关于「实战工具与检查器规格:调度台账、意图编译与任务有向图调度检查器、预注册检验」,《零传令兵调度母典》给出的定量结论是什么?

结论:第八章给出调度台账、意图编译与任务有向图调度检查器、按四件套分组的检查器规格表与预注册检验:差错率的非劣效检验每组需要 1,188 个智能体任务,而把违规率上界压到 0.01 以下需要连续 299 个干净批次。书中以标注框区分已发表的证据、可运行的 Python 计算(22 段程序)、原典与作者的假说。

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: 0.01, 1188

claim A: batches per arm, relay 58.4 vs dispatcher
target mean29.2: 1 batches
target mean40.0: 3 batches
target mean50.0: 13 batches
claim B: agent tasks per arm1,188
= 23 batches of 52 agent tasks
simulated power at that size0.7990
margin 0.01 instead4,749 tasks per arm
查看完整验算程序
# Chapter 8. p82: pre-registered sample sizes for the two core claims.
# Claim A: the dispatcher halves human interventions per batch against
# the hand-carried relay (illustrative baseline 58.4 per batch, from
# p61). Poisson counts, log rate ratio test, two-sided alpha 0.05,
# power 0.8. Claim B: the error rate per agent task does not rise by
# more than 0.02 over a baseline of 0.04 (non-inferiority, one-sided
# alpha 0.05, power 0.8). A seeded simulation checks the power of B.
import math
import random
from statistics import NormalDist

Z = NormalDist()


def n_poisson(m0, m1, alpha=0.05, power=0.8):
    za, zb = Z.inv_cdf(1 - alpha / 2), Z.inv_cdf(power)
    return math.ceil((za + zb) ** 2 * (1 / m0 + 1 / m1)
                     / math.log(m1 / m0) ** 2)


def n_noninf(p, margin, alpha=0.05, power=0.8):
    za, zb = Z.inv_cdf(1 - alpha), Z.inv_cdf(power)
    return math.ceil((za + zb) ** 2 * 2 * p * (1 - p) / margin ** 2)


def power_b(n, p=0.04, margin=0.02, reps=2000):
    rng = random.Random(40782)
    za = Z.inv_cdf(0.95)
    wins = 0
    for _ in range(reps):
        x0 = len([1 for _ in range(n) if rng.random() < p])
        x1 = len([1 for _ in range(n) if rng.random() < p])
        p0, p1 = x0 / n, x1 / n
        se = math.sqrt((p0 * (1 - p0) + p1 * (1 - p1)) / n) or 1e-9
        wins += (p1 - p0 - margin) / se < -za
    return wins / reps


if __name__ == "__main__":
    print("claim A: batches per arm, relay 58.4 vs dispatcher")
    for m1 in (29.2, 40.0, 50.0):
        print("  target mean {:>5}: {} batches".format(
            m1, n_poisson(58.4, m1)))
    nb = n_noninf(0.04, 0.02)
    print("claim B: agent tasks per arm = {:,}".format(nb))
    print("  = {} batches of 52 agent tasks".format(math.ceil(nb / 52)))
    print("  simulated power at that size = {:.4f}".format(power_b(nb)))
    print("  margin 0.01 instead: {:,} tasks per arm".format(
        n_noninf(0.04, 0.01)))

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 8 章「实战工具与检查器规格:调度台账、意图编译与任务有向图调度检查器、预注册检验」(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