tianxuan / stellaris /test_closed_loop_phase2.py
masterofthestars's picture
Tianxuan: full project + data (checkpoints, world_model, data without raw) with correct directory structure
db7bc06
Raw History Blame Contribute Delete
4.22 kB
#!/usr/bin/env python3
"""
群星 P8 · Phase 2 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_phase2
断言 (对应文档 4.3 记忆 + 4.4 世界模型 + 9 认知循环):
1. 记忆引擎始终在线 (stop 被拒)
2. 世界模型引擎可停止 (按需唤醒, 非始终在线)
3. 每事件都编码进记忆 (memory_count == n_events)
4. 世界模型有非零转移数 (update 真实发生)
5. 世界模型命中率在 [0,1] 区间
6. 预测误差信号产生且非空 (Phase 4 Learning 的前置通路)
7. 主观时钟 = 事件数 (文档 SubjectiveClock)
8. 记忆衰减公式: 高意义记忆 (s=0.9) 衰减慢于低意义 (s=0.1)
9. 回忆重演: 检索命中后 d0 增强
"""
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 import run_closed_loop
from stellaris.core.memory.memory_engine import MemoryEngine
from stellaris.core.state.state_engine import StateEngine
from stellaris.core.world_model.world_model_engine import WorldModelEngine
from stellaris.core.engines.interface import EngineRegistry
from stellaris.core.blackboard import Blackboard
def test_phase2() -> int:
cfg = yaml.safe_load(
(Path(__file__).resolve().parent / "configs" / "engine_config.yaml").read_text(encoding="utf-8")
)
s = run_closed_loop(cfg, n_events=10)
# 1 & 2: 始终在线 / 按需唤醒
assert s["stopped"]["memory"] is False, "记忆引擎应始终在线 (文档 4.3)"
assert s["stopped"]["world_model"] is True, "世界模型引擎应按需可停止 (文档 4.4)"
assert s["stopped"]["perception"] is False, "感知引擎应始终在线"
assert s["stopped"]["state"] is False, "状态引擎应始终在线"
# 3: 每事件都编码
assert s["memory_count"] == s["n_events"] == 10, f"记忆数 {s['memory_count']} != 事件数 {s['n_events']}"
# 4: 世界模型有真实转移
assert s["world_model"]["total_transitions"] >= 8, \
f"世界模型转移数 {s['world_model']['total_transitions']} 过少 (n_events=10 应有 ≥8)"
# 5: 命中率合法
hr = s["world_model"]["hit_rate"]
assert 0.0 <= hr <= 1.0, f"世界模型命中率 {hr} 越界"
# 6: 误差信号通路
assert len(s["prediction_errors"]) == s["n_events"] - 1, "误差信号数量 = 事件数 - 1"
assert all(e >= 0.0 for e in s["prediction_errors"]), "误差信号必须非负"
# 7: 主观时钟
assert s["t_subj"] == 10.0, f"主观时钟 t_subj={s['t_subj']} 应等于事件数 10"
# 8: 记忆衰减 (高意义慢衰减, 走黑板托管主观时间)
z = Blackboard()
mem = MemoryEngine({}, blackboard=z)
state = StateEngine({})
state.init({})
z.write_state(state.get())
high_id = mem.encode({"content": "x", "s": 0.9, "d0": 0.5, "source": "test"})
low_id = mem.encode({"content": "y", "s": 0.1, "d0": 0.5, "source": "test"})
# 推进 20 个主观时刻 (每步刷新黑板)
for _ in range(20):
state.advance(event_occurred=True)
z.write_state(state.get())
high_detail = mem._memories[high_id].decayed_detail(state.t_subj)
low_detail = mem._memories[low_id].decayed_detail(state.t_subj)
assert high_detail > low_detail, \
f"高意义记忆衰减后细节 {high_detail:.4f} 应 > 低意义 {low_detail:.4f}"
assert high_detail < 0.5, "高意义记忆 20 步后细节应 < 初始 0.5 (仍在衰减)"
# 9: 回忆重演增强 (无黑板, t_subj=0, 检查 retrieve 后 d0 增加)
mem2 = MemoryEngine({})
mid = mem2.encode({"content": "z", "s": 0.5, "d0": 0.5, "source": "test"})
d0_before = mem2._memories[mid].d0
mem2.retrieve(query="z", top_k=1)
d0_after = mem2._memories[mid].d0
assert d0_after > d0_before, f"回忆重演后 d0 {d0_after} 应 > 检索前 {d0_before}"
print("== P8 Phase 2 闭环不变量: 9/9 通过 ==")
print(f" 世界模型命中率={hr}, 转移数={s['world_model']['total_transitions']}, "
f"误差均值={s['mean_prediction_error']}")
return 0
if __name__ == "__main__":
sys.exit(test_phase2())