关于「滑动记忆窗:遗忘曲线、旁人暴露与一键删除」,《情境感知副官的阈值与预算》给出的定量结论是什么?

结论:不计旁人代价时(\(k = 0\)),最优窗口仍是 50 分钟,此刻持有 12.50 个旁人分钟;开启就地脱敏后,同样 50 分钟窗口只持有 1.10 个旁人分钟。代价系数升到 0.50,不脱敏的最优窗口缩到 29 分钟,效用降到 0.542;开启脱敏后最优窗口仍有 48 分钟,效用 0.559,暴露只有 1.08。系数升到 2.00,不脱敏的窗口只剩 14 分钟,效用 0.475,净值 0.359;脱敏后窗口 43 分钟,净值 0.524。

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。 与结论吻合的打印值: 0.359, 0.475, 0.50, 0.524, 0.542, 0.559

kredactionbest W minUE bystander-minJ
0.00off500.56012.500.560
0.00on500.5601.100.560
0.25off360.5539.000.515
0.25on490.5601.090.555
0.50off290.5427.250.482
0.50on480.5591.080.550
1.00off210.5185.250.430
1.00on460.5591.050.542
2.00off140.4753.500.359
2.00on430.5581.010.524
查看完整验算程序
# Chapter 5. p52: the window shrinks once bystander exposure is priced.
# Utility U(W) as the power form of p51. A share b = 0.25 of captured
# minutes holds other people's voices or faces. Exposure E(W), in
# bystander-minutes held at any moment, is b W without redaction; with
# on-device redaction within 2 minutes and a miss rate r = 0.05 it is
# b (r W + (1 - r) min(W, 2)). Objective J = U - k E / 60, with k the
# utility given up per bystander-hour held. Parameters illustrative.
TAU, D, H, WN = 2.0, 1.3, 10080.0, 240.0
B, R, LAG = 0.25, 0.05, 2.0


def cum_pow(w):
    return TAU / (D - 1.0) * (1.0 - (1.0 + w / TAU) ** (1.0 - D))


def util(w):
    return cum_pow(w) / cum_pow(H) / (1.0 + w / WN)


def exposure(w, redact):
    if not redact:
        return B * w
    return B * (R * w + (1.0 - R) * min(w, LAG))


if __name__ == "__main__":
    print("  k     redaction  best W min   U      E bystander-min   J")
    for k in (0.0, 0.25, 0.5, 1.0, 2.0):
        for red in (False, True):
            best = max(range(1, 10081),
                       key=lambda w: util(w) - k * exposure(w, red) / 60.0)
            e = exposure(best, red)
            print("  {:.2f}  {:<9}  {:>10,}  {:.3f}   {:>12.2f}    {:.3f}"
                  .format(k, "on" if red else "off", best, util(best), e,
                          util(best) - k * e / 60.0))

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 5 章「滑动记忆窗:遗忘曲线、旁人暴露与一键删除」(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