关于「方法:后台的三类损失、四个指标、证据分级与刘晏考辨」,《数据后台减负与防呆表单》给出的定量结论是什么?
结论:第一章是方法:虚构的代购商行海桥每天 1,200 行流水、18 行真异常,六条整行标红规则让 435 行变红,阳性预测值只有 0.037;在每天只能认真查 60 行的预算下,阈值越放越松,漏检率可以从 0.500 升到 0.853;把前后对照放在业务量上升的旺季里,估计出的效果平均只有真实效果的一半不到,改成每周之内随机交替才回到 −0.1040。第二至第七章依次写扫视路径与表格排版、前注意特征与条件格式的层级、信号检测与警报阈值、渐进披露与默认视图、防错表单,以及对账、订单、库存、客户资料四种后台的手册与工具规格。
数据口径:本书不是会计、审计或法律意见。 书中以标注框区分已发表的证据、可运行的计算、原典与作者的假说,每条假说都写明什么结果会推翻它;原典考辨取唐代刘晏的报数制度,原文只转述大意。
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: 0.500, 0.853, 1200, 18, 60
| rows 1200, true anomalies 18, d' 2.0, careful checks per day 60 | ||||||||
| thresh | hit_rate | false_alarm | red_rows | red_share | PPV | caught | missed | miss_rate |
| 3.5 | 0.067 | 0.0002 | 1.5 | 0.001 | 0.814 | 1.20 | 16.80 | 0.933 |
| 3.0 | 0.159 | 0.0013 | 4.5 | 0.004 | 0.642 | 2.86 | 15.14 | 0.841 |
| 2.5 | 0.309 | 0.0062 | 12.9 | 0.011 | 0.431 | 5.55 | 12.45 | 0.691 |
| 2.0 | 0.500 | 0.0228 | 35.9 | 0.030 | 0.251 | 9.00 | 9.00 | 0.500 |
| 1.5 | 0.691 | 0.0668 | 91.4 | 0.076 | 0.136 | 8.17 | 9.83 | 0.546 |
| 1.0 | 0.841 | 0.1587 | 202.7 | 0.169 | 0.075 | 4.48 | 13.52 | 0.751 |
| 0.5 | 0.933 | 0.3085 | 381.5 | 0.318 | 0.044 | 2.64 | 15.36 | 0.853 |
| 0.0 | 0.977 | 0.5000 | 608.6 | 0.507 | 0.029 | 1.73 | 16.27 | 0.904 |
| -0.5 | 0.994 | 0.6915 | 835.2 | 0.696 | 0.021 | 1.29 | 16.71 | 0.929 |
| most anomalies caught at threshold 2.0: 9.00 caught with 35.9 red rows | ||||||||
| threshold where red rows equal the budget: 1.733; caught there 10.89 |
查看完整验算程序
# Chapter 1. p11: Why more red can mean more misses: a fixed careful-check budget meets a loosening rule.
# Equal-variance signal detection: normal rows score N(0,1), true anomalies score N(d',1). A rule paints a
# row red when its score exceeds threshold c. The team can check only B red rows carefully per day and,
# since red rows all look alike, picks them at random. Settings, not measurements: d' = 2.0, B = 60.
import math
ROWS, ANOM = 1200, 18
NORMAL = ROWS - ANOM
D_PRIME = 2.0
BUDGET = 60
def tail(x):
return 0.5 * math.erfc(x / math.sqrt(2.0))
print("rows {}, true anomalies {}, d' {}, careful checks per day {}".format(ROWS, ANOM, D_PRIME, BUDGET))
print("thresh hit_rate false_alarm red_rows red_share PPV caught missed miss_rate")
best = None
for c10 in range(35, -6, -5):
c = c10 / 10.0
hit, fa = tail(c - D_PRIME), tail(c)
tp, fp = ANOM * hit, NORMAL * fa
red = tp + fp
caught = tp * min(1.0, BUDGET / red)
print("{:6.1f} {:8.3f} {:11.4f} {:8.1f} {:9.3f} {:.3f} {:6.2f} {:6.2f} {:9.3f}".format(
c, hit, fa, red, red / ROWS, tp / red, caught, ANOM - caught, 1 - caught / ANOM))
if best is None or caught > best[1]:
best = (c, caught, red)
print("most anomalies caught at threshold {:.1f}: {:.2f} caught with {:.1f} red rows".format(*best))
lo, hi = 0.0, 6.0
for _ in range(80):
mid = (lo + hi) / 2.0
if ANOM * tail(mid - D_PRIME) + NORMAL * tail(mid) > BUDGET:
lo = mid
else:
hi = mid
c = (lo + hi) / 2.0
print("threshold where red rows equal the budget: {:.3f}; caught there {:.2f}".format(
c, ANOM * tail(c - D_PRIME)))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《高密度数据工作台视觉减负、条件格式显著性与零培训防呆表单工程》第 1 章「方法:后台的三类损失、四个指标、证据分级与刘晏考辨」(版本 0.1.0). 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-138 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026138,
author = {Fan, Yafu},
title = {Cognitive Unloading of Dense Data Workbenches, Conditional-Format Salience and Zero-Training Error-Proof Form Engineering},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}