关于「真实反超:目标多变、模型设错与悬崖代价」,《效用、真实与双账本》给出的定量结论是什么?

结论:本章用三段程序给适存胜过真实划了边界。目标从一个增加到两个,真实编码就在 0.703 的随机目标组上不输;收益模型偏 0.15,真实编码反超;悬崖存在时,赢家是能把悬崖边界如实刻画出来的那种编码。适存胜过真实是一个有条件的定理,而不是一条处世格言。

数据口径:书中参数与数据均为示意或合成,全部计算由书中程序生成;本书不是医疗建议,不提供任何临床心理服务,不是个人投资建议,不承诺任何收益;功能性信念只用于自己的行动框架,不用于欺骗他人。 第六章建立双账本:行动框架可按效用编码,概率账以 Brier 分数校准,硬边界账不让步;框架改写概率账时,示意破产比例从 0.0925 升到 0.7055。

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: 0.703

KTRUTHFIT-1truth >= fit-1mean gap
10.56000.56940.260-0.0094
20.56290.55450.703+0.0084
30.56730.55150.873+0.0158
40.56610.54750.903+0.0185
60.56500.54460.953+0.0204
80.56510.54240.980+0.0227
查看完整验算程序
# Chapter 2. p21: when goals change, a truth-coded interface catches up
# After the argument of Berke, Walter-Terrill, Jara-Ettinger and Scholl
# (2022). x is uniform on [0, 1] (a 400-point grid). There are K goals;
# goal k values x by a bump exp(-((x - m_k) / s_k)^2 / 2) with m_k in
# [0, 1] and s_k in [0.08, 0.3] drawn at random. Perception is fixed and
# has four categories (two bits). FIT-1 is tuned to goal 1: its
# categories are the quartiles of goal-1 value. TRUTH uses the quartiles
# of x. Each round one goal is active (all equally often); the agent
# knows which, sees the categories of two territories and takes the one
# whose category has the higher mean value under that goal (a flexible
# decision layer on a fixed perceptual layer). Payoffs are exact
# expectations over the grid; 300 random goal sets per K. Illustrative.
import math
import random

GRID = [(i + 0.5) / 400 for i in range(400)]


def cats_by(score):
    order = sorted(range(len(GRID)), key=lambda i: (score[i], i))
    lab = [0] * len(GRID)
    for r, i in enumerate(order):
        lab[i] = r * 4 // len(GRID)
    return lab


TRUTH = cats_by(GRID)


def payoff(lab, vals):
    means = []
    for c in range(4):
        v = [vals[i] for i in range(len(GRID)) if lab[i] == c]
        means.append(math.fsum(v) / len(v))
    return math.fsum(max(a, b) for a in means for b in means) / 16


rng = random.Random(42721)
print(" K   TRUTH    FIT-1   truth >= fit-1   mean gap")
for k in (1, 2, 3, 4, 6, 8):
    tr, fi, wins, gaps = [], [], 0, []
    for _ in range(300):
        goals = []
        for _ in range(k):
            m, s = rng.random(), rng.uniform(0.08, 0.3)
            goals.append([math.exp(-0.5 * ((x - m) / s) ** 2)
                          for x in GRID])
        fit = cats_by(goals[0])
        a = math.fsum(payoff(TRUTH, v) for v in goals) / k
        b = math.fsum(payoff(fit, v) for v in goals) / k
        tr.append(a)
        fi.append(b)
        gaps.append(a - b)
        wins += a >= b
    print("%2d  %.4f  %.4f  %15.3f  %+9.4f" % (
        k, math.fsum(tr) / 300, math.fsum(fi) / 300, wins / 300,
        math.fsum(gaps) / 300))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《效用优先于字面真实:演化适存度博弈、功能性信念重塑与物理硬边界母典》第 2 章「真实反超:目标多变、模型设错与悬崖代价」(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-427 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026427,
  author    = {Fan, Yafu},
  title     = {When Utility Outperforms Literal Truth: Evolutionary Fitness-Payoff Games, Functional Belief Engineering and Physical Guardrails},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论