关于「实证闭环:每卷双锚流程、回喂坍缩速率计算器与预注册检验」,《抗模型崩溃双锚工程》给出的定量结论是什么?

结论:第八章给出每卷八步双锚流程、蒸馏小工具「回喂坍缩速率计算器」与预注册检验。书中以标注框区分已发表的证据、可运行的 Python 计算(22 段程序)、原典与作者的假说。

数据口径:书中参数为示意;全部计算由书中程序生成,不调用任何模型接口,不使用任何读者或客户的数据;本书是非个人化的经营与工程参考。 书中以标注框区分已发表的证据、可运行的 Python 计算(22 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

N=100 n=100 share=0.0 rule=mle p=0.001
expected_ratio0.3660
fresh_coverage_of_p0.0000
fresh_items0
fresh_needed_952995
half_life_expected_mle68.9676
half_life_typical34.3391
mean_drift_accumulate1.6351
mean_drift_replace101.0000
p_missing_one_gen0.9048
typical_ratio_no_fresh0.1328
N=100 n=100 share=0.1 rule=mle p=0.001
expected_ratio0.9497
fresh_coverage_of_p0.0100
fresh_items10
fresh_needed_952995
half_life_expected_mle68.9676
half_life_typical34.3391
mean_drift_accumulate1.6351
mean_drift_replace101.0000
p_missing_one_gen0.9048
typical_ratio_no_fresh0.1328
查看完整验算程序
# Chapter 8. p81: the core of the feedback collapse rate calculator.
# Model: the formulas of p11, p12, p31, p41 and p43 in one place, as
# used by distill/tool.py. Inputs: N points per generation, n
# generations, fresh share a, refit rule (mle or unbiased), the
# probability p of one rare category. Outputs: the expected variance
# ratio, the typical (geometric-mean) ratio without fresh data, the two
# half-lives, the replace and accumulate mean drift N E[e_n^2], the
# chance p is missing after one generation, and fresh coverage of p.
import math

EULER = 0.5772156649015329


def psi_half(k):
    if k % 2 == 0:
        return -EULER + math.fsum(1.0 / j for j in range(1, k // 2))
    m = (k - 1) // 2
    return (-EULER - 2.0 * math.log(2.0)
            + math.fsum(2.0 / (2 * j - 1) for j in range(1, m + 1)))


def expected_ratio(n_samp, gens, share, rule):
    c = 1.0 if rule == "mle" else n_samp / (n_samp - 1.0)
    v, m = 1.0, 0.0
    for _ in range(gens):
        mix = share + (1.0 - share) * v
        v, m = (c * ((1.0 - 1.0 / n_samp) * mix
                     + share * (1.0 - share) * m),
                (1.0 - share) ** 2 * m + mix / n_samp)
    return v


def drift(n_samp, rule):
    k = n_samp - 1
    div = n_samp if rule == "mle" else k
    return psi_half(k) + math.log(2.0) - math.log(div)


def report(n_samp, gens, share, rule, p):
    d = drift(n_samp, rule)
    fresh = round(share * n_samp)
    out = {
        "expected_ratio": expected_ratio(n_samp, gens, share, rule),
        "typical_ratio_no_fresh": math.exp(gens * d),
        "half_life_expected_mle": math.log(0.5)
        / math.log(1.0 - 1.0 / n_samp),
        "half_life_typical": math.log(0.5) / d,
        "mean_drift_replace": float(gens + 1),
        "mean_drift_accumulate": math.fsum(
            1.0 / (k * k) for k in range(1, gens + 2)),
        "p_missing_one_gen": (1.0 - p) ** n_samp,
        "fresh_items": fresh,
        "fresh_coverage_of_p": 1.0 - (1.0 - p) ** fresh,
        "fresh_needed_95": math.ceil(math.log(0.05)
                                     / math.log(1.0 - p)),
    }
    return out


for share in (0.0, 0.1):
    r = report(100, 100, share, "mle", 0.001)
    print("N=100 n=100 share=%.1f rule=mle p=0.001" % share)
    for key in sorted(r):
        val = r[key]
        txt = "%d" % val if isinstance(val, int) else "%.4f" % val
        print("  %-24s %s" % (key, txt))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《抗模型崩溃双锚真源工程、自产语料回喂方差坍缩速率与标准库验算校勘免疫母典》第 8 章「实证闭环:每卷双锚流程、回喂坍缩速率计算器与预注册检验」(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-265 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026265,
  author    = {Fan, Yafu},
  title     = {Anti-Collapse Engineering with Two Anchors of Truth: the Variance-Collapse Rate of Self-Generated Training Data and Immunity through Standard-Library Recomputation and Textual Collation},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论