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。
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 pass | 95% upper bound on failure rate 3/n |
| 12 | 0.2209 0.2500 |
| 24 | 0.1173 0.1250 |
| 50 | 0.0582 0.0600 |
| 100 | 0.0295 0.0300 |
| 299 | 0.0100 0.0100 |
| 300 | 0.0099 0.0100 |
| samples needed so that zero failures bound the rate below 0.05 | 59 |
| samples needed so that zero failures bound the rate below 0.01 | 299 |
| q | 0.02 per rule per round; share of rules intact after n rounds |
| coverage c n | 1 n=5 n=10 n=20 P(any of 12 broken by n=10) |
| 0.0 | 0.9800 0.9039 0.8171 0.6676 0.9115 |
| 0.5 | 0.9900 0.9510 0.9044 0.8179 0.7006 |
| 0.9 | 0.9980 0.9900 0.9802 0.9608 0.2136 |
| 1.0 | 1.0000 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}
}