关于「惊奇、锚点与完读」,《计算著作学与自校验专著》给出的定量结论是什么?

结论:第二章以比特度量惊奇:一次把 10 倍改为 1.35 倍的修正带来 12.65 比特;玩具读者模型中最佳的稳定剂量是每节 1.2 比特;销售数据把锚点的价值高估 3.39 倍。

数据口径:书中的读者、书、错误率与成本均为示意模型;全部计算由书中程序生成;本书不是法律、财务或出版建议。 书中以标注框区分已发表的证据、可运行的 Python 计算(16 段程序)、原典与作者的假说。

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: 1.2, 1.35, 10, 12.65

surprise of a 10x claim cut to 1.35x (log scale beliefs):
KL = 12.65 bits
scheduletotal bitscompletion
flat 0.2 bit8.00.0036
flat 1.2 bit48.00.6868
flat 2.5 bit100.00.0000
pulse 1.2 every 4th18.00.1239
pulse 2.2 / 0.248.00.0036
best steady dose 1.2 bit a section, completion 0.6868
查看完整验算程序
# Chapter 2. p03: Surprise pulses and completion. A section's surprise is
# the Kullback-Leibler divergence (bits) from what the reader believed to
# what the section shows. Toy reader: engagement e decays by 0.8 a
# section, gains 1.2 per bit of surprise and loses 0.5 per squared bit
# (overload); the reader quits a section with chance 1 / (1 + exp(2 e)).
# 40 sections, five schedules; then the best steady dose.
import math
from math import exp, log

def kl_normal_bits(m0, s0, m1, s1):
    # KL( N(m1, s1) || N(m0, s0) ) in bits: what the reader learns when a
    # belief N(m0, s0) is replaced by N(m1, s1)
    nats = log(s0 / s1) + (s1 ** 2 + (m1 - m0) ** 2) / (2 * s0 ** 2) - 0.5
    return nats / log(2)

def completion(sched):
    e, alive = 0.0, 1.0
    for s in sched:
        e = 0.8 * e + 1.2 * s - 0.5 * s * s
        alive *= 1 - 1 / (1 + exp(2 * e))
    return alive

N = 40
SCHEDULES = (
    ("flat 0.2 bit", [0.2] * N),
    ("flat 1.2 bit", [1.2] * N),
    ("flat 2.5 bit", [2.5] * N),
    ("pulse 1.2 every 4th", [1.2 if i % 4 == 0 else 0.2 for i in range(N)]),
    ("pulse 2.2 / 0.2", [2.2 if i % 2 == 0 else 0.2 for i in range(N)]),
)

if __name__ == "__main__":
    print("   surprise of a 10x claim cut to 1.35x (log scale beliefs):")
    print("   KL = {:.2f} bits".format(
        kl_normal_bits(log(10), 0.5, log(1.35), 0.15)))
    print("   schedule              total bits   completion")
    for name, s in SCHEDULES:
        print("   {:<20}  {:10.1f}   {:10.4f}".format(
            name, math.fsum(s), completion(s)))
    best = max((completion([d / 10] * N), d / 10) for d in range(1, 31))
    print("   best steady dose {:.1f} bit a section, completion {:.4f}"
          .format(best[1], best[0]))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《计算著作学、先证伪后立界修辞与自校验专著编译器》第 2 章「惊奇、锚点与完读」(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-101 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026101,
  author    = {Fan, Yafu},
  title     = {Computational Bibliometrics, Falsification-First Rhetoric and Self-Verifying Monograph Compilers},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论