关于「异议拆解树:覆盖率、深度与补洞次序」,《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:预算 2、4、6 小时时,贪心与最优完全相同,覆盖率依次升到 0.7300、0.7725、0.8325。预算 8 小时,贪心得到 0.8725,最优是 0.8775,比值 0.9943。差别在于选择:贪心选了免费替代品、评价、没时间三项,最优选了免费替代品、评价、分期三项。贪心先拿了一个性价比高的小洞,结果剩下的时间凑不出更大的组合。预算 10 小时与 12 小时时,比值分别是 0.9973 与 0.9948;14 小时全部补完,覆盖率 1.0000。

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: 0.7300, 0.7725, 0.8325, 0.8725, 0.8775, 0.9943

hoursgreedy covoptimum covgreedy/optimum
20.73000.73001.0000
40.77250.77251.0000
60.83250.83251.0000
80.87250.87750.9943
100.91500.91750.9973
120.95500.96000.9948
141.00001.00001.0000
8 h greedy : free_alternative, reviews, no_time
8 h optimum: free_alternative, reviews, instalments
objections closed per hour, first 2 h and last 2 h: 12.00 8.00
查看完整验算程序
# Chapter 4. p42: which objection gaps to close first under a time budget.
# The five gaps of p41 each need a writing task with a cost in hours
# (illustrative) and close a count of logged objections out of 400.
# Greedy by gain per hour (a heap) is compared with the exact optimum
# found by trying every subset. Starting weighted coverage is 0.67.
import heapq
import itertools
import math

BASE = 268
LOG = 400
TASKS = [
    ("free_alternative", 3, 41),
    ("not_for_me", 4, 33),
    ("reviews", 2, 24),
    ("instalments", 3, 18),
    ("no_time", 2, 16),
]


def greedy(budget):
    heap = [(-g / c, name, c, g) for name, c, g in TASKS]
    heapq.heapify(heap)
    used, gain, picked = 0, 0, []
    while heap:
        _, name, c, g = heapq.heappop(heap)
        if used + c <= budget:
            used += c
            gain += g
            picked.append(name)
    return gain, picked


def best(budget):
    top, pick = 0, ()
    for k in range(len(TASKS) + 1):
        for sub in itertools.combinations(TASKS, k):
            c = math.fsum(t[1] for t in sub)
            g = int(math.fsum(t[2] for t in sub))
            if c <= budget and g > top:
                top, pick = g, tuple(t[0] for t in sub)
    return top, pick


if __name__ == "__main__":
    print("hours  greedy cov  optimum cov  greedy/optimum")
    for b in (2, 4, 6, 8, 10, 12, 14):
        g, _ = greedy(b)
        o, _ = best(b)
        print("{:>5}  {:>10.4f}  {:>11.4f}  {:>14.4f}".format(
            b, (BASE + g) / LOG, (BASE + o) / LOG,
            (BASE + g) / (BASE + o)))
    g, gp = greedy(8)
    o, op = best(8)
    print("8 h greedy : " + ", ".join(gp))
    print("8 h optimum: " + ", ".join(op))
    print("objections closed per hour, first 2 h and last 2 h: {:.2f} {:.2f}"
          .format(
        (best(2)[0]) / 2.0, (best(14)[0] - best(12)[0]) / 2.0))

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 4 章「异议拆解树:覆盖率、深度与补洞次序」(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