关于「分级:分几级、切在哪里,以及分数膨胀为什么先甜后苦」,《达人认证的信号与分级》给出的定量结论是什么?
结论:级数由打分精度决定。p31 说明,误差标准差 0.3 时最好分三级,0.5 时两级,1.0 时只能有一个级别;p32 说明,申请成本达到 0.241 以后两级方案胜过精确分数,达到 0.315 以后连只分通过与否都胜过精确分数;p33 说明,分数膨胀只在第二季度略占上风,九个季度的溢价总量从 305,243 降到 134,913。
数据口径:书中的达人人数、品牌数、打分误差、成本、溢价与审计参数均为示意值,全部计算由书中程序生成;本书是非个人化经营建议,不构成法律、税务或投资建议;书中的认证标准公开可测,费用与是否通过无关,不暗示任何政府或法定资格,可以经审计撤销,不制造虚假稀缺,推荐奖励只限单层、只给直接推荐人,不涉及任何多层分润。 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。
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 标准库与结论吻合的打印值: 0.3, 0.5, 1.0
| score noise sigma | 0.3, pass rate 0.20 |
| K | 1 R^2 0.000 misgrade 0.000 net 0.0 |
| K | 2 R^2 0.477 misgrade 0.127 net 39.2 share of holders above each cut 0.28 |
| K | 3 R^2 0.594 misgrade 0.243 net 42.8 share of holders above each cut 0.46 0.12 |
| K | 4 R^2 0.640 misgrade 0.342 net 40.0 share of holders above each cut 0.56 0.24 0.06 |
| K | 5 R^2 0.663 misgrade 0.423 net 35.5 share of holders above each cut 0.63 0.34 0.15 0.04 |
| best number of holder levels | 3 |
| score noise sigma | 0.5, pass rate 0.20 |
| K | 1 R^2 0.000 misgrade 0.000 net 0.0 |
| K | 2 R^2 0.314 misgrade 0.187 net 20.8 share of holders above each cut 0.28 |
| K | 3 R^2 0.391 misgrade 0.339 net 19.2 share of holders above each cut 0.45 0.12 |
| K | 4 R^2 0.421 misgrade 0.449 net 14.4 share of holders above each cut 0.56 0.24 0.06 |
| K | 5 R^2 0.436 misgrade 0.530 net 9.1 share of holders above each cut 0.63 0.34 0.15 0.04 |
| best number of holder levels | 2 |
| score noise sigma | 1.0, pass rate 0.20 |
| K | 1 R^2 0.000 misgrade 0.000 net 0.0 |
| K | 2 R^2 0.121 misgrade 0.260 net -1.0 share of holders above each cut 0.28 |
| K | 3 R^2 0.150 misgrade 0.431 net -8.1 share of holders above each cut 0.45 0.12 |
| K | 4 R^2 0.162 misgrade 0.538 net -14.6 share of holders above each cut 0.56 0.24 0.06 |
| K | 5 R^2 0.168 misgrade 0.608 net -20.5 share of holders above each cut 0.62 0.34 0.14 0.03 |
| best number of holder levels | 1 |
查看完整验算程序
# Chapter 3. p31: how many holder levels a noisy audit can support.
# Quality q ~ N(0, 1), audited score s = q + e, e ~ N(0, sigma^2). The top
# 20 per cent of scores pass (the pass rate of Chapter 2). Holders are
# split into K levels (K = 1 is a single badge) with cut points set by the
# Lloyd-Max rule on the posterior mean among holders, which maximises the
# share of holders' quality variance the level explains (R^2 within).
# Misgrade = share of holders whose level differs from the level their
# true quality would earn at the same cut values. Net value per holder
# (illustrative weights): 100 * R^2 - 4 * (K - 1) for running each extra
# level - 35 * misgrade for appeals and lost trust.
import math
from statistics import NormalDist
ND = NormalDist()
PASS = 0.20
STEP = 0.004
GRID = [-6.0 + i * STEP for i in range(3001)]
def levels(sigma, k):
v = 1 / (1 + sigma * sigma) # variance of E[q | s]
t_m = ND.inv_cdf(1 - PASS) * math.sqrt(v)
mg = [m for m in GRID if m > t_m]
w = [ND.pdf(m / math.sqrt(v)) for m in mg]
tot = math.fsum(w)
w = [x / tot for x in w]
mean_m = math.fsum(m * x for m, x in zip(mg, w))
var_m = math.fsum((m - mean_m) ** 2 * x for m, x in zip(mg, w))
var_q = var_m + sigma * sigma / (1 + sigma * sigma)
qs = [mg[int(len(mg) * j / k)] for j in range(1, k)]
cuts = list(qs)
for _ in range(200):
edges = [t_m] + cuts + [math.inf]
cent = []
for a, b in zip(edges, edges[1:]):
sel = [(m, x) for m, x in zip(mg, w) if a < m <= b]
mass = math.fsum(x for _, x in sel)
cent.append(math.fsum(m * x for m, x in sel) / mass)
cuts = [(x + y) / 2 for x, y in zip(cent, cent[1:])]
edges = [t_m] + cuts + [math.inf]
between = []
for a, b, c in zip(edges, edges[1:], cent):
mass = math.fsum(x for m, x in zip(mg, w) if a < m <= b)
between.append(mass * (c - mean_m) ** 2)
r2 = math.fsum(between) / var_q
# misgrade among holders: cut values on the score scale s = m / v
s_edges = [e / v for e in edges]
terms, hold = [], []
for x in GRID:
f = ND.pdf(x) * STEP
j = 0
while j < k - 1 and x > s_edges[j + 1]:
j += 1
p_hold = 1 - ND.cdf((s_edges[0] - x) / sigma)
same = ND.cdf((s_edges[j + 1] - x) / sigma) - \
ND.cdf((s_edges[j] - x) / sigma)
hold.append(f * p_hold)
terms.append(f * (p_hold - same))
mis = math.fsum(terms) / math.fsum(hold)
shares = []
for c in cuts:
above = math.fsum(x for m, x in zip(mg, w) if m > c)
shares.append(above)
return r2, mis, shares
for sigma in (0.3, 0.5, 1.0):
print("score noise sigma = %.1f, pass rate %.2f" % (sigma, PASS))
best = (1, 0.0)
print(" K=1 R^2 0.000 misgrade 0.000 net 0.0")
for k in range(2, 6):
r2, mis, shares = levels(sigma, k)
net = 100 * r2 - 4 * (k - 1) - 35 * mis
if net > best[1]:
best = (k, net)
print(" K=%d R^2 %.3f misgrade %.3f net %.1f share of holders "
"above each cut %s" % (k, r2, mis, net,
" ".join("%.2f" % s for s in shares)))
print(" best number of holder levels: %d" % best[0])唯一数据源与引用
范雅福(Fan, Y.). (2026). 《行业认证标准自举、斯宾塞信号分级与达人主动申领动力学母典》第 3 章「分级:分几级、切在哪里,以及分数膨胀为什么先甜后苦」(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-358 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026358,
author = {Fan, Yafu},
title = {Bootstrapping Industry Accreditation Standards, Spence Signalling Tiers and Inbound Creator Certification Magnetics},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}