关于「关键事实复述位:零损耗只能来自模板复制与校验」,《中段遗忘与接力阈值》给出的定量结论是什么?
结论:第六章说明零损耗只能来自模板复制加校验码的关键事实复述位,它把关键事实的存活从 0.601 提到九成以上,但只容得下约七行。
数据口径:书中任务、材料与参数均为示意,模型常数来自 2023 年的一个模型;全部计算由书中程序生成,不调用任何模型接口;本书是非个人化经营建议,不评测、不推荐任何模型或服务。 书中以标注框区分已发表的证据、可运行的 Python 计算(21 段程序)、原典与作者的假说。
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: 0.601
| no slot: key fact survives (q c(4))^4 = 0.601 | ||||
| lines | head: worst | mean | split: worst | mean |
| 5 | 0.924 | 0.940 | 0.860 | 0.913 |
| 10 | 0.878 | 0.910 | 0.835 | 0.884 |
| 20 | 0.810 | 0.860 | 0.795 | 0.848 |
| 40 | 0.732 | 0.789 | 0.735 | 0.794 |
| 80 | 0.663 | 0.711 | 0.665 | 0.726 |
| 160 | 0.608 | 0.644 | 0.608 | 0.657 |
| largest slot with every line read at 0.9 or better, head: 7 | ||||
| largest slot with every line read at 0.9 or better, split: 1 |
查看完整验算程序
# Chapter 6. p61: a key-fact slot carried by the template, not the model.
# Relay of D = 256 units with segment s = 4 (L = 4 layers), per-call
# rewriting fidelity q = 0.97, reader from the p32 joint fit (rounded).
# Without a slot a key fact must survive every rewrite: (q c(4))^4.
# With a slot, the orchestration code copies k key-fact lines verbatim
# into every call (checked by checksum, see p63), so the carry itself
# loses nothing; the only loss left is the final read of the slot. A
# line is 0.1 unit (about 10 tokens); positions are continuous. The
# slot sits at the head, or is split between head and tail.
import math
TAU, M0, K, A, B = 2.4, 0.752, 0.0505, 0.251, 0.128
S, L, Q, W = 4, 4, 0.97, 0.1
def rho(x, n):
v = (M0 - K * math.log(n) + A * math.exp(-(x - 1) / TAU)
+ B * math.exp(-(n - x) / TAU))
return min(1.0, v)
def slot_read(k, split):
n = S + k * W
head = (k + 1) // 2 if split else k
xs = [1 + j * W for j in range(head)]
xs += [n - j * W for j in range(k - head)]
vals = [rho(x, n) for x in xs]
return min(vals), math.fsum(vals) / k
cap4 = math.fsum(rho(i, S) for i in range(1, S + 1)) / S
print("no slot: key fact survives (q c(4))^4 = %.3f" % ((Q * cap4) ** L))
print(" lines head: worst mean split: worst mean")
for k in (5, 10, 20, 40, 80, 160):
hw, hm = slot_read(k, False)
sw, sm = slot_read(k, True)
print("%6d %11.3f %6.3f %12.3f %6.3f" % (k, hw, hm, sw, sm))
for split in (False, True):
kmax = max(k for k in range(1, 400) if slot_read(k, split)[0] >= 0.9)
print("largest slot with every line read at 0.9 or better, %s: %d"
% ("split" if split else "head", kmax))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《长上下文注意力中段遗忘曲线、位置偏置与零损耗分段摘要接力阈值母典》第 6 章「关键事实复述位:零损耗只能来自模板复制与校验」(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-255 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026255,
author = {Fan, Yafu},
title = {Mid-Context Forgetting in Long-Context Attention: Positional Bias and the Threshold for Lossless Segment-Summary Relays},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}