关于「方法:技能包、四个保留指标、证据分级与蒸馏四步」,《专著蒸馏与技能包工程》给出的定量结论是什么?

结论:第一章证明两个反常识结论:把整本书贴进提示词,覆盖率是 1.000,有效保留率却是 0.000;十二个样本全部通过时,失败率的 95% 上限仍有 0.2209,要压到百分之一以下需要 299 个;按写明的假设,没有样本的十二条规则包十轮修改之内至少坏一条的概率是 0.9115。第二至第七章依次写抽取、编译、护栏与校验、四件套统一规格、自进化与治理。

数据口径:书中样本全部虚构,定价与成交规则只写一般框架,不是法律或合规意见。 全书按证据分级标注每一条说法,以《荀子·劝学》考辨工具与积累,所有计算由书中程序生成,作者的判断一律写进假说框并写明什么结果会推翻它。

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: 0.2209, 0.9115, 299, 95

golden n all pass95% upper bound on failure rate 3/n
120.2209 0.2500
240.1173 0.1250
500.0582 0.0600
1000.0295 0.0300
2990.0100 0.0100
3000.0099 0.0100
samples needed so that zero failures bound the rate below 0.0559
samples needed so that zero failures bound the rate below 0.01299
q0.02 per rule per round; share of rules intact after n rounds
coverage c n1 n=5 n=10 n=20 P(any of 12 broken by n=10)
0.00.9800 0.9039 0.8171 0.6676 0.9115
0.50.9900 0.9510 0.9044 0.8179 0.7006
0.90.9980 0.9900 0.9802 0.9608 0.2136
1.01.0000 1.0000 1.0000 1.0000 0.0000
with no tests, half of the rules are gone after 34.3 rounds
查看完整验算程序
# Chapter 1. p12: What "all golden samples pass" proves, and how fast an untested pack decays (closed form).
# Part 1, zero numerator: if n independent golden samples all pass, the one-sided 95% upper bound on
# the per-case failure rate is 1 - 0.05 ** (1 / n), close to 3 / n (the "rule of three").
# Part 2, silent decay (a model, parameters assumed, not measured): each edit round breaks each rule
# with probability q = 0.02, independently; a golden suite notices a broken rule with probability c
# and the edit is reverted. A rule survives one round with probability 1 - q * (1 - c), so after
# n rounds the expected share of rules intact is (1 - q * (1 - c)) ** n. For a 12-rule pack the chance
# that at least one rule is broken is 1 - (1 - q * (1 - c)) ** (12 * n). Chapter 6 simulates this.
from math import ceil, log

Q = 0.02
RULES = 12


def upper95(n):
    return 1 - 0.05 ** (1 / n)


if __name__ == "__main__":
    print("   golden n   all pass -> 95% upper bound on failure rate   3/n")
    for n in (12, 24, 50, 100, 299, 300):
        print("   {:8d}   {:44.4f}   {:.4f}".format(n, upper95(n), 3 / n))
    for target in (0.05, 0.01):
        need = ceil(log(0.05) / log(1 - target))
        print("   samples needed so that zero failures bound the rate below {:.2f}: {}".format(target, need))
    print("   q = {:.2f} per rule per round; share of rules intact after n rounds".format(Q))
    print("   coverage c     n=1     n=5    n=10    n=20   P(any of 12 broken by n=10)")
    for c in (0.0, 0.5, 0.9, 1.0):
        s = 1 - Q * (1 - c)
        row = ["{:7.4f}".format(s ** n) for n in (1, 5, 10, 20)]
        print("   {:10.1f} {}   {:25.4f}".format(c, " ".join(row), 1 - s ** (RULES * 10)))
    half = log(0.5) / log(1 - Q)
    print("   with no tests, half of the rules are gone after {:.1f} rounds".format(half))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《专著向智能体技能包无损蒸馏、规则可执行化编译与系统自进化工程》第 1 章「方法:技能包、四个保留指标、证据分级与蒸馏四步」(版本 0.1.0). 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-150 · v0.1.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026150,
  author    = {Fan, Yafu},
  title     = {Lossless Distillation of Monographs into Agent Skill Packs, Executable Rule Compilation and Self-Evolving Systems},
  year      = {2026},
  publisher = {5D Library},
  edition   = {0.1.0},
  note      = {DOI pending}
}

本书其他章节结论