candidates" % 0.10)
print("%5s %12s %12s %10s" % ("w", "commoner up", "clan in low",
"rho(T)"))
for w in (0.0, 1.0, 2.0, 3.0):
_, cu, cl, rho, _ = report(w, BASE)
print("%5.1f %12.3f %12.4f %10.3f" % (w, cu, cl, rho))
print("B. w = 2, the upper quota grows (grade inflation)")
print("%6s %12s %10s %12s" % ("upper", "commoner up", "rho(T)",
"talent of up"))
for up in (0.10, 0.25, 0.40, 0.55):
rest = 1.0 - up
q = (up * 0.1, up * 0.3, up * 0.6) + tuple(
rest * x / 0.90 for x in BASE[3:])
share, cu, _, rho, t_up = report(2.0, q)
print("%6.2f %12.3f %10.3f %12.3f" % (share, cu, rho, t_up))Single source and citation
Fan, Y. (2026). J-Shaped Self-Selection in Star Ratings, Bayesian-Average Ranking and Rating Inflation from the Nine-Rank System to Online Marketplaces (Version 1.0.0), Chapter 7: Classic Texts East and West: Liu Yi on the Nine-Rank System and Galton's "Vox Populi". 5D Library. DOI pending
DOI pending · monograph page goes live after publication · 5DL-B-212 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026212,
author = {Fan, Yafu},
title = {J-Shaped Self-Selection in Star Ratings, Bayesian-Average Ranking and Rating Inflation from the Nine-Rank System to Online Marketplaces},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}Other chapter answers in this book
- Chapter 1 · Where the J Comes From: a U-Shaped Urge to Post and Two Self-Selection Biases
- Chapter 2 · Five Stars Do Not Sell Best: the Spiegel Curve at Source and a Small-Sample Reading
- Chapter 3 · The Wilson Lower Bound: Why Three Five-Star Reviews Rank Below Two Hundred at 4.6
- Chapter 4 · The Dirichlet-Prior Bayesian Average: Pseudo-Reviews Borrowed from the Shop
- Chapter 5 · Sorting Is Soliciting: the Shelf Feedback Loop, the Cost of Exploring and the First Five Reviews
- Chapter 6 · Rating Inflation: Norm Chasing, Information Decay and the Friction of Asking
- Chapter 8 · Closing the Loop: Shop Sorting, the Timing of Review Requests and a Falsification Plan