关于「蒸馏的天花板:压缩只能保留,合成才能生成」,《超越蒸馏:商业器官合成与自检》给出的定量结论是什么?

结论:再看表格。保留全部十二条规则,描述长度 522 字节,目标市场零遗漏。砍到十条,描述只剩 415 字节,压缩比 0.795,原市场覆盖率仍有 0.9439,看上去几乎无损;可目标市场已有 0.2266 的场景无规则可依,硬约束场景漏掉 0.6045。砍到八条,原市场覆盖率还有 0.8758,硬约束场景却漏掉 1.0000,也就是全部漏掉。砍到两条,描述只剩 75 字节,压缩比 0.144,目标市场有 0.6673 的场景无人处理。

数据口径:本书主张向上合成:把专著定理连同目标市场的约束(计价单位、价格带、合规红线、界面门禁)正向编译成自带规格与不变量的小器官。 书中参数为示意,全部计算由书中程序生成。

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: 0.144, 0.2266, 0.6045, 0.6673, 0.795, 0.8758

source share of hard rules 0.0920
target share of hard rules 0.3780
kbytesratiosrc-covertgt-misshard-miss
125221.0001.00000.00000.0000
104150.7950.94390.22660.6045
83180.6090.87580.39581.0000
62410.4620.78950.45461.0000
41560.2990.67130.53661.0000
2750.1440.48340.66731.0000
查看完整验算程序
# Chapter 1. p12: distilling harder drops the rare rules that carry limits.
# Model: a mother book yields 12 operating rules. In the source market the
# rule frequencies follow a Zipf law (rank r has weight 1 / r). Three rare
# rules (ranks 9, 11, 12) are hard limits: a compliance red line, a margin
# floor in a second unit, a tap-target gate. "Distilling harder" keeps
# only the top-k rules by frequency. In the target market the hard limits
# fire far more often (weight x6). We draw 20,000 target cases (seeded)
# and count cases the kept rules cannot handle, and hard-limit misses.
# Rule sizes in bytes and all weights are illustrative.
import math
import random

SIZES = [34, 41, 29, 52, 38, 47, 33, 44, 61, 36, 58, 49]
HARD = {9, 11, 12}


def weights(boost):
    w = []
    for r in range(1, 13):
        x = 1.0 / r
        if r in HARD:
            x *= boost
        w.append(x)
    t = math.fsum(w)
    return [x / t for x in w]


if __name__ == "__main__":
    rng = random.Random(378012)
    src = weights(1.0)
    tgt = weights(6.0)
    cases = rng.choices(range(1, 13), weights=tgt, k=20000)
    n_hard = 0
    for c in cases:
        if c in HARD:
            n_hard += 1
    print("source share of hard rules {:.4f}".format(
        math.fsum(src[r - 1] for r in sorted(HARD))))
    print("target share of hard rules {:.4f}".format(
        math.fsum(tgt[r - 1] for r in sorted(HARD))))
    print(" k  bytes  ratio  src-cover  tgt-miss  hard-miss")
    total_bytes = math.fsum(SIZES)
    for k in (12, 10, 8, 6, 4, 2):
        kept = set(range(1, k + 1))
        nbytes = math.fsum(SIZES[:k])
        cover = math.fsum(src[r - 1] for r in sorted(kept))
        miss = 0
        hard_miss = 0
        for c in cases:
            if c not in kept:
                miss += 1
                if c in HARD:
                    hard_miss += 1
        print("{:>2}  {:>5.0f}  {:>5.3f}  {:>9.4f}  {:>8.4f}  {:>9.4f}"
              .format(k, nbytes, nbytes / total_bytes, cover,
                      miss / len(cases),
                      hard_miss / n_hard if n_hard else 0.0))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《超越蒸馏:可执行商业器官合成、生成式代码结晶与活体系统自举母典》第 1 章「蒸馏的天花板:压缩只能保留,合成才能生成」(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-378 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026378,
  author    = {Fan, Yafu},
  title     = {Beyond Distillation: Executable Business-Organ Synthesis, Generative Code Crystallisation and Living-System Bootstrapping},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论