关于「不容定价的红灯:说停就停、弱势转人工与人手上限」,《三级灯号切分点》给出的定量结论是什么?

结论:不设上限时,最优切分点就是第二章的 0.400 与 0.950,负荷 22.1 个单位,每百段损失 295.2 元,这里包含了硬规则对话的人工代价。上限降到 20 时,影子价格 1.96 元,绿灯门限从 0.950 降到 0.938,红灯门限从 0.400 降到 0.253,绿灯比例升到 0.572,损失只增加到 297.1 元。上限降到 16 时,影子价格 11.05 元,红灯门限降到零,即除了硬规则,不再有任何对话因为置信度低而整段转人工,绿灯门限降到 0.881,损失 320.0 元。上限降到 12 时,影子价格 40.14 元,绿灯门限 0.699,绿灯占 0.911,损失 408.9 元。

数据口径:本书用纯标准库程序把命题写成可计算的模型,并修正蓝图三处:带拒识通道的最优识别早见于 Chow 1957 年的论文,1970 年的论文给出误差与拒判的一般关系;规则只对校准过的概率成立,把成本比算出的切分点直接套在过度自信机器人的自报分数上,每百段对话损失 265.4 元,等渗回归校准后 239.8 元;私域客服有「机器人起草、真人过目」的第三种动作,需要六格损失矩阵与两条切分点,示意参数下答对概率低于 0.4000 亮红灯、不低于 0.9500 亮绿灯。 书中参数为示意,对话全部虚构;全部计算由书中程序生成;程序不读写任何文件,不连接任何客服系统;本书是非个人化经营建议,任何线上灯号的改动须由主事者人工复核放行。

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: 0.253, 0.400, 0.572, 0.699, 0.881, 0.911

caplamp_lop_higreenamberredloadcost/100
250.000.4000.9500.5140.4700.01722.1295.2
201.960.2530.9380.5720.4220.00520.0297.1
185.250.0060.9170.6460.3540.00018.0304.1
1611.050.0000.8810.7340.2660.00016.0320.0
1420.680.0000.8210.8230.1770.00014.0350.7
1240.140.0000.6990.9110.0890.00012.0408.9
9below the hard-rule load 10: add staff, never drop the rules
查看完整验算程序
# Chapter 6. p62: cut points when human hours are capped.
# Model (illustrative): 10 % of chats are red by hard rule (stop
# requests and vulnerable customers, p61) and are never displaced. The
# other 90 % follow the p21 loss matrix (LW 50, CA 2, M 0.8, CR 8) on
# the p11 bot. Human load counts a take-over as 1 unit and an amber
# review as KAPPA = 0.25 unit, per 100 chats. A cap on load is met by
# a shadow price lam per unit (Lagrange): red costs CR + lam, amber
# CA + KAPPA lam; lam is found by bisection. Cost excludes lam.
import math

LW, CA, M, CR, KAPPA, HARD = 50.0, 2.0, 0.8, 8.0, 0.25, 0.10


def nodes(mu=3.0, sd=1.6, n=4001):
    zs = [mu + sd * (-8.0 + 16.0 * i / (n - 1)) for i in range(n)]
    ws = [math.exp(-0.5 * ((z - mu) / sd) ** 2) for z in zs]
    tot = math.fsum(ws)
    return [(1.0 / (1.0 + math.exp(-z)), w / tot) for z, w in zip(zs, ws)]


PW = nodes()


def solve(lam):
    hi = 1 - (CA + KAPPA * lam) / (M * LW)
    lo = 1 - (CR + lam - CA - KAPPA * lam) / ((1 - M) * LW)
    sh = {0: [], 1: [], 2: []}
    cost = []
    for p, w in PW:
        k = 0 if p >= hi else (1 if p >= lo else 2)
        sh[k].append(w)
        cost.append(w * ((1 - p) * LW, CA + (1 - p) * (1 - M) * LW,
                         CR)[k])
    g, a, r = (math.fsum(sh[k]) for k in range(3))
    load = 100 * (HARD + (1 - HARD) * (r + KAPPA * a))
    money = 100 * (HARD * CR + (1 - HARD) * math.fsum(cost))
    return max(lo, 0.0), hi, g, a, r, load, money


print("  cap    lam   p_lo   p_hi  green  amber  red   load  cost/100")
for cap in (25.0, 20.0, 18.0, 16.0, 14.0, 12.0, 9.0):
    if cap < 100 * HARD:
        print("%5.0f  below the hard-rule load %.0f: add staff, never"
              " drop the rules" % (cap, 100 * HARD))
        continue
    lo_l, hi_l = 0.0, 0.0
    if solve(0.0)[5] > cap:
        hi_l = 1.0
        while solve(hi_l)[5] > cap:
            hi_l *= 2
        for _ in range(60):
            mid = (lo_l + hi_l) / 2
            if solve(mid)[5] > cap:
                lo_l = mid
            else:
                hi_l = mid
    out = solve(hi_l)
    print("%5.0f  %5.2f  %5.3f  %5.3f  %5.3f  %5.3f  %5.3f  %5.1f  %8.1f"
          % ((cap, hi_l) + out))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《私域客服人机交接三级灯号精算、误判损失矩阵与最优置信度切分点母典》第 6 章「不容定价的红灯:说停就停、弱势转人工与人手上限」(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-266 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026266,
  author    = {Fan, Yafu},
  title     = {Three Lights for Handing Over between Bot and Human in Private Customer Service: Misclassification Loss Matrices, Chow's Reject Rule and Optimal Cut Points on Calibrated Confidence},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论