tianxuan / stellaris /test_closed_loop_phase3.py
masterofthestars's picture
Tianxuan: full project + data (checkpoints, world_model, data without raw) with correct directory structure
db7bc06
Raw History Blame Contribute Delete
4.61 kB
#!/usr/bin/env python3
"""
群星 P8 · Phase 3 闭环不变量测试 (宪章第 7/8 条: 机制可量化)
用法: python3 -m stellaris.test_closed_loop_phase3
断言 (对应文档 4.5 推理 + 4.6 认知控制 + 9 认知循环 Phase 3):
1. 六引擎全部启动成功 (perception/state/memory/world_model/reasoning/cognitive_control)
2. 认知控制始终在线 (stop 被拒), 推理按需可停止
3. 策略路由: 任务类型分布与 task_cycle 一致
4. 策略路由: 每条 policy 的 route 与 engines 一致 (order == engines)
5. 推理: 每事件都有推理结果, avg_chain_len >= 2 (读状态+预测)
6. 推理: 结论含「下一事件」或「基于世界模型」 (真实通路, 非写死)
7. 世界模型命中率在 [0,1] 且 > 0 (转移矩阵已学习)
8. 预测误差信号产生 (Phase 4 Learning 前置通路)
9. 行动占位: active_process 分区有 last_reasoning (不外发, record-only)
10. 主观时钟 = 事件数
"""
from __future__ import annotations
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
import yaml
from stellaris.main_phase3 import run_closed_loop
def test_phase3() -> int:
cfg = yaml.safe_load(
(Path(__file__).resolve().parent / "configs" / "engine_config.yaml").read_text(encoding="utf-8")
)
n = 10
s = run_closed_loop(cfg, n_events=n)
# 1. 六引擎全部启动
started_engines = [e for e in ("perception", "state", "memory", "world_model",
"reasoning", "cognitive_control")]
health_names = set(s["health"].keys())
for e in started_engines:
assert e in health_names, f"引擎 {e} 未启动 (health 缺失)"
# 2. 认知控制始终在线 (stop_all 后 world_model 可停, 但 control 拒 stop)
assert s["health"]["cognitive_control"]["healthy"] is True
assert s["health"]["reasoning"]["healthy"] is True
# 3. 策略任务类型分布 (task_cycle=[fact,causal,novel,fact] × n 事件)
# n=10: fact=5(0,4,8 位置? 不对, 循环 fact,causal,novel,fact 周期4)
# i%4: 0=fact,1=causal,2=novel,3=fact → 0..9: fact=5(0,3,4,7,8? 不对)
# 精确: i=0,1,2,3,4,5,6,7,8,9 → fact,causal,novel,fact,fact,causal,novel,fact,fact,causal
# fact=5(0,3,4,7,8? 让我数: i%4==0→fact(0,4,8)=3, i%4==3→fact(3,7)=2, 合计fact=5
# causal: i%4==1(1,5,9)=3, novel: i%4==2(2,6)=2
dist = s["policy_task_distribution"]
assert dist.get("fact") == 5, f"fact 策略数 {dist.get('fact')} != 5 (n={n})"
assert dist.get("causal") == 3, f"causal 策略数 {dist.get('causal')} != 3 (n={n})"
assert dist.get("novel") == 2, f"novel 策略数 {dist.get('novel')} != 2 (n={n})"
# 4. 策略 route == engines (order 与 engines 一致)
for p in s["policies"]:
assert p["route"] == p["engines"], f"策略 route {p['route']} != engines {p['engines']}"
# 5. 推理: 每事件都有结果
assert s["reasoning"]["n_results"] == n, "推理未逐事件生成"
assert s["reasoning"]["avg_chain_len"] >= 2, \
f"推理链平均 {s['reasoning']['avg_chain_len']} < 2 (应含读状态+预测)"
# 6. 推理结论非写死 (含真实通路关键词)
sample = s["reasoning"]["sample"]
for r in sample:
c = str(r["conclusion"])
assert ("下一事件" in c) or ("基于世界模型" in c) or ("基于" in c) \
or ("信息不足" in c), f"推理结论疑似写死: {c}"
# 7. 世界模型命中率合法且 > 0
hr = s["world_model"]["hit_rate"]
assert 0.0 < hr <= 1.0, f"世界模型命中率 {hr} 应 > 0 且 ≤ 1"
assert s["world_model"]["total_transitions"] >= n - 2, "世界模型转移数过少"
# 8. 预测误差信号
assert len(s["prediction_errors"]) == n - 1, "误差信号数量 = 事件数 - 1"
assert s["mean_prediction_error"] >= 0.0, "误差均值必须非负"
# 9. 行动占位: active_process 分区有 last_reasoning
ap = s["active_process"]
assert ap is not None, "active_process 分区未写入"
assert "last_reasoning" in ap, "行动占位缺少 last_reasoning"
assert ap.get("action_taken") == "record-only", "Phase 5 前应只记录不外发"
# 10. 主观时钟
assert s["t_subj"] == float(n), f"主观时钟 {s['t_subj']} != {n}"
print("== P8 Phase 3 闭环不变量: 10/10 通过 ==")
print(f" 策略分布={dist}, 推理 avg_conf={s['reasoning']['avg_confidence']}, "
f"世界模型命中率={hr}, 误差均值={s['mean_prediction_error']}")
return 0
if __name__ == "__main__":
sys.exit(test_phase3())