12.1 目标与假设

#!/usr/bin/env python3
# file: sim/nm_12_1_hypothesis.py
# tier: S
# verify: /opt/homebrew/bin/python3 sim/nm_12_1_hypothesis.py
"""卷三 12.1 目标与假设（S 层）：把「这次可玩性测试要验什么」写成可证伪的命题。

一次可玩性测试如果没有假设，就只是一次「点着玩」。本节把要验的东西落成四条**可证伪**
的命题：每条都给出可测的整数、判据（op + 阈值）与**证伪条件**——只要批跑结果越过阈值，
命题就被证伪，脚本与人都无法含糊过去。

四条假设（全部挂在叙事路径的**结构量**上，不碰任何主观体验）：
  H1 覆盖率：N=24 条采样路径走过的去重对话边 ≥ 850‰（规格目标 coverage_target_permille）
  H2 结局覆盖：N=24 条路径到达的**不同结局数** ≥ 4（5 个结局里至少摸到 4 个）
  H3 跨宿主一致：24 条路径在浏览器注入后的终局摘要与 Python 侧**逐位一致**（= 24）
  H4 分布熵：24 条路径的结局分布熵 ≥ 1500 毫比特（规格阈值 ending_entropy_min_milli_bits）

脚本只读 config/story_spec.json，算出**零假设基线**（确定性单路径能覆盖多少），再把四条
假设与阈值落成 out/nm_12_1_hypothesis.json，供 12.2 的批跑与 12.3 的决策读。公共底座见
sim/nm_12_common.py。全部整数；熵公式用 math.log2（不进入状态，同卷三 3.1 口径）。
"""
from __future__ import annotations

import json
import os
import sys

import nm_12_common as cm

BATCH_N, SEED = 24, 20261016


def null_baseline(spec: dict) -> tuple[int, int]:
    """零假设：每步选第一个可用选项走一条路径，返回（走过的去重边数, 到达结局数）。"""
    s, edges = cm.init_state(spec), set()
    for _ in range(cm.STEP_CAP):
        cur = cm.node_of(spec, s["node"])
        if cur.get("ending"):
            break
        opts = [c for c in (cur.get("choices") or []) if cm.cond_ok(c.get("when"), s)]
        if not opts:
            break
        edges.add((s["node"], opts[0]["goto"]))
        s = cm.step(s, opts[0])
    return len(edges), 1 if cm.node_of(spec, s["node"]).get("ending") else 0


def main() -> int:
    spec = cm.load_spec()
    E = len(cm.dedup_edges(spec))
    m = spec["metrics"]
    det_edges, det_ends = null_baseline(spec)
    det_permille = det_edges * 1000 // E

    hyps = [
        {"id": "H1", "metric": "coverage_permille", "op": ">=", "threshold": m["coverage_target_permille"],
         "statement": f"N={BATCH_N} 条采样路径的去重对话边覆盖率 ≥ {m['coverage_target_permille']}‰",
         "falsified_if": f"批跑覆盖率 < {m['coverage_target_permille']}‰",
         "source": "config/story_spec.json#metrics.coverage_target_permille"},
        {"id": "H2", "metric": "endings_reached", "op": ">=", "threshold": 4,
         "statement": f"N={BATCH_N} 条路径到达的不同结局数 ≥ 4（共 {len(spec['endings'])} 个结局）",
         "falsified_if": "批跑到达的不同结局数 < 4",
         "source": "设计约定：批跑须覆盖多数结局"},
        {"id": "H3", "metric": "browser_python_match", "op": ">=", "threshold": BATCH_N,
         "statement": f"浏览器注入的 {BATCH_N} 条路径终局摘要与 Python 侧逐位一致（= {BATCH_N}）",
         "falsified_if": f"一致条数 < {BATCH_N}",
         "source": "跨语言 digest 同源（卷二 13.1）"},
        {"id": "H4", "metric": "ending_entropy_milli_bits", "op": ">=",
         "threshold": m["ending_entropy_min_milli_bits"],
         "statement": f"N={BATCH_N} 条路径的结局分布熵 ≥ {m['ending_entropy_min_milli_bits']} 毫比特",
         "falsified_if": f"批跑分布熵 < {m['ending_entropy_min_milli_bits']} 毫比特",
         "source": "config/story_spec.json#metrics.ending_entropy_min_milli_bits"},
    ]

    report = {
        "seed": SEED, "batch_n": BATCH_N, "step_cap": cm.STEP_CAP, "edges_total": E,
        "null_baseline": {"strategy": "deterministic-first", "edges": det_edges,
                          "permille": det_permille, "endings_reached": det_ends},
        "targets": {"coverage_target_permille": m["coverage_target_permille"],
                    "ending_entropy_min_milli_bits": m["ending_entropy_min_milli_bits"],
                    "endings_total": len(spec["endings"])},
        "hypotheses": hyps,
        "proxy": "覆盖率 / 结局数 / 分布熵均为结构量代理指标，只做相对比较，不声称体验好坏",
    }

    print(f"故事图      {len(spec['dialogue'])} 节点  去重对话边 {E}  结局 {len(spec['endings'])}")
    print(f"零假设基线  确定性单路径 {det_edges}/{E} = {det_permille}‰  到达结局 {det_ends}")
    print(f"批跑规模    N={BATCH_N}  seed={SEED}")
    print("假设        判据                      阈值   证伪条件")
    for h in hyps:
        print(f"  {h['id']}  {h['metric']:<26}{h['op']}{h['threshold']:>6}   {h['falsified_if']}")

    assert det_permille < m["coverage_target_permille"], "零假设基线应低于目标，否则假设无张力"
    assert 0 < det_ends < 4, "确定性单路径不应已覆盖 4 个结局"
    assert len({h["id"] for h in hyps}) == 4, "四条假设 id 应互不相同"

    os.makedirs(cm.ROOT + "/out", exist_ok=True)
    with open(cm.ROOT + "/out/nm_12_1_hypothesis.json", "w", encoding="utf-8") as fh:
        json.dump(report, fh, ensure_ascii=False, indent=2)
        fh.write("\n")
    print("wrote out/nm_12_1_hypothesis.json")
    print(f"OK nm_12_1_hypothesis：4 条可证伪假设已落盘，零假设基线 {det_permille}‰ "
          f"（目标 {m['coverage_target_permille']}‰）")
    return 0


if __name__ == "__main__":
    sys.exit(main())

故事图      48 节点  去重对话边 56  结局 5
零假设基线  确定性单路径 25/56 = 446‰  到达结局 1
批跑规模    N=24  seed=20261016
假设        判据                      阈值   证伪条件
  H1  coverage_permille         >=   850   批跑覆盖率 < 850‰
  H2  endings_reached           >=     4   批跑到达的不同结局数 < 4
  H3  browser_python_match      >=    24   一致条数 < 24
  H4  ending_entropy_milli_bits >=  1500   批跑分布熵 < 1500 毫比特
wrote out/nm_12_1_hypothesis.json
OK nm_12_1_hypothesis：4 条可证伪假设已落盘，零假设基线 446‰ （目标 850‰）