测试结果总表

Harness 启动

• Node v24.20.0 LTS
• npm 11.19.0
• pnpm 11.25.0
• npx 路径卡住
• pnpm dlx @deepseek-ai/dsh@latest web 成功
• Web UI：127.0.0.1:3080

Workspace Write

• harness-test.md 写入/读取成功
• 说明 model → tools → Windows file write/read 链路成立

Test Fixture

已知：

• task_count = 15
• planned = 1380
• actual = 1410
• workflow optimization occurrences = 4
• high algebra occurrences = 3
• missing Time = 11

Skill-only 测试

发现：

• task_count 被写成 14
• 高等代数被写成 2 结论：
• Grounding/Skill 本身不足以保证 deterministic counting

weekly_validator.py

通过：

• task_count 15
• high algebra 3/3
• workflow occurrences 4
• missing Time 11
• bad path exit 1

End-to-End

Skill → Validator → JSON → Raw Markdown → Model Analysis 通过。

Experiment Governance

proposal → approval → running → completed → kept → promotion eligibility → test rule 所有 Gate 按用户授权执行。

Real Collector

范围：2026-08-27~2026-08-30

• project_files = 3
• task_files = 8
• daily_files = 12
• daily_files_in_period = 2
• try_events_in_period = 1
• missing Daily = 08-27, 08-29
• PyYAML date 对象问题已修复

Real Validator

• validator_status = passed_with_warnings
• expected_days = 4
• present_days = 2
• missing_days = 2
• try_event_count = 1
• task_count = 8
• source consistency 4/4 match
• warnings = 5
• errors = 0

Real Dry Run

通过：

• 08-27 Daily missing + Try evidence 被同时保留
• 08-29 保持 Unknown
• Growth Archive 未当作当周事实
• Git 不可用时跳过
• 无正式写入

暴露：

1. UPDATE 压缩漏 08-27 Try
2. Task date 被误推成“已经可以开始”
3. “系统记录未闭环”被错标为 Fact

v0.3.2 Grounding Regression

• Test 1 Summary/UPDATE Compression: PASS
• Test 2 Unknown Field Semantics: PASS
• Test 3 Evidence Label Discipline: PASS
• Overall: PASS

当前结论

Weekly Review Harness Core Phase 1 完成。 Architecture Freeze。 下一阶段：First Real Baseline Weekly。