关于「风险分级确认:确认疲劳与撤销窗口」,《情境感知副官的阈值与预算》给出的定量结论是什么?

结论:暂存 5 秒几乎拦不住什么(0.014),净值为负。10 秒拦下 0.062,30 秒拦下 0.290,60 秒拦下 0.493,120 秒拦下 0.648,300 秒拦下 0.734,但五分钟的延误代价已经超过收益,净值转负。最优暂存是 74 秒,拦下 0.548,每千条净值 9.04。程序 p62 里可撤销动作用的 60 秒撤销窗口,拦截率 0.49,正来自这里。

数据口径:语音端点等待是延迟的地板,示意架构下端到端延迟中位数 1,083 毫秒、95 分位 1,540 毫秒,一秒内完成的只有 0.323,「零延迟」「人类开口前」「电影级」都被改写为分位数与推翻条件。 书中参数为示意,数据为合成;全部计算由书中 23 段标准库程序生成。

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: 0.014, 0.062, 0.290, 0.49, 0.493, 0.548

hold w sP(caught)net per 1,000 sends
50.014-0.08
100.0620.86
200.1833.49
300.2905.70
600.4938.79
1200.6487.44
3000.734-7.99
best hold 74 s, P(caught) 0.548, net 9.04 per 1,000 sends
60 s window catch rate used in p62: 0.49
查看完整验算程序
# Chapter 6. p63: sizing the hold-and-undo window for an outbound send.
# A wrong send is noticed after T seconds: with probability 0.75 T is
# log-normal (median 40 s, sigma 1.0); otherwise it is not noticed
# within the hour. Holding every send for w seconds catches a wrong one
# with P(T <= w) and delays every send by w. Per send:
#   net(w) = e P(T <= w) - h w,  e = 0.03 wrong sends, h = 0.0001 per s
# (the error rate and the price of delay in caught-error units are
# illustrative). Undo is a fail-safe default in the sense of Saltzer and
# Schroeder (1975): the safe state is the one reached by doing nothing.
import math
from statistics import NormalDist

E, H = 0.03, 0.0001
Z = NormalDist()


def caught(w):
    return 0.75 * Z.cdf(math.log(w / 40.0) / 1.0)


if __name__ == "__main__":
    print("  hold w s   P(caught)   net per 1,000 sends")
    for w in (5, 10, 20, 30, 60, 120, 300):
        print("  {:>8}    {:.3f}       {:>7.2f}".format(
            w, caught(w), 1000.0 * (E * caught(w) - H * w)))
    best = max(range(1, 601), key=lambda w: E * caught(w) - H * w)
    print("best hold {} s, P(caught) {:.3f}, net {:.2f} per 1,000 sends"
          .format(best, caught(best), 1000.0 * (E * caught(best) - H * best)))
    print("60 s window catch rate used in p62: {:.2f}".format(caught(60)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《全天候多模态情境感知智能副官、低延迟视听记忆流与主动预判辅助母典》第 6 章「风险分级确认:确认疲劳与撤销窗口」(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-409 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026409,
  author    = {Fan, Yafu},
  title     = {Ambient Multimodal Cognitive Adjutants: Low-Latency Audio-Visual Memory Streams, Proactive Intent Anticipation and Real-Time Decision Support},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论