关于「例外呈报:人工介入次数、预算超支、截止概率与利特尔定律」,《零传令兵调度母典》给出的定量结论是什么?

结论:例外分四类:不可逆动作、预算超支、验收三连败、锁等待超时,另有兜底的未分类错误;呈报自带五槽上下文,截止只来自真实时限。模拟中每批介入均值 9.951 次,比人工传话下降 0.830,但签字是地板,永远降不到零。预算与截止是两个独立的风险,要分开定、分开报;利特尔定律让台账能查出丢失与卡住的任务。

数据口径:在四卷一批的示意模拟里,改用意图编译器加调度器之后,每批人工介入均值为 9.951 次,比人工传话的 58.4 次下降 0.830;但不可逆与对外动作必须由人签字,仅签字一项就占旧口径的 0.137,是永远降不到零的地板。 书中参数为示意,全部计算由书中程序生成;全书只写调度原理与公开流程形态,不含服务器、端口、账户、密钥与任何部署命令;本书是非个人化经营分析,不构成经营、法律或运维建议。

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: 0.830, 9.951

interventions per batch (4 volumes, 20,000 batches)
  floor (signatures only)  8
  mean 9.951  P50 10  P90 12  P99 14  max 17
  batches with no exception at all 0.1364
  exceptions per agent task 0.0375
old relay hand-carries per batch 58.4
reduction of the mean 0.830, never below floor share 0.137
查看完整验算程序
# Chapter 6. p61: how many times a human must step in per batch.
# Model (illustrative): one overnight batch = 4 volumes, each the 15-
# node graph of chapter 2. Escalation rules: (1) every irreversible or
# outward step needs a signature (2 per volume, never automated);
# (2) budget overrun, chance 0.10 per volume (budget set at the P90
# of p62, about 66 % above the nominal plan);
# (3) an acceptance check failing three times in a row, per attempt
# failure 0.3, on 5 checked nodes; (4) a lock wait timing out, 0.02 on
# each of 6 locking nodes; (5) an unclassified agent error, 0.01 on
# each of 13 agent nodes. The old relay: a person carries every
# hand-off between windows, 0.6 of 21 edges, plus the signatures.
import math
import random

N = 20000


def batch(rng):
    sig = exc = 0
    for _ in range(4):
        sig += 2
        exc += rng.random() < 0.10
        for _ in range(5):
            exc += all(rng.random() < 0.3 for _ in range(3))
        for _ in range(6):
            exc += rng.random() < 0.02
        for _ in range(13):
            exc += rng.random() < 0.01
    return sig, exc


if __name__ == "__main__":
    rng = random.Random(40761)
    tot, excs = [], []
    for _ in range(N):
        s, e = batch(rng)
        tot.append(s + e)
        excs.append(e)
    tot.sort()
    q = [tot[int(p * N) - 1] for p in (0.5, 0.9, 0.99)]
    mean = math.fsum(tot) / N
    print("interventions per batch (4 volumes, {:,} batches)".format(N))
    print("  floor (signatures only)  8")
    print("  mean {:.3f}  P50 {}  P90 {}  P99 {}  max {}".format(
        mean, q[0], q[1], q[2], tot[-1]))
    zero = len([e for e in excs if e == 0]) / N
    print("  batches with no exception at all {:.4f}".format(zero))
    print("  exceptions per agent task {:.4f}".format(
        math.fsum(excs) / N / (4 * 13)))
    relay = 4 * (2 + 0.6 * 21)
    print("old relay hand-carries per batch {:.1f}".format(relay))
    print("reduction of the mean {:.3f}, never below floor share "
          "{:.3f}".format(1 - mean / relay, 8 / relay))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《总司令自然语言意图编译、有向无环任务自动拆解与零传令兵自治调度母典》第 6 章「例外呈报:人工介入次数、预算超支、截止概率与利特尔定律」(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-407 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026407,
  author    = {Fan, Yafu},
  title     = {Commander Intent Compilation, Automated DAG Task Decomposition and Zero-Courier Multi-Agent Dispatch},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论