关于「台阶为何变成斜坡:有限批次、logistic 晋级与瓶颈门」,《短视频流量池冷启动》给出的定量结论是什么?

结论:第三章说明硬门槛在数据上是斜坡,批次从 100 人增至 3,000 人时放行区间宽度从 0.124 收窄到 0.023;示意基准视频的破圈概率为 0.03585,瓶颈在第四道门。

数据口径:书中所有门槛、池子规模与权重均为示意,平台规则以其官方公示为准;全部计算由书中程序生成;本书是非个人化经营建议,只讨论真实内容质量与合规测试,不涉及任何刷量或规避平台规则的做法。 第三章说明硬门槛在数据上是斜坡,批次从 100 人增至 3,000 人时放行区间宽度从 0.124 收窄到 0.023;示意基准视频的破圈概率为 0.03585,瓶颈在第四道门。

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: 0.023, 0.124, 100, 3000

true cn=100 exact/logitn=300 exact/logitn=3000 exact
0.340.1230.0980.0170.0210.000
0.360.2320.1850.0840.0710.000
0.380.3760.3230.2560.2170.013
0.400.5380.5000.5220.5000.507
0.420.6920.6770.7760.7830.988
0.440.8170.8150.9270.9291.000
0.460.9040.9020.9840.9791.000
n =100: pass 0.1 -> 0.9 between c = 0.335 and 0.459, width 0.124
n =300: pass 0.1 -> 0.9 between c = 0.363 and 0.435, width 0.072
n = 3000: pass 0.1 -> 0.9 between c = 0.389 and 0.412, width 0.023

核心公式

\[s = \frac{\sqrt{3}}{\pi}\sqrt{\frac{t(1-t)}{n}},\]
查看完整验算程序
# Chapter 3. p31: a hard threshold seen through a small batch.
# Gate 1 shows a new video to n viewers and promotes it if the
# measured completion rate X / n reaches the threshold 0.40, where X
# is binomial(n, c) and c is the video's true completion rate. All
# numbers are illustrative; no platform publishes its gate rules.
# The exact pass probability is the binomial upper tail. A logistic
# curve with the same centre and the same spread, slope
#     s = sqrt(3) / pi * sqrt(t (1 - t) / n), t = 0.40 the threshold,
# approximates it; the step becomes a ramp whose width shrinks
# like 1 / sqrt(n).
import math

THR = 0.40


def tail(n, c, k):
    # P(X >= k), X ~ binomial(n, c), terms added with math.fsum
    lp, lq = math.log(c), math.log1p(-c)
    terms = [math.exp(math.lgamma(n + 1) - math.lgamma(j + 1)
                      - math.lgamma(n - j + 1) + j * lp + (n - j) * lq)
             for j in range(k, n + 1)]
    return math.fsum(terms)


def logistic(n, c):
    s = math.sqrt(3.0) / math.pi * math.sqrt(THR * (1.0 - THR) / n)
    return 1.0 / (1.0 + math.exp(-(c - THR) / s))


print("true c    n=100 exact/logit   n=300 exact/logit   n=3000 exact")
for c in (0.34, 0.36, 0.38, 0.40, 0.42, 0.44, 0.46):
    row = []
    for n in (100, 300):
        k = math.ceil(THR * n - 1e-9)
        row.append("%7.3f %6.3f" % (tail(n, c, k), logistic(n, c)))
    k3 = math.ceil(THR * 3000 - 1e-9)
    print("%6.2f   %s   %s   %10.3f" % (c, row[0], row[1],
                                        tail(3000, c, k3)))
for n in (100, 300, 3000):
    k = math.ceil(THR * n - 1e-9)
    lo = [c / 1000.0 for c in range(250, 551)
          if tail(n, c / 1000.0, k) >= 0.1][0]
    hi = [c / 1000.0 for c in range(250, 551)
          if tail(n, c / 1000.0, k) >= 0.9][0]
    print("n = %4d: pass 0.1 -> 0.9 between c = %.3f and %.3f, width %.3f"
          % (n, lo, hi, hi - lo))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《全域短视频赛马池冷启动动力学、完播互动复播阶跃阈值与零投流破圈测试矩阵母典》第 3 章「台阶为何变成斜坡:有限批次、logistic 晋级与瓶颈门」(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-273 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026273,
  author    = {Fan, Yafu},
  title     = {Cold-Start Dynamics of Short-Video Traffic Pools: Step Thresholds in Completion, Engagement and Replay and a Zero-Ad-Spend Breakout Test Matrix},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论