关于「命中率阈值:主动插话何时净为正」,《情境感知副官的阈值与预算》给出的定量结论是什么?

结论:A 段四行,对应本书反复回到的四种提示。带盘参谋的追问提示(用户问完命盘结构,副官想补一句术语解释)收益只有 20 秒,打断 6 秒,阈值是 0.417;换到边界处说,降到 0.267。跨战线切换时的上下文递送收益大得多,阈值只有 0.188,边界处 0.113。订单群摘要的阈值是 0.371,边界处 0.234。门禁红线提醒收益最大,阈值低到 0.073,边界处 0.048。

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: 0.048, 0.073, 0.113, 0.188, 0.234, 0.267

A. p* = (I + F) / (B + F); seconds, illustrative
scenario B I F p* mid p* boundary
chart follow-up hint20 6 4 0.417 0.267
front-switch context90 12 6 0.188 0.113
order-group digest30 8 5 0.371 0.234
gate red-line note240 10 8 0.073 0.048
B. net seconds per 100 prompts, chart follow-up hint, mid-task
p0.2: -520.0
p0.3: -280.0
p0.4: -40.0
p0.5: 200.0
p0.7: 680.0
p0.9: 1,160.0
C. largest interruption cost I that still pays at p (B90,F=6)
p0.3: I max = 22.8 s
p0.5: I max = 42.0 s
p0.7: I max = 61.2 s

核心公式

\[p^* = \frac{I + F}{B + F}\]
查看完整验算程序
# Chapter 1. p11: break-even hit rate p* for one proactive prompt.
# A prompt that turns out useful (a hit) saves B seconds but costs the
# interruption I; a useless prompt (a miss) costs I plus F, the time to
# dismiss it and find the place again. With hit probability p:
#   E[net] = p (B - I) - (1 - p) (I + F) = p (B + F) - (I + F)
#   so p* = (I + F) / (B + F).
# At a task boundary the interruption cost is scaled by D = 2/5. All
# seconds are illustrative; the direction of the boundary effect follows
# Adamczyk and Bailey (2004) and Bailey and Konstan (2006), not their
# numbers.
from fractions import Fraction

D_BOUND = Fraction(2, 5)
SCENARIOS = [
    ("chart follow-up hint", 20, 6, 4),
    ("front-switch context", 90, 12, 6),
    ("order-group digest", 30, 8, 5),
    ("gate red-line note", 240, 10, 8),
]


def p_star(b, i, f):
    return (Fraction(i) + f) / (Fraction(b) + f)


def net(p, b, i, f):
    return p * (b + f) - (i + f)


if __name__ == "__main__":
    print("A. p* = (I + F) / (B + F); seconds, illustrative")
    print("  scenario               B    I   F  p* mid  p* boundary")
    for name, b, i, f in SCENARIOS:
        mid = p_star(b, i, f)
        bd = p_star(b, D_BOUND * i, f)
        print("  {:<21}{:>4} {:>4} {:>3}  {:.3f}   {:.3f}".format(
            name, b, i, f, float(mid), float(bd)))
    print("B. net seconds per 100 prompts, chart follow-up hint, mid-task")
    for p in (0.2, 0.3, 0.4, 0.5, 0.7, 0.9):
        print("  p = {:.1f}: {:>8,.1f}".format(p, 100 * net(p, 20, 6, 4)))
    print("C. largest interruption cost I that still pays at p (B=90,F=6)")
    for p in (0.3, 0.5, 0.7):
        print("  p = {:.1f}: I max = {:.1f} s".format(p, p * 96 - 6))

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 1 章「命中率阈值:主动插话何时净为正」(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