关于「实证闭环:团队 AI 复盘流程与出版局决策质量审计」,《围棋 AI 后的人类决策质量》给出的定量结论是什么?

结论:第八章给出团队 AI 复盘的七条规则、出版方 AI 写作的决策质量审计、蒸馏小工具「复盘质量评分器」与预注册方案。书中以标注框区分已发表的证据、可运行的 Python 计算(17 段程序)、原典与作者的假说。

数据口径:书中除论文数字外的参数均为示意,全部计算由书中程序生成;本书是非个人化经营建议,不针对任何个人决定。 第四章用香农熵度量定式集中度,示意角部局面的策略熵从 1.608 升到 2.360 比特,从众学习模型显示外部评估者让群体提早约 12 年离开局部最优。

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

player eramean DQImedianmatchblundersnovelgap
P1before98.39799.480.49410.083+0.60
P1after98.49699.760.54410.156-0.17
P2before97.82299.080.42240.078-2.65
P2after98.20699.100.41740.150-0.29
P3before97.96599.280.46750.072-1.38
P3after98.38799.340.46120.150-0.39
P4before98.13299.060.45010.089-0.40
P4after98.62399.560.52800.172-0.02
team mean change in DQI: +0.349 (players up: 4 of 4)
design value, novelty scaling ignored: +0.390
查看完整验算程序
# Chapter 8. p82: scoring a team's AI-assisted reviews with the rules
# of the Review Quality Scorer (distill/tool.py uses this function)
# Model (synthetic team, illustrative): 4 players, 6 reviewed games
# each before and after a team adopts a routine AI review, 30 own
# decisions per game. A decision matches the AI's best choice with
# probability m, else loses an exponential amount of win rate with
# mean L; a share q of decisions is novel and novel losses are scaled
# by f. Before: m 0.35, L 3.0, q 0.08, f 1.5. After: m 0.40, L 2.6,
# q 0.14, f 1.1. Rules: loss = best - played (never below 0), DQI =
# 100 - loss, match if loss <= 0.5, blunder if loss >= 10.
import math
import random
import statistics

SEED = 20261112
MATCH_TOL = 0.5
BLUNDER = 10.0


def score(moves):
    loss = [max(0.0, b - p) for b, p, _ in moves]
    dqi = [100.0 - x for x in loss]
    nov = [d for d, (_, _, n) in zip(dqi, moves) if n]
    old = [d for d, (_, _, n) in zip(dqi, moves) if not n]
    match = len([x for x in loss if x <= MATCH_TOL])
    bad = len([x for x in loss if x >= BLUNDER])
    gap = (math.fsum(nov) / len(nov) - math.fsum(old) / len(old)
           if nov and old else 0.0)
    return {"n": len(moves), "mean_dqi": math.fsum(dqi) / len(dqi),
            "median_dqi": statistics.median(dqi),
            "match": match / len(moves), "blunders": bad,
            "novel": len(nov) / len(moves), "novel_gap": gap}


def game(rng, m, mean_loss, q, f):
    out = []
    for _ in range(30):
        best = rng.uniform(35.0, 65.0)
        nv = rng.random() < q
        if rng.random() < m:
            lost = 0.0
        else:
            lost = rng.expovariate(1.0 / (mean_loss * (f if nv else 1.0)))
        out.append((best, best - lost, nv))
    return out


if __name__ == "__main__":
    rng = random.Random(20261112)
    eras = {"before": (0.35, 3.0, 0.08, 1.5), "after": (0.40, 2.6, 0.14, 1.1)}
    res = {}
    print("player era     mean DQI  median  match  blunders  novel  gap")
    for p in range(1, 5):
        for era in ("before", "after"):
            mv = []
            for _ in range(6):
                mv.extend(game(rng, *eras[era]))
            s = score(mv)
            res[(p, era)] = s
            print("  P{}   {:<6}  {:7.3f}  {:6.2f}  {:.3f}    {:2d}"
                  "     {:.3f}  {:+.2f}".format(
                      p, era, s["mean_dqi"], s["median_dqi"], s["match"],
                      s["blunders"], s["novel"], s["novel_gap"]))
    diffs = [res[(p, "after")]["mean_dqi"] - res[(p, "before")]["mean_dqi"]
             for p in range(1, 5)]
    print("team mean change in DQI: {:+.3f} (players up: {} of 4)".format(
        math.fsum(diffs) / 4, len([d for d in diffs if d > 0])))
    print("design value, novelty scaling ignored: {:+.3f}".format(
        (1 - 0.35) * 3.0 - (1 - 0.40) * 2.6))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《围棋人工智能后人类决策质量跃迁、新颖着法率与棋经十三篇策略熵母典》第 8 章「实证闭环:团队 AI 复盘流程与出版局决策质量审计」(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-230 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026230,
  author    = {Fan, Yafu},
  title     = {Post-AI Leaps in Human Go Decision Quality, Novel-Move Rates and Strategic Entropy in the Thirteen-Chapter Classic of Go},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论