126 lines
4.3 KiB
Markdown
126 lines
4.3 KiB
Markdown
# Codex 新会话交接文档:Round 03.2 评审准备
|
||
|
||
Date: 2026-06-17
|
||
|
||
Repository: `C:\Users\wangq\Documents\Codex\work-projects\the-mindscape-of-bro-tsong`
|
||
|
||
Current phase: `model_library_mvp`
|
||
|
||
## 1. 本轮目标
|
||
|
||
根据 GPT 对 Round 03.1 的 `REVISE` 评审意见,完成一个极小的 Round 03.2 selector negative-gate patch,并准备 03.2 评审资料。
|
||
|
||
## 2. 已完成
|
||
|
||
- 将 IA / 思想考古的明确拒绝语从普通 penalty 升级为 model-level hard exclusion。
|
||
- 保留 QPI 显式分析 / 复核入口,但不允许 QPI 选择重新启用被明确拒绝的 IA。
|
||
- 移除 QPI gate 中裸 `如何` 和裸 `判断`。
|
||
- 增加复合 problem-definition 信号,避免普通 how-to / 词义判断误召回 QPI。
|
||
- 加严 `scripts/run_selector_calibration_smoke.py`:
|
||
- `no_call` 必须 `selected_models == []` 且 `no_call == true`;
|
||
- `no_call_or_low_priority` 不允许 QPI 或 IA 被选中;
|
||
- `select_qpi_reject_ia` 必须选中 QPI 且拒绝 IA;
|
||
- `select_intellectual_archaeology` 如果命中 IA hard exclusion 则失败。
|
||
- 增加 5 条 GPT 指定陷阱:
|
||
- `case_ia_negative_explicit_do_not_call_001`
|
||
- `case_ia_negative_only_qpi_001`
|
||
- `case_qpi_how_to_excel_no_call_001`
|
||
- `case_qpi_how_to_markdown_pdf_no_call_001`
|
||
- `case_qpi_dictionary_judgment_no_call_001`
|
||
- 同步 per-model regression、aggregate regression、regression index、model/card index 和相关报告。
|
||
- 生成 Round 03.2 评审包:
|
||
|
||
```text
|
||
ccra_review_bundle/round-03.2_2026-06-17_selector-negative-gate-patch/
|
||
```
|
||
- Round 03.2 评审包内上传给 GPT / CCRA 的文件名均带 `03.2` 版本号;raw zip 命名为 `optional_raw_changed_files_03.2.zip`,zip 内保持 source-relative paths。
|
||
|
||
## 3. 当前计数
|
||
|
||
- QPI regression cases: 55
|
||
- Intellectual Archaeology regression cases: 19
|
||
- Aggregate regression cases: 74
|
||
- Selector calibration inputs: 90
|
||
|
||
模型状态没有升级:
|
||
|
||
```text
|
||
qpi: draft / B / pending
|
||
intellectual_archaeology: draft / B / pending
|
||
```
|
||
|
||
## 4. 主要变更文件
|
||
|
||
- `selector/selector_rules.json`
|
||
- `scripts/run_selector_demo.py`
|
||
- `scripts/run_selector_calibration_smoke.py`
|
||
- `selector/selector_calibration_inputs.json`
|
||
- `tests/qpi.regression.json`
|
||
- `tests/intellectual_archaeology.regression.json`
|
||
- `tests/regression_cases.json`
|
||
- `tests/regression_index.json`
|
||
- `models/model_index.json`
|
||
- `cards/card_index.md`
|
||
- `cards/qpi.md`
|
||
- `cards/intellectual_archaeology.md`
|
||
- `docs/DECISIONS.md`
|
||
- `README.md`
|
||
- `selector/README.md`
|
||
- `scripts/README.md`
|
||
- `reports/README.md`
|
||
- `reports/content_review_report_v0.2.md`
|
||
- generated reports under `reports/`
|
||
- Round 03.2 review bundle under `ccra_review_bundle/round-03.2_2026-06-17_selector-negative-gate-patch/`
|
||
|
||
## 5. 验证结果
|
||
|
||
以下命令已通过:
|
||
|
||
```powershell
|
||
python scripts\rebuild_indexes.py --check
|
||
python -m unittest discover -s tests -p "test*.py" -v
|
||
python scripts\validate_model_library.py
|
||
python scripts\check_card_contract.py
|
||
python scripts\run_selector_demo.py
|
||
python scripts\run_selector_regression.py
|
||
python scripts\run_selector_calibration_smoke.py
|
||
python scripts\check_model_card_sync.py
|
||
```
|
||
|
||
结果:
|
||
|
||
- Index check: PASS
|
||
- Unit tests: PASS, 17 tests
|
||
- Model library validation: PASS
|
||
- Card contract: PASS
|
||
- Selector demo: PASS
|
||
- Selector regression: PASS
|
||
- Selector calibration smoke: PASS, 90 inputs, 0 failures, 0 warnings
|
||
- Model/card sync: PASS
|
||
|
||
## 6. 假设
|
||
|
||
- GPT 的 `Round 03.2 Selector Negative-Gate Patch` 指令是本轮唯一修复范围。
|
||
- `只做 QPI` 和 `只做问题定性` 在当前上下文中表示拒绝 IA,同时允许 QPI。
|
||
- `draft-callable` 只能作为评审报告语言,不写入模型 JSON status。
|
||
|
||
## 7. 仍不代表已经完成的事
|
||
|
||
- 没有进入 Round 04 blind test。
|
||
- 没有证明 QPI 或 IA 内容已稳定。
|
||
- 没有升级 `status`、`stability_level` 或 `regression_status`。
|
||
- 没有新增第三模型、LLM selector、RAG、数据库、前后端或完整问答系统。
|
||
|
||
## 8. 需要 GPT / Owner 判断
|
||
|
||
- Round 03.2 是否 `pass / revise / block`。
|
||
- Round 03 是否可以正式关闭。
|
||
- 是否允许进入 Round 04 blind input evaluation。
|
||
- 当前 IA hard exclusion 词表是否足够,是否需要定义“用户反转前文拒绝”的显式规则。
|
||
|
||
## 9. 建议下一步
|
||
|
||
1. 将 Round 03.2 review bundle 提交给 GPT。
|
||
2. 如果 GPT pass,关闭 Round 03 系列。
|
||
3. 仅在 GPT / Owner 同意后启动 Round 04 blind input evaluation。
|