关于「早期录取率:打分有误差时,门槛定在哪里」,《达人认证的信号与分级》给出的定量结论是什么?

结论:打分有误差时,录取率越低,精度越高,这一点流行说法说对了;可是召回、可用人数与信息量都在下降。p21 说明,信息量在录取率接近强者基础比例时最大,误差标准差 0.5 时峰值在 0.35;p22 说明,加入品牌方的使用决策以后,最优录取率是 0.22,质量提升在九成以内的区间只是从 0.18 到 0.23;p23 说明,早期重审 10 人零失败,误收率的上界仍有 0.238,平台不能据此宣称精度。

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。

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 标准库与结论吻合的打印值: 0.18, 0.22, 0.23

passprecisionbrands filteringdemandholdersfilleduplift
0.050.995200.0600.0100100.0184.5
0.100.975197.1591.4200200.0313.9
0.150.939178.6535.8300300.0417.1
0.200.894154.5463.5400400.0500.8
0.250.841127.0380.9500380.9433.0
0.300.78697.8293.4600293.4304.2
0.400.67639.9119.8800119.8103.5
0.500.5770.00.010000.00.0
0.600.4940.00.012000.00.0
best pass rate 0.22: uplift 519.2, precision 0.873, 440 holders
within 90% of best: pass rates 0.18 to 0.23
at 0.05 holders are too few; at 0.50 brands stop filtering
查看完整验算程序
# Chapter 2. p22: neither the lowest pass rate nor the most holders is best.
# 2,000 applicants a quarter; 200 brands, each filling 3 campaign slots.
# A brand filters on the certificate only if its precision P(strong | pass)
# reaches the brand's own requirement, spread uniformly on [0.60, 0.98].
# Slots that filtering brands fill from holders yield a quality uplift of
# E[q | pass] standard units each. Holders are scarce when r * 2,000 is
# below the filtering demand. Quality model as in p21 (sigma 0.5, base
# rate 0.30). All parameters are illustrative.
import math
from statistics import NormalDist

ND = NormalDist()
BASE, SIGMA = 0.30, 0.5
C_STRONG = ND.inv_cdf(1 - BASE)
GRID = [-6.0 + i * 0.002 for i in range(6001)]
APPLICANTS, BRANDS, SLOTS = 2000, 200, 3
REQ_LO, REQ_HI = 0.60, 0.98


def stats(r):
    t = ND.inv_cdf(1 - r) * math.sqrt(1 + SIGMA * SIGMA)
    joint = [(x, ND.pdf(x) * 0.002 * (1 - ND.cdf((t - x) / SIGMA)))
             for x in GRID]
    p_pass = math.fsum(v for _, v in joint)
    p_ps = math.fsum(v for x, v in joint if x > C_STRONG)
    return p_ps / p_pass, math.fsum(x * v for x, v in joint) / p_pass


def value(r):
    prec, eq = stats(r)
    share = min(1.0, max(0.0, (prec - REQ_LO) / (REQ_HI - REQ_LO)))
    demand = BRANDS * share * SLOTS
    holders = r * APPLICANTS
    filled = min(demand, holders)
    return prec, share, demand, holders, filled, filled * eq


print("pass  precision  brands filtering  demand  holders  filled  uplift")
rows = []
for k in range(2, 61):
    r = k / 100
    rows.append((r,) + value(r))
for row in rows:
    r = row[0]
    if round(r * 100) in (5, 10, 15, 20, 25, 30, 40, 50, 60):
        print("%.2f   %.3f       %5.1f           %5.1f   %5.0f   %5.1f  %6.1f" % (
            r, row[1], row[2] * BRANDS, row[3], row[4], row[5], row[6]))
best = max(rows, key=lambda x: x[6])
band = [x[0] for x in rows if x[6] >= 0.9 * best[6]]
print("best pass rate %.2f: uplift %.1f, precision %.3f, %d holders" % (
    best[0], best[6], best[1], round(best[4])))
print("within 90%% of best: pass rates %.2f to %.2f" % (band[0], band[-1]))
print("at 0.05 holders are too few; at 0.50 brands stop filtering")

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 2 章「早期录取率:打分有误差时,门槛定在哪里」(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}
}

本书其他章节结论