关于「承诺强度的操作化:每周探针数、首探延迟与停止规则」,《先手探针与隐藏事实》给出的定量结论是什么?

结论:第五章用信息价值给出停止规则,探针停在 375 位访客;每周探针数有上限,超过信息峰值即过度扰动;首探延迟有价格,立刻开始的 90 天遗憾是 773.6,一直不动是 5,804.9。

数据口径:书中参数为示意,全部计算由书中程序生成,数据全部是合成数据。 本书是非个人化经营建议,不构成法律、税务或投资意见,不向任何人承诺收入。

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: 5804.9, 773.6, 90

delay(days)regret/runends rightbits learned
0773.60.9400.795
71,003.30.9380.782
141,348.50.9200.751
302,494.00.9450.772
603,923.40.9220.722
905,804.90.7000.222
查看完整验算程序
# Chapter 5. p53: what waiting before the first probe costs.
# Illustrative. Default tier A (price 30) versus a new real tier B
# (price 40). True conversions are drawn per run from the priors
# qA ~ Beta(24, 76) and qB ~ Beta(17, 83). 100 visitors a day for 90
# days. Before day d every visitor sees A; from day d each day's share
# shown B equals the posterior chance that B earns more (probability
# matching, normal approximation). Reported over 400 runs: regret
# against the better tier, share of runs that end on the better tier,
# and bits learned about "which tier is better" (entropy drop).
import math
import random

DAYS, PER_DAY, RUNS = 90, 100, 400


def p_b_better(a1, b1, a2, b2):
    m1, m2 = 30 * a1 / (a1 + b1), 40 * a2 / (a2 + b2)
    v1 = 900 * a1 * b1 / ((a1 + b1) ** 2 * (a1 + b1 + 1))
    v2 = 1600 * a2 * b2 / ((a2 + b2) ** 2 * (a2 + b2 + 1))
    z = (m2 - m1) / math.sqrt(v1 + v2)
    return 0.5 * (1.0 + math.erf(z / math.sqrt(2.0)))


def h2(p):
    p = min(max(p, 1e-12), 1 - 1e-12)
    return -(p * math.log2(p) + (1 - p) * math.log2(1 - p))


def run(rng, delay):
    qa, qb = rng.betavariate(24, 76), rng.betavariate(17, 83)
    best = max(30 * qa, 40 * qb)
    st = [24.0, 76.0, 17.0, 83.0]
    lost = []
    h0 = h2(p_b_better(*st))
    for day in range(DAYS):
        share = p_b_better(*st) if day >= delay else 0.0
        for _ in range(PER_DAY):
            if rng.random() < share:
                buy = rng.random() < qb
                st[2 if buy else 3] += 1
                lost.append(best - 40 * qb)
            else:
                buy = rng.random() < qa
                st[0 if buy else 1] += 1
                lost.append(best - 30 * qa)
    pb = p_b_better(*st)
    right = (pb > 0.5) == (40 * qb > 30 * qa)
    return math.fsum(lost), right, h0 - h2(pb)


if __name__ == "__main__":
    rng = random.Random(39553)
    print("delay(days)  regret/run  ends right  bits learned")
    for d in (0, 7, 14, 30, 60, 90):
        regs, rights, bits = [], 0, []
        for _ in range(RUNS):
            r, ok, b = run(rng, d)
            regs.append(r)
            rights += 1 if ok else 0
            bits.append(b)
        print("{:>11}  {:>10,.1f}  {:>10.3f}  {:>12.3f}".format(
            d, math.fsum(regs) / RUNS, rights / RUNS,
            math.fsum(bits) / RUNS))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《高能动性主动推理、意念强度先手扰动与商业隐藏事实勘探母典》第 5 章「承诺强度的操作化:每周探针数、首探延迟与停止规则」(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-395 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026395,
  author    = {Fan, Yafu},
  title     = {High-Agency Active Inference, Intentional Perturbation Probes and Hidden Market-Truth Exploration},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论