- commit
- 6e1fbb9
- parent
- 4785d11
- author
- im_wower
- date
- 2026-03-31 16:20:32 +0800 CST
phase6: 涌现结构分析+沉积窗口修复+归巢mu耦合 (42/42)
4 files changed,
+373,
-13
+2,
-0
1@@ -1,3 +1,5 @@
2 __pycache__/
3 *.pyc
4 .DS_Store
5+tests/cie_analysis.json
6+__pycache__/
+72,
-0
1@@ -180,3 +180,75 @@ d554217 branch-b: Phase 0-3 实现 + Phase 4 测试全过 (12/12)
2 4c794d4 docs: 锁定施工合同
3 28c1126 docs: CIE统一设计文档
4 ```
5+
6+---
7+
8+## Phase 6: 涌现结构分析 + Bug 修复
9+
10+> 2026-03-31 | 42/42 测试全绿
11+
12+### 修复内容
13+
14+1. **沉积 last_sed_hits 未初始化**: 空输入时 for 循环不执行导致 UnboundLocalError
15+2. **沉积滑动窗口**: 引入 prev_hits 跟踪，合并需要 recent_growth >= threshold
16+3. **归巢→mu 耦合**: 归巢力除了更新 phi，也微弱传播激活到能力核邻居
17+
18+### 涌现结构分析结果
19+
20+用 5 本真实课本（小学语文/数学、初中语文/数学、高中语文）灌入 CIE runtime 后分析图上涌现了什么。
21+
22+#### 全5本课本灌入同一 runtime
23+
24+| 指标 | 数值 |
25+|------|------|
26+| 节点数 | 758 |
27+| 边数 | 2,114 |
28+| Hub-高频字一致性 | 70% |
29+| 非对称边比例 | 100% |
30+| 闭环数 | 31 |
31+| 锚点核 | 398 个 |
32+| 能力核 | 3 个 (44 节点) |
33+| 沉积事件 | 1,452 |
34+| 合并事件 | 4,487 |
35+| 衰减事件 | 35,763 |
36+
37+#### Top Hub 节点（全语料）
38+
39+| 节点 | κ (路径汇聚度) | hits |
40+|------|---------------|------|
41+| (空格) | 185.96 | 48,692 |
42+| 的 | 35.50 | 7,664 |
43+| 学 | 31.75 | 6,578 |
44+| 一 | 17.49 | 4,533 |
45+| 。 | 16.88 | 4,152 |
46+
47+**结论：高频功能词（的、一、。）自然成为图上的 hub 节点，κ 最高。** 这不是预设的——系统从纯拓扑的流动中涌现出了"哪些字最重要"。
48+
49+#### 非对称环流（闭环/技能）
50+
51+Top 闭环 `上→册→ →上` 的正向环流 24.10，反向 2.11，非对称度 21.99。
52+
53+**结论：非对称权重产生了单向优势环流，验证了 README 的核心命题——非对称→极限环→技能。**
54+
55+#### 单本 vs 全灌对比
56+
57+| 课本 | Hub一致性 | 闭环数 | 能力核 |
58+|------|----------|--------|--------|
59+| 小学语文一上 | 50% | 31 | 1 |
60+| 小学数学一上 | 80% | 31 | 3 |
61+| 初中语文七上 | 60% | 31 | 1 |
62+| 初中数学七上 | 80% | 33 | 2 |
63+| 高中语文必修上 | 50% | 32 | 1 |
64+| **全灌** | **70%** | **31** | **3** |
65+
66+数学课本 hub 一致性更高（80%），因为数学文本词汇集中度更高。
67+
68+### 已知局限（Phase 7 待解决）
69+
70+1. **置信度未分化**: 所有节点 c=0.333（均匀先验），因为 `update_confidence` 总是更新 category 0。需要基于上下文自动选择 Dirichlet 分量。
71+
72+2. **锚点核过多**: 398/758 = 52% 的节点被判为锚点核。`anchor_epsilon=0.005` 阈值太宽松，需要动态调整。
73+
74+3. **能力核过少**: 只有 3 个能力核（44 节点），相对 758 节点来说太少。merge_threshold=60 可能过高。
75+
76+这三个问题互相关联——置信度不分化导致衰减率均匀化，导致锚点核阈值不精确，导致能力核检测不准。**修好 Dirichlet 分类是下一步的关键。**
+28,
-13
1@@ -233,9 +233,15 @@ class Dynamics:
2 # 长程归巢——拉向锚点核
3 pull2 = self.homing_lambda2 * (anchor_center_phi - phi_v)
4
5- # 更新
6+ # 更新 phi
7 self.state.phi[node_id] = phi_v + pull1 + pull2
8
9+ # 归巢也微弱影响 mu：向能力核方向的节点获得微量激活
10+ if nearest_core and nearest_core in self.state.ability_cores:
11+ for cn in list(self.state.ability_cores[nearest_core])[:3]:
12+ if cn != node_id:
13+ self.state.mu[cn] = self.state.mu.get(cn, 0.0) + abs(pull1) * 0.01
14+
15 if nearest_core:
16 self.state.bound_ability_core = nearest_core
17
18@@ -245,12 +251,17 @@ class Dynamics:
19 """
20 沉积路径：记忆层 → 经验层 → 技能带 → 能力核
21
22- 基于 experience_hits 检测：
23- - hits > sediment_threshold → 进入经验层
24- - hits > skill_belt_threshold → 成为技能带候选
25- - hits > merge_threshold → 并入能力核
26+ 使用 experience_hits 检测，但引入"最近沉积步数"避免饱和：
27+ 只有当 hits 在最近 window 步内增长了才记录新的 trace。
28 """
29+ window = 50 # 滑动窗口：最近50步内的新增才算
30+ last_sed_hits = getattr(self, "_last_sed_hits", {})
31+
32 for node_id, hits in list(self.state.experience_hits.items()):
33+ # 检查该节点上次沉积时的 hits
34+ prev_hits = last_sed_hits.get(node_id, 0)
35+ recent_growth = hits - prev_hits
36+
37 # 记忆层 → 经验层
38 if hits >= self.sediment_threshold:
39 if 'experience' not in self.state.experience_regions:
40@@ -263,12 +274,13 @@ class Dynamics:
41 'transition': 'memory -> experience',
42 'hits': hits,
43 })
44+ last_sed_hits[node_id] = hits
45
46- # 经验层 → 技能带候选
47+ # 经验层 → 技能带候选（需要持续增长）
48 if hits >= self.skill_belt_threshold:
49 old_score = self.state.skill_belt_candidates.get(node_id, 0.0)
50 new_score = hits / self.merge_threshold
51- if new_score > old_score:
52+ if new_score > old_score + 0.05: # 需要显著增长
53 self.state.skill_belt_candidates[node_id] = new_score
54 if old_score == 0.0:
55 self.state.sedimentation_trace.append({
56@@ -277,14 +289,13 @@ class Dynamics:
57 'transition': 'experience -> skill_belt',
58 'hits': hits,
59 })
60+ last_sed_hits[node_id] = hits
61
62- # 技能带 → 能力核
63- if hits >= self.merge_threshold:
64- # 找到或创建最近的能力核
65+ # 技能带 → 能力核（需要持续增长，且最近有新激活）
66+ if hits >= self.merge_threshold and recent_growth >= self.sediment_threshold:
67 merged = False
68 for core_id, core_nodes in self.state.ability_cores.items():
69- # 如果节点与现有能力核有边连接
70- for cn in core_nodes:
71+ for cn in list(core_nodes):
72 if (self.graph.get_edge_weight(node_id, cn) > 0 or
73 self.graph.get_edge_weight(cn, node_id) > 0):
74 core_nodes.add(node_id)
75@@ -300,7 +311,6 @@ class Dynamics:
76 break
77
78 if not merged:
79- # 创建新能力核
80 core_id = f'core_{len(self.state.ability_cores)}'
81 self.state.ability_cores[core_id] = {node_id}
82 self.state.merge_events.append({
83@@ -309,6 +319,11 @@ class Dynamics:
84 'core': core_id,
85 'transition': 'new_ability_core',
86 })
87+ last_sed_hits[node_id] = hits
88+
89+ if not hasattr(self, '_last_sed_hits'):
90+ self._last_sed_hits = {}
91+ self._last_sed_hits.update(last_sed_hits)
92
93 # ── 边流衰减 ──
94
+271,
-0
1@@ -0,0 +1,271 @@
2+"""
3+CIE 涌现结构分析
4+=================
5+用真实课本数据跑 runtime，分析图上涌现了什么结构。
6+
7+验证核心命题：
8+ - 词/属性/概念边界不是预制的，是流动反复重叠后涌现的稳定结构
9+ - 对称→不动点(知识)，非对称→极限环(技能)
10+ - 高频字成为 hub（路径汇聚度 κ 高）
11+ - 锚点核自动浮出
12+"""
13+
14+import sys
15+import os
16+import json
17+import math
18+from collections import Counter
19+
20+sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
21+from cie import CIERuntime
22+
23+DATA_DIR = "/Users/george/code/china-text-book-md"
24+
25+
26+def load_textbook_clean(filename):
27+ """加载课本，提取纯中文段落"""
28+ path = os.path.join(DATA_DIR, filename)
29+ with open(path, "r", encoding="utf-8") as f:
30+ raw = f.read()
31+ paragraphs = []
32+ for line in raw.split("\n"):
33+ line = line.strip()
34+ if not line or line.startswith("#") or line.startswith("**") or line.startswith("---"):
35+ continue
36+ if line.startswith("!["):
37+ continue
38+ ctrl = sum(1 for c in line if ord(c) < 32 and c not in '\n\t')
39+ if ctrl > len(line) * 0.3:
40+ continue
41+ cn = sum(1 for c in line if '\u4e00' <= c <= '\u9fff')
42+ if cn >= 2:
43+ paragraphs.append(line)
44+ return paragraphs
45+
46+
47+def analyze_graph(rt, title):
48+ """分析图的涌现结构"""
49+ g = rt.graph
50+ s = rt.state
51+
52+ result = {"title": title}
53+
54+ # ── 基本统计 ──
55+ result["nodes"] = g.node_count
56+ result["edges"] = g.edge_count
57+
58+ # ── Hub 节点：路径汇聚度 κ 最高的节点 ──
59+ kappas = {}
60+ for nid in g.nodes:
61+ kappas[nid] = g.convergence(nid)
62+ top_hubs = sorted(kappas.items(), key=lambda x: -x[1])[:20]
63+ result["top_hubs"] = [(nid, round(k, 2)) for nid, k in top_hubs]
64+
65+ # ── 高频字 vs Hub 的一致性 ──
66+ # 高频字应该成为 hub（κ 高）
67+ hits = sorted(s.experience_hits.items(), key=lambda x: -x[1])[:20]
68+ result["top_hits"] = [(nid, h) for nid, h in hits]
69+
70+ # hub 和高频字的重叠度
71+ hub_set = set(nid for nid, _ in top_hubs[:10])
72+ hit_set = set(nid for nid, _ in hits[:10])
73+ overlap = hub_set & hit_set
74+ result["hub_hit_overlap"] = len(overlap)
75+ result["hub_hit_overlap_pct"] = round(len(overlap) / max(len(hub_set), 1) * 100)
76+
77+ # ── 非对称度分布 ──
78+ asymmetries = []
79+ for src_edges in g.fwd_edges.values():
80+ for dst, edge in src_edges.items():
81+ bwd = g.get_bwd_weight(edge.src, dst)
82+ asym = abs(edge.weight - bwd)
83+ asymmetries.append(asym)
84+ if asymmetries:
85+ result["asym_mean"] = round(sum(asymmetries) / len(asymmetries), 4)
86+ result["asym_max"] = round(max(asymmetries), 4)
87+ result["asym_nonzero_pct"] = round(sum(1 for a in asymmetries if a > 0.01) / len(asymmetries) * 100)
88+
89+ # ── 闭环检测：找长度3-4的环并计算环流 ──
90+ circuits = []
91+ nodes_list = list(g.nodes.keys())
92+ checked = 0
93+ for i, a in enumerate(nodes_list[:50]): # 只检查前50个节点
94+ for b in g.neighbors_fwd(a)[:10]:
95+ for c in g.neighbors_fwd(b)[:10]:
96+ if a in g.neighbors_fwd(c):
97+ circ = g.circulation([a, b, c, a])
98+ rev_circ = g.circulation([a, c, b, a])
99+ if abs(circ - rev_circ) > 0.1: # 非对称环流
100+ circuits.append({
101+ 'path': f"{a}→{b}→{c}→{a}",
102+ 'circulation': round(circ, 2),
103+ 'reverse': round(rev_circ, 2),
104+ 'asymmetry': round(abs(circ - rev_circ), 2)
105+ })
106+ checked += 1
107+ if checked > 30:
108+ break
109+ if checked > 30:
110+ break
111+
112+ circuits.sort(key=lambda x: -x['asymmetry'])
113+ result["top_circuits"] = circuits[:10]
114+ result["total_circuits_found"] = len(circuits)
115+
116+ # ── 锚点核 ──
117+ result["anchor_count"] = len(s.anchor_nodes)
118+ result["anchor_nodes"] = list(s.anchor_nodes)[:20]
119+
120+ # ── 能力核 ──
121+ result["ability_cores"] = {
122+ cid: list(nodes)[:10] for cid, nodes in s.ability_cores.items()
123+ }
124+ result["ability_core_count"] = len(s.ability_cores)
125+ result["total_core_nodes"] = sum(len(n) for n in s.ability_cores.values())
126+
127+ # ── 经验层 / 技能带 ──
128+ exp = s.experience_regions.get('experience', set())
129+ result["experience_count"] = len(exp)
130+ result["skill_belt_count"] = len(s.skill_belt_candidates)
131+ result["top_skill_belts"] = sorted(
132+ s.skill_belt_candidates.items(), key=lambda x: -x[1])[:10]
133+
134+ # ── 置信度分布 ──
135+ confidences = [(nid, s.get_confidence(nid)) for nid in g.nodes]
136+ confidences.sort(key=lambda x: -x[1])
137+ result["top_confidence"] = [(nid, round(c, 3)) for nid, c in confidences[:10]]
138+ c_vals = [c for _, c in confidences]
139+ if c_vals:
140+ result["confidence_mean"] = round(sum(c_vals) / len(c_vals), 3)
141+ result["confidence_high_pct"] = round(sum(1 for c in c_vals if c > 0.5) / len(c_vals) * 100)
142+
143+ # ── φ 分布 ──
144+ phi_vals = sorted(s.phi.items(), key=lambda x: -abs(x[1]))
145+ result["top_phi"] = [(nid, round(v, 4)) for nid, v in phi_vals[:10]]
146+ result["phi_positive"] = sum(1 for _, v in phi_vals if v > 0.01)
147+ result["phi_negative"] = sum(1 for _, v in phi_vals if v < -0.01)
148+
149+ # ── 沉积统计 ──
150+ transitions = Counter(t['transition'] for t in s.sedimentation_trace)
151+ result["sedimentation_transitions"] = dict(transitions)
152+ result["total_sed_events"] = len(s.sedimentation_trace)
153+ result["total_merge_events"] = len(s.merge_events)
154+ result["total_decay_events"] = len(s.decay_events)
155+
156+ return result
157+
158+
159+def run_analysis():
160+ """对每本课本单独分析 + 全部灌入分析"""
161+
162+ textbooks = [
163+ ("小学语文一上", "小学_语文_统编版_义务教育教科书·语文一年级上册.md"),
164+ ("小学数学一上", "小学_数学_人教版_义务教育教科书 · 数学一年级上册.md"),
165+ ("初中语文七上", "初中_语文_统编版-人民教育出版社_七年级_义务教育教科书·语文七年级上册.md"),
166+ ("初中数学七上", "初中_数学_人教版-人民教育出版社_七年级_义务教育教科书·数学七年级上册.md"),
167+ ("高中语文必修上", "高中_语文_统编版-人民教育出版社_普通高中教科书·语文必修 上册.md"),
168+ ]
169+
170+ all_results = []
171+
172+ # ── 单本分析 ──
173+ for name, filename in textbooks:
174+ print(f"\n{'='*60}")
175+ print(f" 分析: {name}")
176+ print(f"{'='*60}")
177+
178+ rt = CIERuntime(seed=42)
179+ paras = load_textbook_clean(filename)
180+ print(f" 段落数: {len(paras)}")
181+
182+ # 喂入（前100段，每段截60字）
183+ feed_count = min(100, len(paras))
184+ for p in paras[:feed_count]:
185+ rt.ingest(p[:60])
186+ rt.step(n=2)
187+
188+ result = analyze_graph(rt, name)
189+ result["paragraphs_fed"] = feed_count
190+ all_results.append(result)
191+
192+ print_result(result)
193+
194+ # ── 全灌分析 ──
195+ print(f"\n{'='*60}")
196+ print(f" 分析: 全5本课本灌入同一 runtime")
197+ print(f"{'='*60}")
198+
199+ rt_all = CIERuntime(seed=42)
200+ total_fed = 0
201+ for name, filename in textbooks:
202+ paras = load_textbook_clean(filename)
203+ for p in paras[:60]:
204+ rt_all.ingest(p[:50])
205+ rt_all.step(n=2)
206+ total_fed += 1
207+
208+ result_all = analyze_graph(rt_all, "全5本课本")
209+ result_all["paragraphs_fed"] = total_fed
210+ all_results.append(result_all)
211+ print_result(result_all)
212+
213+ # ── 输出 JSON ──
214+ with open("/tmp/cie_analysis.json", "w") as f:
215+ json.dump(all_results, f, ensure_ascii=False, indent=2, default=str)
216+ print(f"\n\n完整数据已写入 /tmp/cie_analysis.json")
217+
218+ # ── 核心结论 ──
219+ print(f"\n{'='*60}")
220+ print(f" 核心结论")
221+ print(f"{'='*60}")
222+
223+ for r in all_results:
224+ hub_pct = r.get("hub_hit_overlap_pct", 0)
225+ asym_pct = r.get("asym_nonzero_pct", 0)
226+ circuits = r.get("total_circuits_found", 0)
227+ anchors = r.get("anchor_count", 0)
228+ cores = r.get("ability_core_count", 0)
229+
230+ print(f"\n {r['title']}:")
231+ print(f" Hub=高频字一致性: {hub_pct}%")
232+ print(f" 非对称边比例: {asym_pct}%")
233+ print(f" 闭环数: {circuits}")
234+ print(f" 锚点核: {anchors} 个")
235+ print(f" 能力核: {cores} 个")
236+
237+ return all_results
238+
239+
240+def print_result(r):
241+ print(f"\n 图规模: {r['nodes']} nodes, {r['edges']} edges")
242+
243+ print(f"\n Top Hub 节点 (κ 最高):")
244+ for nid, k in r['top_hubs'][:10]:
245+ h = dict(r['top_hits']).get(nid, 0)
246+ print(f" '{nid}': κ={k}, hits={h}")
247+
248+ print(f"\n Hub-高频字重叠: {r['hub_hit_overlap']}/10 ({r['hub_hit_overlap_pct']}%)")
249+
250+ if r.get('asym_nonzero_pct') is not None:
251+ print(f" 非对称边: {r['asym_nonzero_pct']}%, mean={r['asym_mean']}, max={r['asym_max']}")
252+
253+ if r['top_circuits']:
254+ print(f"\n Top 闭环 (非对称环流):")
255+ for c in r['top_circuits'][:5]:
256+ print(f" {c['path']}: fwd={c['circulation']}, rev={c['reverse']}, asym={c['asymmetry']}")
257+
258+ print(f" 闭环总数: {r['total_circuits_found']}")
259+ print(f" 锚点核: {r['anchor_count']} 个 — {r['anchor_nodes'][:10]}")
260+ print(f" 能力核: {r['ability_core_count']} 个, 含 {r['total_core_nodes']} 节点")
261+
262+ print(f"\n 沉积统计: {r['sedimentation_transitions']}")
263+ print(f" 总沉积事件: {r['total_sed_events']}, 合并: {r['total_merge_events']}, 衰减: {r['total_decay_events']}")
264+
265+ print(f"\n Top 置信度:")
266+ for nid, c in r['top_confidence'][:5]:
267+ print(f" '{nid}': c={c}")
268+ print(f" 高置信度(>0.5)占比: {r.get('confidence_high_pct', 0)}%")
269+
270+
271+if __name__ == '__main__':
272+ run_analysis()