- commit
- 70cfdf3
- parent
- fabbb69
- author
- codex@macbookpro
- date
- 2026-04-01 11:21:57 +0800 CST
integration phase2.3: validation/context/attention sanity cleanup
6 files changed,
+414,
-114
+28,
-6
1@@ -4,11 +4,11 @@
2 > 决策依据: Branch A 6/8, Branch B 8/8 (同一 cie-datasets formal dataset)
3 > 策略: 方案 C — 核取 B，壳取 A
4 >
5-> 审核更新（2026-04-01）：
6-> `origin/integration/merge-plan @ 423d812` 已落地到 Phase 2，
7-> 但 **Phase 3 前必须暂停**，先通过 Phase 2.1 gate：
8-> PendingSignal 单路径收口、移除旧直接回灌路径、补 exactly-once 回归测试、
9-> 并明确 pytest gate 与 formal validation script gate 的区别。
10+> 状态更新（2026-04-01）：
11+> `origin/integration/merge-plan @ fabbb69` 已完成 Phase 2.2（Kernel Sanity Gate）；
12+> `integration/phase2.3-sanity-cleanup` 已完成 Phase 2.3（validation/context/attention sanity cleanup），
13+> 并通过 quick gate、Phase 2.3 gate、broader regression gate、formal validation smoke。
14+> **Phase 3 尚未开始**；只有在 Phase 2.3 gates 全通过后，才允许进入 Phase 3。
15
16 ---
17
18@@ -136,7 +136,29 @@ reset_session()
19 `python3 tests/formal_validation.py`
20 6. 只有 Phase 2.1 通过，后续实现才允许进入 Phase 3
21
22-### Phase 3: 沉积 Profile 并行观测（Day 2）
23+### Phase 2.2: Kernel Sanity Gate（已完成）
24+
25+1. 修正 Branch B 内核中 backward-weight 读取方向
26+2. 确认 Dirichlet per-node wiring 不再退化成均匀先验占位
27+3. 确认 context 参数已被 runtime 真实消费
28+4. 补 attention ledger 的基础 sanity gate
29+5. 通过 gate：`python3 -m pytest tests/test_kernel_sanity.py -q`
30+
31+### Phase 2.3: Validation / Context / Attention Sanity Cleanup（已完成）
32+
33+1. 修正 validation/test 层残留的 stale `get_bwd_weight(...)` 调用顺序
34+2. 明确 context 语义：
35+ `str` 按字符拆分；`list/tuple` 逐项消费；`dict` 稳定归一化；不支持类型显式报错
36+3. 加强 attention ledger gate，覆盖 multi-step / decay / feedback / reset cleanup 场景
37+4. 将 `tests/test_comprehensive.py` 与 `tests/formal_validation.py` 改成 repo-relative / env-aware 路径解析
38+5. 通过 gate：
39+ `python3 -m pytest tests/test_smoke.py tests/test_dynamics.py tests/test_exactly_once.py tests/test_kernel_sanity.py -q`
40+ `python3 -m pytest tests/test_phase23_validation_sanity.py -q`
41+ `python3 -m pytest tests/test_smoke.py tests/test_dynamics.py tests/test_exactly_once.py tests/test_kernel_sanity.py tests/test_comprehensive.py tests/test_phase23_validation_sanity.py -q`
42+ `python3 tests/formal_validation.py`
43+6. Phase 2.3 完成后，Phase 3 才允许开始；当前 **Phase 3 仍未开始**
44+
45+### Phase 3: 沉积 Profile 并行观测（未开始，Day 2）
46
47 1. 从 Branch A 移植 `SedimentationProfile` dataclass 到 `cie/state.py`
48 2. 保留 Branch B 的 `experience_hits` 作为已验证行为基线；先把 Profile 做成 dual-write / parallel observation
+39,
-11
1@@ -97,26 +97,33 @@ class Dynamics:
2
3 spread_ratio = 0.3 # 每步传播 30% 的激活
4 spread_amount = mu_v * spread_ratio
5+ released = self.state.attention.release(node_id, spread_amount)
6+ if released <= 1e-10:
7+ continue
8+ moved = 0.0
9
10 for nb in neighbors:
11 w = self.graph.get_edge_weight(node_id, nb)
12- flow = spread_amount * (w / total_weight)
13- new_mu[nb] = new_mu.get(nb, 0.0) + flow
14+ desired_flow = released * (w / total_weight)
15+ actual_flow = self.state.attention.allocate(nb, desired_flow)
16+ if actual_flow <= 1e-10:
17+ continue
18+ new_mu[nb] = new_mu.get(nb, 0.0) + actual_flow
19 # 更新边流 J
20 self.state.J[(node_id, nb)] = (
21- self.state.J.get((node_id, nb), 0.0) * 0.9 + flow
22+ self.state.J.get((node_id, nb), 0.0) * 0.9 + actual_flow
23 )
24- # attention accounting: 传播到邻居的激活也要记账
25- self.state.attention.allocate(nb, flow)
26+ moved += actual_flow
27 # 记录经验命中
28 self.state.experience_hits[nb] = (
29 self.state.experience_hits.get(nb, 0) + 1
30 )
31 self.state.active_region.add(nb)
32
33- new_mu[node_id] = mu_v - spread_amount
34- # attention accounting: 释放从源节点流出的激活
35- self.state.attention.release(node_id, spread_amount)
36+ leftover = max(0.0, released - moved)
37+ if leftover > 1e-10:
38+ self.state.attention.allocate(node_id, leftover)
39+ new_mu[node_id] = max(0.0, new_mu.get(node_id, mu_v) - moved)
40
41 self.state.mu.update(new_mu)
42 # 清理极低激活
43@@ -250,9 +257,30 @@ class Dynamics:
44
45 # 归巢也微弱影响 mu：向能力核方向的节点获得微量激活
46 if nearest_core and nearest_core in self.state.ability_cores:
47- for cn in list(self.state.ability_cores[nearest_core])[:3]:
48- if cn != node_id:
49- self.state.mu[cn] = self.state.mu.get(cn, 0.0) + abs(pull1) * 0.01
50+ targets = [cn for cn in list(self.state.ability_cores[nearest_core])[:3] if cn != node_id]
51+ if targets:
52+ desired_total = min(
53+ self.state.mu.get(node_id, 0.0),
54+ abs(pull1) * 0.01 * len(targets),
55+ )
56+ released = self.state.attention.release(node_id, desired_total)
57+ moved = 0.0
58+ if released > 1e-10:
59+ per_target = released / len(targets)
60+ for cn in targets:
61+ actual = self.state.attention.allocate(cn, per_target)
62+ if actual <= 1e-10:
63+ continue
64+ self.state.mu[cn] = self.state.mu.get(cn, 0.0) + actual
65+ self.state.active_region.add(cn)
66+ moved += actual
67+ leftover = max(0.0, released - moved)
68+ if leftover > 1e-10:
69+ self.state.attention.allocate(node_id, leftover)
70+ self.state.mu[node_id] = max(
71+ 0.0,
72+ self.state.mu.get(node_id, 0.0) - moved,
73+ )
74
75 if nearest_core:
76 self.state.bound_ability_core = nearest_core
+48,
-13
1@@ -6,7 +6,6 @@ CIE Runtime — 统一接口（SPEC §5）
2 """
3
4 import random
5-import math
6 from typing import Optional
7 from .graph import Graph
8 from .state import CIEState, PendingSignal
9@@ -40,13 +39,14 @@ class CIERuntime:
10 """
11 接收新的输入。创建 PendingSignal 排队，在下一步 step 时消费。
12 PendingSignal 是唯一的状态写入入口。
13+
14+ context 语义在 Phase 2.3 明确化：
15+ - str: 按字符拆成 context tokens
16+ - list/tuple: 逐项转成 token
17+ - dict: 按 key 排序后稳定归一化为 "key=value" token
18+ - 其他非 None 类型: 显式报 TypeError，避免静默忽略
19 """
20- if isinstance(input_data, str):
21- tokens = list(input_data)
22- elif isinstance(input_data, (list, tuple)):
23- tokens = list(input_data)
24- else:
25- tokens = [str(input_data)]
26+ tokens = self._normalize_input_tokens(input_data)
27
28 anchor_tokens = []
29 if anchors:
30@@ -55,12 +55,7 @@ class CIERuntime:
31 elif isinstance(anchors, (list, tuple)):
32 anchor_tokens = list(anchors)
33
34- context_tokens = []
35- if context:
36- if isinstance(context, str):
37- context_tokens = list(context)[:10]
38- elif isinstance(context, (list, tuple)):
39- context_tokens = list(context)[:10]
40+ context_tokens = self._normalize_context_tokens(context)
41
42 signal = PendingSignal(
43 source="external",
44@@ -72,6 +67,46 @@ class CIERuntime:
45 )
46 self.state.pending_signals.append(signal)
47
48+ def _normalize_input_tokens(self, input_data):
49+ if isinstance(input_data, str):
50+ return list(input_data)
51+ if isinstance(input_data, (list, tuple)):
52+ return list(input_data)
53+ return [str(input_data)]
54+
55+ def _normalize_context_tokens(self, context):
56+ if context is None:
57+ return []
58+ if isinstance(context, str):
59+ return list(context)[:10]
60+ if isinstance(context, (list, tuple)):
61+ return [str(item) for item in list(context)[:10]]
62+ if isinstance(context, dict):
63+ tokens = []
64+ for key in sorted(context.keys(), key=lambda item: str(item)):
65+ value = context[key]
66+ if isinstance(value, str):
67+ normalized = value[:16]
68+ elif isinstance(value, (int, float, bool)) or value is None:
69+ normalized = str(value)
70+ elif isinstance(value, (list, tuple)):
71+ preview = ",".join(str(item)[:8] for item in list(value)[:3])
72+ normalized = f"[{preview}]"
73+ elif isinstance(value, dict):
74+ normalized = "dict"
75+ else:
76+ normalized = type(value).__name__
77+
78+ token = f"{str(key)[:12]}={normalized}"
79+ if token:
80+ tokens.append(token[:32])
81+ if len(tokens) >= 10:
82+ break
83+ return tokens
84+ raise TypeError(
85+ "context must be one of: str, list, tuple, dict, or None"
86+ )
87+
88 def step(self, n: int = 1):
89 """
90 推进 n 步内部动力学演化。
+151,
-80
1@@ -1,88 +1,159 @@
2-import sys, os, json, math, time
3-sys.path.insert(0, "/Users/george/code/CIE-Unified")
4+import json
5+import math
6+import os
7+import sys
8+import time
9+from pathlib import Path
10+
11+REPO_ROOT = Path(__file__).resolve().parents[1]
12+if str(REPO_ROOT) not in sys.path:
13+ sys.path.insert(0, str(REPO_ROOT))
14+
15 from cie import CIERuntime
16
17-DS = "/Users/george/code/cie-datasets/china_text_book_md/v2026-03-28"
18
19-def load_recs(stage, subject, n=80):
20- path = os.path.join(DS, "splits", "by_stage_subject", stage, f"{subject}.jsonl")
21+def resolve_dataset_dir() -> Path:
22+ env_path = os.environ.get("CIE_FORMAL_DATASET_DIR")
23+ candidates = []
24+ if env_path:
25+ candidates.append(Path(env_path))
26+ candidates.append(REPO_ROOT.parent / "cie-datasets" / "china_text_book_md" / "v2026-03-28")
27+
28+ for candidate in candidates:
29+ if candidate.is_dir():
30+ return candidate
31+
32+ raise FileNotFoundError(
33+ "Formal validation dataset not found. Set CIE_FORMAL_DATASET_DIR or place "
34+ "the dataset at ../cie-datasets/china_text_book_md/v2026-03-28 relative "
35+ "to the repo root."
36+ )
37+
38+
39+def load_recs(dataset_dir: Path, stage: str, subject: str, n: int = 80) -> list[str]:
40+ path = dataset_dir / "splits" / "by_stage_subject" / stage / f"{subject}.jsonl"
41 recs = []
42- if not os.path.exists(path): return recs
43- with open(path) as f:
44+ if not path.exists():
45+ return recs
46+ with open(path, encoding="utf-8") as f:
47 for line in f:
48 rec = json.loads(line)
49- if not rec.get("is_content"): continue
50- t = rec.get("text","")
51- if len(t) >= 4: recs.append(t)
52- if len(recs) >= n: break
53+ if not rec.get("is_content"):
54+ continue
55+ text = rec.get("text", "")
56+ if len(text) >= 4:
57+ recs.append(text)
58+ if len(recs) >= n:
59+ break
60 return recs
61
62-combos = [("小学","语文"),("小学","数学"),("初中","语文"),("初中","数学"),("高中","语文")]
63-tests = []
64-
65-# Pipeline + Stability per combo
66-for stage, subj in combos:
67- label = stage + subj
68- recs = load_recs(stage, subj, 80)
69- if not recs:
70- tests.append({"name": label, "status": "FAIL", "detail": "no data"})
71- continue
72- rt = CIERuntime(seed=42)
73- t0 = time.time()
74- for r in recs:
75- rt.ingest(r[:60])
76- rt.step(n=1)
77- elapsed = time.time() - t0
78- ot = rt.emit()
79- if ot["activated"]:
80- rt.commit_feedback({"correct": [ot["activated"][0]["node"]], "reward": 1.0})
81- snap = rt.snapshot_state()
82- ok = snap["phi_summary"]["count"] > 20 and abs(snap["phi_summary"]["max"]) <= 10.1
83- ok &= snap["attention"]["used"] <= snap["attention"]["total"] + 0.01
84- ok &= all(math.isfinite(v) for v in rt.state.phi.values())
85-
86- # Emergence: top words
87- g = rt.graph
88- cn_bg = []
89- for se in g.fwd_edges.values():
90- for dst, edge in se.items():
91- if "\u4e00" <= edge.src <= "\u9fff" and "\u4e00" <= dst <= "\u9fff":
92- bwd = g.get_bwd_weight(edge.src, dst)
93- ratio = edge.weight / bwd if bwd > 0.01 else edge.weight * 100
94- cn_bg.append((edge.src+dst, round(ratio,1)))
95- cn_bg.sort(key=lambda x: -x[1])
96-
97- d = f"n={snap['phi_summary']['count']}, e={snap['graph']['edge_count']}, phi=[{snap['phi_summary']['min']:.3f},{snap['phi_summary']['max']:.3f}], mode={ot['mode']}, t={elapsed:.1f}s, words={cn_bg[:5]}"
98- tests.append({"name": label, "status": "PASS" if ok else "FAIL", "detail": d})
99-
100-# Cross-stage
101-rt2 = CIERuntime(seed=42)
102-for stage in ["小学","初中","高中"]:
103- for r in load_recs(stage, "语文", 30):
104- rt2.ingest(r[:50])
105- rt2.step(n=1)
106-s2 = rt2.snapshot_state()
107-ok2 = abs(s2["phi_summary"]["max"]) <= 10.1 and s2["phi_summary"]["count"] > 30
108-tests.append({"name": "cross_stage", "status": "PASS" if ok2 else "FAIL",
109- "detail": f"n={s2['phi_summary']['count']}, phi={s2['phi_summary']['max']:.3f}"})
110-
111-# Cross-subject
112-rt3 = CIERuntime(seed=42)
113-for subj in ["语文","数学","科学"]:
114- for r in load_recs("小学", subj, 30):
115- rt3.ingest(r[:50], anchors=[subj])
116- rt3.step(n=1)
117-s3 = rt3.snapshot_state()
118-ok3 = abs(s3["phi_summary"]["max"]) <= 10.1
119-tests.append({"name": "cross_subject", "status": "PASS" if ok3 else "FAIL",
120- "detail": f"n={s3['phi_summary']['count']}, phi={s3['phi_summary']['max']:.3f}, cores={len(rt3.state.ability_cores)}"})
121-
122-# Summary
123-passed = sum(1 for t in tests if t["status"]=="PASS")
124-failed = sum(1 for t in tests if t["status"]=="FAIL")
125-for t in tests:
126- print(f"[{t['status']}] {t['name']}: {t['detail']}")
127-print(f"\n总计: {passed}/{len(tests)} PASS")
128-
129-with open("/tmp/formal_val_results.json", "w") as f:
130- json.dump({"tests": tests, "summary": {"passed": passed, "failed": failed, "total": len(tests), "dataset": DS}}, f, ensure_ascii=False, indent=2, default=str)
131+
132+def run_validation(dataset_dir: Path | None = None) -> dict:
133+ dataset_dir = Path(dataset_dir) if dataset_dir is not None else resolve_dataset_dir()
134+
135+ combos = [("小学", "语文"), ("小学", "数学"), ("初中", "语文"), ("初中", "数学"), ("高中", "语文")]
136+ tests = []
137+
138+ # Pipeline + Stability per combo
139+ for stage, subj in combos:
140+ label = stage + subj
141+ recs = load_recs(dataset_dir, stage, subj, 80)
142+ if not recs:
143+ tests.append({"name": label, "status": "FAIL", "detail": "no data"})
144+ continue
145+
146+ rt = CIERuntime(seed=42)
147+ t0 = time.time()
148+ for rec in recs:
149+ rt.ingest(rec[:60])
150+ rt.step(n=1)
151+ elapsed = time.time() - t0
152+
153+ output = rt.emit()
154+ if output["activated"]:
155+ rt.commit_feedback({"correct": [output["activated"][0]["node"]], "reward": 1.0})
156+ snap = rt.snapshot_state()
157+
158+ ok = snap["phi_summary"]["count"] > 20 and abs(snap["phi_summary"]["max"]) <= 10.1
159+ ok &= snap["attention"]["used"] <= snap["attention"]["total"] + 0.01
160+ ok &= all(math.isfinite(v) for v in rt.state.phi.values())
161+
162+ g = rt.graph
163+ cn_bg = []
164+ for src_edges in g.fwd_edges.values():
165+ for dst, edge in src_edges.items():
166+ if "\u4e00" <= edge.src <= "\u9fff" and "\u4e00" <= dst <= "\u9fff":
167+ bwd = g.get_bwd_weight(dst, edge.src)
168+ ratio = edge.weight / bwd if bwd > 0.01 else edge.weight * 100
169+ cn_bg.append((edge.src + dst, round(ratio, 1)))
170+ cn_bg.sort(key=lambda item: -item[1])
171+
172+ detail = (
173+ f"n={snap['phi_summary']['count']}, e={snap['graph']['edge_count']}, "
174+ f"phi=[{snap['phi_summary']['min']:.3f},{snap['phi_summary']['max']:.3f}], "
175+ f"mode={output['mode']}, t={elapsed:.1f}s, words={cn_bg[:5]}"
176+ )
177+ tests.append({"name": label, "status": "PASS" if ok else "FAIL", "detail": detail})
178+
179+ # Cross-stage
180+ rt2 = CIERuntime(seed=42)
181+ for stage in ["小学", "初中", "高中"]:
182+ for rec in load_recs(dataset_dir, stage, "语文", 30):
183+ rt2.ingest(rec[:50])
184+ rt2.step(n=1)
185+ snap2 = rt2.snapshot_state()
186+ ok2 = abs(snap2["phi_summary"]["max"]) <= 10.1 and snap2["phi_summary"]["count"] > 30
187+ tests.append({
188+ "name": "cross_stage",
189+ "status": "PASS" if ok2 else "FAIL",
190+ "detail": f"n={snap2['phi_summary']['count']}, phi={snap2['phi_summary']['max']:.3f}",
191+ })
192+
193+ # Cross-subject
194+ rt3 = CIERuntime(seed=42)
195+ for subj in ["语文", "数学", "科学"]:
196+ for rec in load_recs(dataset_dir, "小学", subj, 30):
197+ rt3.ingest(rec[:50], anchors=[subj])
198+ rt3.step(n=1)
199+ snap3 = rt3.snapshot_state()
200+ ok3 = abs(snap3["phi_summary"]["max"]) <= 10.1
201+ tests.append({
202+ "name": "cross_subject",
203+ "status": "PASS" if ok3 else "FAIL",
204+ "detail": (
205+ f"n={snap3['phi_summary']['count']}, phi={snap3['phi_summary']['max']:.3f}, "
206+ f"cores={len(rt3.state.ability_cores)}"
207+ ),
208+ })
209+
210+ passed = sum(1 for test in tests if test["status"] == "PASS")
211+ failed = sum(1 for test in tests if test["status"] == "FAIL")
212+ return {
213+ "tests": tests,
214+ "summary": {
215+ "passed": passed,
216+ "failed": failed,
217+ "total": len(tests),
218+ "dataset": str(dataset_dir),
219+ },
220+ }
221+
222+
223+def main() -> int:
224+ try:
225+ results = run_validation()
226+ except FileNotFoundError as exc:
227+ print(f"[FAIL] dataset: {exc}")
228+ return 1
229+
230+ for test in results["tests"]:
231+ print(f"[{test['status']}] {test['name']}: {test['detail']}")
232+ print(f"\n总计: {results['summary']['passed']}/{results['summary']['total']} PASS")
233+
234+ with open("/tmp/formal_val_results.json", "w", encoding="utf-8") as f:
235+ json.dump(results, f, ensure_ascii=False, indent=2, default=str)
236+ return 0 if results["summary"]["failed"] == 0 else 1
237+
238+
239+if __name__ == "__main__":
240+ sys.exit(main())
+20,
-4
1@@ -24,12 +24,28 @@ import time
2 import math
3 import json
4 import traceback
5+from pathlib import Path
6
7-sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
8+REPO_ROOT = Path(__file__).resolve().parents[1]
9+sys.path.insert(0, str(REPO_ROOT))
10 from cie import CIERuntime
11
12 # ── 数据路径 ──
13-DATA_DIR = "/Users/george/code/china-text-book-md"
14+def resolve_textbook_dir() -> Path:
15+ env_path = os.environ.get("CIE_TEXTBOOK_DATA_DIR")
16+ candidates = []
17+ if env_path:
18+ candidates.append(Path(env_path))
19+ candidates.append(REPO_ROOT.parent / "china-text-book-md")
20+
21+ for candidate in candidates:
22+ if candidate.is_dir():
23+ return candidate
24+
25+ raise FileNotFoundError(
26+ "Textbook dataset not found. Set CIE_TEXTBOOK_DATA_DIR or place "
27+ "the dataset at ../china-text-book-md relative to the repo root."
28+ )
29
30 TEXTBOOKS = {
31 "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",
32@@ -42,7 +58,7 @@ TEXTBOOKS = {
33
34 def load_textbook(name):
35 """加载课本，提取纯文本段落（跳过 markdown 标记和乱码）"""
36- path = os.path.join(DATA_DIR, TEXTBOOKS[name])
37+ path = resolve_textbook_dir() / TEXTBOOKS[name]
38 with open(path, "r", encoding="utf-8") as f:
39 raw = f.read()
40
41@@ -154,7 +170,7 @@ def test_A04_chuzhong_shuxue_formula():
42 total_edges = 0
43 for src_edges in rt.graph.fwd_edges.values():
44 for dst, edge in src_edges.items():
45- bwd_w = rt.graph.get_bwd_weight(edge.src, dst)
46+ bwd_w = rt.graph.get_bwd_weight(dst, edge.src)
47 if abs(edge.weight - bwd_w) > 0.01:
48 asym_count += 1
49 total_edges += 1
+128,
-0
1@@ -0,0 +1,128 @@
2+import math
3+from importlib import import_module
4+from pathlib import Path
5+
6+import pytest
7+
8+from cie import CIERuntime
9+
10+
11+REPO_ROOT = Path(__file__).resolve().parents[1]
12+
13+
14+def assert_attention_matches_mu(rt: CIERuntime, tol: float = 1e-6):
15+ mu_positive = {
16+ node: value for node, value in rt.state.mu.items()
17+ if value > tol
18+ }
19+ attn_positive = {
20+ node: value for node, value in rt.state.attention.allocated.items()
21+ if value > tol
22+ }
23+
24+ assert set(mu_positive) == set(attn_positive), (
25+ f"attention/mu node mismatch: mu_only={sorted(set(mu_positive) - set(attn_positive))}, "
26+ f"attn_only={sorted(set(attn_positive) - set(mu_positive))}"
27+ )
28+
29+ for node, mu_value in mu_positive.items():
30+ assert math.isclose(mu_value, attn_positive[node], rel_tol=1e-9, abs_tol=tol), (
31+ f"ledger mismatch at {node!r}: mu={mu_value}, attention={attn_positive[node]}"
32+ )
33+
34+ assert math.isclose(
35+ sum(mu_positive.values()),
36+ rt.state.attention.used,
37+ rel_tol=1e-9,
38+ abs_tol=tol,
39+ ), f"total mismatch: mu={sum(mu_positive.values())}, attention={rt.state.attention.used}"
40+
41+
42+def test_validation_call_sites_use_correct_backward_direction():
43+ targets = [
44+ REPO_ROOT / "tests" / "test_comprehensive.py",
45+ REPO_ROOT / "tests" / "formal_validation.py",
46+ ]
47+
48+ for path in targets:
49+ text = path.read_text(encoding="utf-8")
50+ assert "get_bwd_weight(edge.src, dst)" not in text, f"stale reversed call remains in {path.name}"
51+ assert "get_bwd_weight(dst, edge.src)" in text, f"corrected call missing in {path.name}"
52+
53+
54+def test_dict_context_is_normalized_and_consumed_deterministically():
55+ rt = CIERuntime(seed=42)
56+ context = {
57+ "topic": "背景",
58+ "stage": "小学",
59+ "tags": ["A", "B"],
60+ "meta": {"nested": True},
61+ }
62+
63+ rt.ingest("你好", context=context)
64+ pending = rt.state.pending_signals[-1]
65+ assert pending.context_tokens == [
66+ "meta=dict",
67+ "stage=小学",
68+ "tags=[A,B]",
69+ "topic=背景",
70+ ]
71+
72+ rt.step(n=1)
73+ for token in pending.context_tokens:
74+ assert rt.graph.has_node(token), f"context token missing from graph: {token}"
75+
76+ assert rt.graph.get_edge_weight("stage=小学", "你") > 0
77+ assert rt.graph.get_edge_weight("topic=背景", "好") > 0
78+
79+
80+def test_unsupported_context_type_fails_fast():
81+ rt = CIERuntime(seed=42)
82+ with pytest.raises(TypeError, match="context must be one of"):
83+ rt.ingest("你好", context=object())
84+
85+
86+def test_attention_ledger_matches_mu_across_steps_feedback_and_cleanup():
87+ rt = CIERuntime(seed=42)
88+
89+ for idx in range(6):
90+ rt.ingest(
91+ f"轮次{idx}",
92+ context={"stage": "小学", "round": idx},
93+ anchors=["锚点"],
94+ )
95+ rt.step(n=3)
96+ assert_attention_matches_mu(rt)
97+
98+ output = rt.emit()
99+ assert output is not None
100+ rt.step(n=2)
101+ assert_attention_matches_mu(rt)
102+
103+ rt.commit_feedback({"correct": list(rt.state.active_region)[:2], "reward": 1.0})
104+ rt.step(n=2)
105+ assert_attention_matches_mu(rt)
106+
107+ rt.step(n=25)
108+ assert_attention_matches_mu(rt)
109+
110+ rt.reset_session()
111+ assert_attention_matches_mu(rt)
112+ assert rt.state.attention.used == 0.0
113+ assert all(value == 0.0 for value in rt.state.mu.values())
114+
115+
116+def test_validation_path_helpers_are_env_aware(monkeypatch, tmp_path):
117+ textbook_dir = tmp_path / "textbooks"
118+ textbook_dir.mkdir()
119+ formal_dir = tmp_path / "formal"
120+ formal_dir.mkdir()
121+
122+ monkeypatch.setenv("CIE_TEXTBOOK_DATA_DIR", str(textbook_dir))
123+ monkeypatch.setenv("CIE_FORMAL_DATASET_DIR", str(formal_dir))
124+
125+ comprehensive = import_module("tests.test_comprehensive")
126+ formal_validation = import_module("tests.formal_validation")
127+
128+ assert comprehensive.resolve_textbook_dir() == textbook_dir
129+ assert formal_validation.resolve_dataset_dir() == formal_dir