关于「实战:海桥对账后台从诊断到重排的完整案例」,《数据后台减负与防呆表单》给出的定量结论是什么?

结论:第八章让海桥的对账后台走完全程:两周埋雷审计的基线漏检率 0.683;红色只留给四条对账规则后,红行从 435 降到 40,阳性预测值 0.425,模型预期的漏检率从 0.664 降到 0.089,即使新规则只有设定灵敏度的六成也是 0.339;状态机、金额掩码、币种默认值与批量操作的报数撤销,把返工率从 0.0064 降到 0.0013;每行处理秒数只少 0.49 秒,在八周内测不准,所以只作次要指标。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。

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。

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: 0.089, 0.339, 0.425, 0.664, 40

new red rule                    fires  anomalies  PPV
N1 difference beyond RM 1.00       20          9  0.450
N2 duplicate pair                   9          3  0.333
N3 credit without order             8          2  0.250
N4 paid, no credit after 2 days     3          3  1.000
new: red rows 40 (share 0.033), anomalies in red 17 of 18, PPV 0.425, red per screen 0.83
icon rows 412, folded clean S7/S9 rows 180, anomalies inside folded rows 0
expected anomalies caught a day: old 6.04, new 16.40; miss rate old 0.664, new 0.089
proper checks left for icon rows after red: 20; anomalies reached through icons 0
seconds per row: old 12.34, new 11.85; change -0.49
budget  30: new view miss rate 0.283
budget  45: new view miss rate 0.089
budget  60: new view miss rate 0.089
budget  90: new view miss rate 0.089
sensitivity of the new red rules scaled by m (expected values, all red rows checked):
  m 1.0: expected miss rate 0.066
  m 0.8: expected miss rate 0.202
  m 0.6: expected miss rate 0.339
  m 0.4: expected miss rate 0.476
查看完整验算程序
# Chapter 8. p81: The redesigned view on the same day: four reconciliation-identity red rules, icons, budget.
# Rebuilds the p10 day (seed 20261112). New red rules test the reconciliation identity itself; their firing
# probabilities are settings, not measurements. The old R2, R3, R4, R5, R6 become icons in one narrow column.
# Default view "to handle": red rows first, then rows ranked by icons (R6, R2 first); fully clean rows in
# S7 or S9 fold into one summary line. Same detection and time settings as p80 for a fair comparison.
import math
import random

SEED, ROWS, SCREEN, BUDGET, P_CHECK, P_GLANCE = 20261112, 1200, 25, 60, 0.95, 0.25
OPS = ["Ah Rong", "Chee Ming", "Mei Ling", "Ka Ho"]
STATUSES = [("S1", 0.06), ("S2", 0.14), ("S3", 0.10), ("S4", 0.12), ("S5", 0.11), ("S6", 0.17),
            ("S7", 0.20), ("S8", 0.04), ("S9", 0.06)]
CHANNELS = [("FPX transfer", 0.46), ("card", 0.24), ("e-wallet", 0.22), ("direct bank transfer", 0.08)]
CURRENCIES = [("JPY", 0.45), ("KRW", 0.25), ("TWD", 0.15), ("MYR", 0.15)]
RM_PER_UNIT = {"JPY": 0.030, "KRW": 0.0033, "TWD": 0.14, "MYR": 1.0}   # sample rates (assumed)
ANOMALIES = [("A1", 6), ("A2", 3), ("A3", 4), ("A4", 2), ("A5", 3)]
RULES = [("R1", 0.12, {"A1": 1.0, "A3": 1.0, "A2": 0.3, "A4": 0.3, "A5": 0.3}),
         ("R2", 0.10, {"A1": 0.5, "A2": 0.5, "A3": 0.5, "A4": 0.5, "A5": 0.9}),
         ("R3", None, None),
         ("R4", 0.08, {"A1": 0.2, "A2": 0.2, "A3": 0.2, "A4": 0.4, "A5": 0.2}),
         ("R5", None, None),
         ("R6", 0.04, {"A1": 0.3, "A2": 0.3, "A3": 0.3, "A4": 0.3, "A5": 0.3})]


def pick(rng, table):
    u, acc = rng.random(), 0.0
    for row in table:
        acc += row[-1]
        if u < acc:
            return row[0]
    return table[-1][0]


def make_day(seed):
    rng = random.Random(seed)
    rows = []
    for _ in range(ROWS):
        amt = round(math.exp(math.log(180.0) + 0.9 * rng.gauss(0.0, 1.0)), 2)
        rows.append({"op": OPS[rng.randrange(4)], "st": pick(rng, STATUSES), "ch": pick(rng, CHANNELS),
                     "cur": pick(rng, CURRENCIES), "amt": amt, "an": None})
    idx = rng.sample(range(ROWS), sum(n for _, n in ANOMALIES))
    k = 0
    for code, n in ANOMALIES:
        for _ in range(n):
            rows[idx[k]]["an"] = code
            k += 1
    for r in rows:
        r["fired"] = []
        for name, p0, p1 in RULES:
            if name == "R3":
                hit = r["amt"] >= 1000
            elif name == "R5":
                hit = r["ch"] == "e-wallet" and r["st"] in ("S1", "S2")
            else:
                hit = rng.random() < (p1[r["an"]] if r["an"] else p0)
            if hit:
                r["fired"].append(name)
    return rows


NEW = [("N1 difference beyond RM 1.00", 0.010, {"A1": 1.0, "A3": 0.9}),
       ("N2 duplicate pair", 0.003, {"A2": 1.0}),
       ("N3 credit without order", 0.002, {"A4": 1.0}),
       ("N4 paid, no credit after 2 days", 0.002, {"A5": 1.0})]


def catch_and_time(rows, red, ranked, folded, budget):
    checked = set(id(r) for r in red[:budget])
    for r in ranked[:max(0, budget - len(red))]:
        checked.add(id(r))
    caught = 0.0
    for r in rows:
        if r["an"]:
            if id(r) in checked:
                caught += P_CHECK
            elif id(r) not in folded:
                caught += P_GLANCE
    return caught, checked


rows = make_day(SEED)
n_an = sum(1 for r in rows if r["an"])
rng = random.Random(SEED + 81)
for r in rows:
    r["new"] = [nm for nm, p0, p1 in NEW if rng.random() < (p1.get(r["an"], 0.0) if r["an"] else p0)]
    r["tie"] = rng.random()
print("new red rule                    fires  anomalies  PPV")
for nm, _, _ in NEW:
    f = [r for r in rows if nm in r["new"]]
    tp = sum(1 for r in f if r["an"])
    print("{:31s} {:5d}  {:9d}  {:.3f}".format(nm, len(f), tp, tp / len(f) if f else 0.0))
red_new = [r for r in rows if r["new"]]
tp_new = sum(1 for r in red_new if r["an"])
rest = [r for r in rows if not r["new"]]
ranked = sorted([r for r in rest if r["fired"]],
                key=lambda r: (-("R6" in r["fired"]) - ("R2" in r["fired"]), -len(r["fired"]), r["tie"]))
folded = set(id(r) for r in rest if not r["fired"] and r["st"] in ("S7", "S9"))
print("new: red rows {} (share {:.3f}), anomalies in red {} of {}, PPV {:.3f}, red per screen {:.2f}".format(
    len(red_new), len(red_new) / ROWS, tp_new, n_an, tp_new / len(red_new), len(red_new) / (ROWS / SCREEN)))
print("icon rows {}, folded clean S7/S9 rows {}, anomalies inside folded rows {}".format(
    len(ranked), len(folded), sum(1 for r in rows if id(r) in folded and r["an"])))
red_old = [r for r in rows if r["fired"]]
old_caught = 0.0
p_chk = min(1.0, BUDGET / len(red_old))
for r in rows:
    if r["an"]:
        p = p_chk if r["fired"] else 0.0
        old_caught += p * P_CHECK + (1 - p) * P_GLANCE
new_caught, checked = catch_and_time(rows, red_new, ranked, folded, BUDGET)
print("expected anomalies caught a day: old {:.2f}, new {:.2f}; miss rate old {:.3f}, new {:.3f}".format(
    old_caught, new_caught, 1 - old_caught / n_an, 1 - new_caught / n_an))
spare = BUDGET - len(red_new)
print("proper checks left for icon rows after red: {}; anomalies reached through icons {}".format(
    spare, sum(1 for r in ranked[:spare] if r["an"])))
old_sec = ROWS * 4.0 + BUDGET * 120.0 + (len(red_old) - BUDGET) * 0.5 * 15.0
new_sec = (ROWS - len(folded)) * 4.0 + BUDGET * 120.0 + (len(ranked) - max(0, spare)) * 0.5 * 15.0
print("seconds per row: old {:.2f}, new {:.2f}; change {:+.2f}".format(
    old_sec / ROWS, new_sec / ROWS, (new_sec - old_sec) / ROWS))
for budget in (30, 45, 60, 90):
    c, _ = catch_and_time(rows, red_new, ranked, folded, budget)
    print("budget {:3d}: new view miss rate {:.3f}".format(budget, 1 - c / n_an))
print("sensitivity of the new red rules scaled by m (expected values, all red rows checked):")
for m in (1.0, 0.8, 0.6, 0.4):
    found = 0.0
    for code, n in ANOMALIES:
        p_red = 1.0
        for _, _, p1 in NEW:
            p_red *= 1.0 - m * p1.get(code, 0.0)
        p_red = 1.0 - p_red
        found += n * (p_red * P_CHECK + (1 - p_red) * P_GLANCE)
    print("  m {:.1f}: expected miss rate {:.3f}".format(m, 1 - found / n_an))

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 8 章「实战:海桥对账后台从诊断到重排的完整案例」(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}
}

本书其他章节结论