关于「作为有噪声信道的指令」,《人机共生、并行智能体与指令信息论》给出的定量结论是什么?

结论:第六章把指令看作有噪声的信道:误读率为 0.15 时,覆盖 0.9 的决定的检查以每个决定 2.06 次的用量得到 0.6355,胜过重复陈述五次的 0.4452。

数据口径:书中的智能体、比率、工时与价值均为示意模型;全部计算由书中程序生成;本书不是法律、财务或雇佣建议,也不对任何具体的 AI 产品作出断言。 书中以标注框区分已发表的证据、可运行的 Python 计算(16 段程序)、原典与作者的假说。

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: 0.4452, 0.6355, 0.9, 2.06

optionuses per decisionerrorP(all 30 right)
restate r = 11.000.1500.0076
restate r = 33.000.0610.1526
restate r = 55.000.0270.4452
restate r = 77.000.0120.6940
test c = 0.51.590.0750.0964
test c = 0.92.060.0150.6355
test c = 1.02.180.0001.0000
查看完整验算程序
# Chapter 6. p12: Restating versus testing. Same channel as p11: each of
# 30 decisions is misread with chance 0.15. Option R: state each decision
# r times and let the agent follow the majority (cost r uses). Option T:
# state it once and add an executable check that covers a share c of
# decisions; a misread covered decision fails its check and is redone
# until it passes (cost 1 + c uses, plus expected redo uses). Residual
# error per decision for T is p (1 - c).
import math
from math import comb

P, D = 0.15, 30

def maj(p, r):
    return math.fsum(comb(r, k) * p ** k * (1 - p) ** (r - k)
               for k in range(r // 2 + 1, r + 1))

if __name__ == "__main__":
    print("   option            uses per decision   error   P(all 30 right)")
    for r in (1, 3, 5, 7):
        e = maj(P, r)
        print("   restate r = {:<4}  {:17.2f}   {:5.3f}   {:15.4f}".format(
            r, r, e, (1 - e) ** D))
    for c in (0.5, 0.9, 1.0):
        e = P * (1 - c)
        redo = c * P / (1 - P)
        print("   test c = {:<6}   {:17.2f}   {:5.3f}   {:15.4f}".format(
            c, 1 + c + redo, e, (1 - e) ** D))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《人机共生认知架构、多智能体并行流水线与指令信息论下界》第 6 章「作为有噪声信道的指令」(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-103 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026103,
  author    = {Fan, Yafu},
  title     = {Human-AI Symbiotic Cognition, Parallel Agent Pipelines and Prompt Information Bounds},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论