专题研修
实际结果 · 记录与轨迹全部展开
49项离线测试通过，0失败、0错误、0跳过。CPU小模型训练另外执行。商业模型、GPU基准、生产安全审计未执行。下面是实际报告，不是预期值。
原始JSON / 测试日志 / 训练JSON / 如何解释这些结果
本章目录
执行环境与验证范围
{
"python": "3.13.5",
"platform": "Linux-6.18.55-x86_64-with-glibc2.41",
"test_count": 49,
"failures": 0,
"errors": 0,
"skipped": 0,
"seconds": 2.300336534000053,
"live_llm": "NOT_RUN",
"gpu_benchmark": "NOT_RUN",
"production_security_audit": "NOT_RUN",
"scope": "机制/协议/真实本地文件及子进程；所有任务与数值数据为作者构建"
}真实小模型CPU训练
{
"kind": "真实小型模型CPU训练；合成数据；未运行商业模型或GPU吞吐基准",
"torch_version": "2.10.0+cpu",
"parameter_count": 18320,
"steps": 100,
"initial_train_loss": 2.846790313720703,
"final_train_loss": 0.003446191782131791,
"train_token_accuracy": 1.0,
"causal_prefix_error": 0.0,
"elapsed_seconds": 0.3893101959999967,
"history": [
{
"step": 0,
"loss": 2.846790313720703
},
{
"step": 20,
"loss": 0.1237737238407135
},
{
"step": 40,
"loss": 0.013883101753890514
},
{
"step": 60,
"loss": 0.006276607979089022
},
{
"step": 80,
"loss": 0.004390936344861984
}
],
"limitation": "测试集泛化尚未评价；高训练准确率不是通用语言理解"
}数值前向与采样
{
"kind": "数值机制实验；非商用模型评测",
"cache_max_error": 1.5543122344752192e-15,
"wrong_rope_offset_error": 3.1744545678365306,
"online_softmax_error": 3.552713678800501e-15,
"speculative": {
"empirical": [
0.5978666666666667,
0.30333333333333334,
0.0988
],
"target": [
0.6,
0.3,
0.1
],
"acceptance": 0.6037333333333333,
"trials": 30000
},
"byte_bpe_roundtrip": "工程lower"
}解析反向与有限差分
{
"kind": "解析反向与有限差分；非训练能力/非GPU性能评测",
"max_absolute_errors": {
"dQ": 9.143890158935797e-11,
"dK": 1.9576201770732382e-10,
"dV": 1.1711515091050728e-10
}
}权限检索
{
"kind": "权限过滤先于评分的真实词项检索",
"result": [
[
"current",
0.8630462173553426
]
],
"denied": [],
"symbols": [
{
"name": "choose",
"kind": "FunctionDef",
"start": 1,
"end": 2
}
],
"rrf": [
[
"current",
0.03252247488101534
],
[
"old",
0.03252247488101534
]
]
}上下文与记忆
{
"kind": "确定性元数据约束，不证明自然语言语义保真",
"pack": {
"selected": [
"definition",
"policy",
"usage"
],
"cost": 70,
"budget": 75
},
"historically_unknown": {
"status": "not_found",
"claims": []
},
"now_known": {
"status": "found",
"claims": [
"repository-relative"
]
}
}SSE协议分片
{
"kind": "协议字节分片实验，非真实网络请求",
"calls": [
{
"id": "c1",
"name": "read_file",
"args": {
"path": "中文.py"
}
}
]
}合成统计实验
{
"kind": "合成评测数据",
"bootstrap": {
"difference": 0.2222222222222222,
"interval_95": [
-1.850371707708594e-17,
0.3333333333333333
],
"resampling_unit": "task"
},
"pass_at_3": 0.5333333333333333,
"wilson_8_of_10": [
0.49016247153665027,
0.943317848545623
],
"mcnemar_b1_c4": 0.375,
"independent_p08_all3": 0.5120000000000001
}贯穿案例：实际差异
--- a/matcher.py
+++ b/matcher.py
@@ -1,7 +1,22 @@
from pathlib import PurePosixPath
def should_exclude(path, patterns):
- # 契约：路径相对仓库根；末尾/代表目录及后代；其他项精确匹配。
- # 历史错误：只比较basename，导致nested路径失配，也误排除同名文件。
- path = PurePosixPath(path).name
- return path in patterns
+ if not isinstance(path, str) or not all(isinstance(p, str) for p in patterns):
+ raise TypeError("path和patterns必须是字符串")
+ normalized = path.replace("\\", "/")
+ parts = PurePosixPath(normalized).parts
+ if PurePosixPath(normalized).is_absolute() or ".." in parts:
+ raise ValueError("只支持仓库相对路径，拒绝父级跳转")
+ value = str(PurePosixPath(normalized))
+ for pattern in patterns:
+ rule = pattern.replace("\\", "/")
+ rule_parts = PurePosixPath(rule).parts
+ if PurePosixPath(rule).is_absolute() or ".." in rule_parts:
+ raise ValueError("排除规则不能越界")
+ if rule.endswith("/"):
+ prefix = str(PurePosixPath(rule))
+ if value == prefix or value.startswith(prefix + "/"):
+ return True
+ elif value == str(PurePosixPath(rule)):
+ return True
+ return False
贯穿案例：实际调用、stdout与退出码
步骤1
{
"call": {
"id": "t0",
"name": "test",
"args": {}
},
"result": {
"version": "9889901cd19183eb5a6fcdb45fd82538ad8e85ed15c768678af5dfe7ea1a6b4d",
"passed": false,
"exit_code": 1,
"log": "F..FFF.F\n======================================================================\nFAIL: test_directory (__main__.MatchingTests.test_directory)\n----------------------------------------------------------------------\nTraceback (most recent call last):\n File \"test_matcher.py\", line 7, in test_directory\n def test_directory(self): self.assertTrue(f(\"vendor/a.py\",[\"vendor/\"]))\n ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\nAssertionError: False is not true\n\n======================================================================\nFAIL: test_nested (__main__.MatchingTests.test_nested)\n----------------------------------------------------------------------\nTraceback (most recent call last):\n File \"test_matcher.py\", line 5, in test_nested\n def test_nested(self): self.assertTrue(f(\"src/a.py\",[\"src/a.py\"]))\n ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^\nAssertionError: False is not true\n\n======================================================================\nFAIL: test_parent (__main__.MatchingTests.test_parent)\n----------------------------------------------------------------------\nTraceback (most recent call last):\n File \"test_matcher.py\", line 12, in test_parent\n with self.assertRaises(ValueError): f(\"../a.py\",[\"a.py\"])\n ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^\nAssertionError: ValueError not raised\n\n======================================================================\nFAIL: test_same_name (__main__.MatchingTests.test_same_name)\n----------------------------------------------------------------------\nTraceback (most recent call last):\n File \"test_matcher.py\", line 6, in test_same_name\n def test_same_name(self): self.assertFalse(f(\"src/a.py\",[\"a.py\"]))\n ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^\nAssertionError: True is not false\n\n======================================================================\nFAIL: test_windows (__main__.MatchingTests.test_windows)\n----------------------------------------------------------------------\nTraceback (most recent call last):\n File \"test_matcher.py\", line 9, in test_windows\n def test_windows(self): self.assertTrue(f(\"src\\\\a.py\",[\"src/a.py\"]))\n ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\nAssertionError: False is not true\n\n----------------------------------------------------------------------\nRan 8 tests in 0.001s\n\nFAILED (failures=5)\n"
}
}步骤2
{
"call": {
"id": "r1",
"name": "read",
"args": {
"path": "matcher.py"
}
},
"result": {
"text": "from pathlib import PurePosixPath\n\ndef should_exclude(path, patterns):\n # 契约：路径相对仓库根；末尾/代表目录及后代；其他项精确匹配。\n # 历史错误：只比较basename，导致nested路径失配，也误排除同名文件。\n path = PurePosixPath(path).name\n return path in patterns\n",
"version": "9889901cd19183eb5a6fcdb45fd82538ad8e85ed15c768678af5dfe7ea1a6b4d"
}
}步骤3
{
"call": {
"id": "p1",
"name": "replace",
"args": {
"expected": "[源码内容]",
"new": "[源码内容]"
}
},
"result": {
"status": "applied",
"wrote": true,
"version": "383da239221fd2a5d1d19cfc232ebe6403e3730c9ed9ae15f1f58322a9e8b769"
}
}步骤4
{
"call": {
"id": "t1",
"name": "test",
"args": {}
},
"result": {
"version": "383da239221fd2a5d1d19cfc232ebe6403e3730c9ed9ae15f1f58322a9e8b769",
"passed": true,
"exit_code": 0,
"log": "........\n----------------------------------------------------------------------\nRan 8 tests in 0.000s\n\nOK\n"
}
}全部测试名称与结果
test_all_gradients (test_suite.AttentionBackwardTests.test_all_gradients) ... ok
test_future_gradient_zero (test_suite.AttentionBackwardTests.test_future_gradient_zero) ... ok
test_mask_validation (test_suite.AttentionBackwardTests.test_mask_validation) ... ok
test_upstream_shape (test_suite.AttentionBackwardTests.test_upstream_shape) ... ok
test_cycle (test_suite.ContextTests.test_cycle) ... ok
test_dependency (test_suite.ContextTests.test_dependency) ... ok
test_memory (test_suite.ContextTests.test_memory) ... ok
test_memory_conflict (test_suite.ContextTests.test_memory_conflict) ... ok
test_missing_dependency (test_suite.ContextTests.test_missing_dependency) ... ok
test_required_overflow (test_suite.ContextTests.test_required_overflow) ... ok
test_stale_evidence (test_suite.ContextTests.test_stale_evidence) ... ok
test_unbacked_claim (test_suite.ContextTests.test_unbacked_claim) ... ok
test_all_masked_rejected (test_suite.NumericsTests.test_all_masked_rejected) ... ok
test_bpe_merge_shortens (test_suite.NumericsTests.test_bpe_merge_shortens) ... ok
test_bpe_roundtrip (test_suite.NumericsTests.test_bpe_roundtrip) ... ok
test_cache_chunk (test_suite.NumericsTests.test_cache_chunk) ... ok
test_cache_token (test_suite.NumericsTests.test_cache_token) ... ok
test_causal (test_suite.NumericsTests.test_causal) ... ok
test_dpo_gradient (test_suite.NumericsTests.test_dpo_gradient) ... ok
test_gradient (test_suite.NumericsTests.test_gradient) ... ok
test_gradient_sum_zero (test_suite.NumericsTests.test_gradient_sum_zero) ... ok
test_negative_mask_detected (test_suite.NumericsTests.test_negative_mask_detected) ... ok
test_online (test_suite.NumericsTests.test_online) ... ok
test_rope_norm (test_suite.NumericsTests.test_rope_norm) ... ok
test_softmax_stable (test_suite.NumericsTests.test_softmax_stable) ... ok
test_speculative_distribution (test_suite.NumericsTests.test_speculative_distribution) ... ok
test_wrong_offset_detected (test_suite.NumericsTests.test_wrong_offset_detected) ... ok
test_failed_response (test_suite.ProtocolTests.test_failed_response) ... ok
test_id_conflict (test_suite.ProtocolTests.test_id_conflict) ... ok
test_not_executed_on_delta (test_suite.ProtocolTests.test_not_executed_on_delta) ... ok
test_truncated_sse (test_suite.ProtocolTests.test_truncated_sse) ... ok
test_utf8_sse (test_suite.ProtocolTests.test_utf8_sse) ... ok
test_capstone_actual_patch (test_suite.RecoveryTests.test_capstone_actual_patch) ... ok
test_changed_after_applied (test_suite.RecoveryTests.test_changed_after_applied) ... ok
test_id_reuse (test_suite.RecoveryTests.test_id_reuse) ... ok
test_outside_root (test_suite.RecoveryTests.test_outside_root) ... ok
test_real_crash_and_reconcile (test_suite.RecoveryTests.test_real_crash_and_reconcile) ... ok
test_stale_precondition (test_suite.RecoveryTests.test_stale_precondition) ... ok
test_acl (test_suite.RetrievalTests.test_acl) ... ok
test_ast (test_suite.RetrievalTests.test_ast) ... ok
test_empty (test_suite.RetrievalTests.test_empty) ... ok
test_principal_cache_isolation (test_suite.RetrievalTests.test_principal_cache_isolation) ... ok
test_rrf_duplicate (test_suite.RetrievalTests.test_rrf_duplicate) ... ok
test_update_invalidates (test_suite.RetrievalTests.test_update_invalidates) ... ok
test_version (test_suite.RetrievalTests.test_version) ... ok
test_mcnemar_zero (test_suite.StatisticsTests.test_mcnemar_zero) ... ok
test_paired_identity (test_suite.StatisticsTests.test_paired_identity) ... ok
test_pass_at_k (test_suite.StatisticsTests.test_pass_at_k) ... ok
test_wilson (test_suite.StatisticsTests.test_wilson) ... ok
----------------------------------------------------------------------
Ran 49 tests in 2.298s
OK