research(rectification): baseline runner never rewrites the published 09-14 report
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
This commit is contained in:
co-authored by
Claude Fable 5.1
parent
88dca263fe
commit
06ea1f4c6d
@@ -39,6 +39,7 @@ v5 已建:**77 例**(v4 20 例原样继承 + 57 例新增公开 Rodden-AA)
|
||||
- `scripts/research/holdout_v5_baseline.py`:把 `minute_resolution_sweep` / `cluster_width_probe` / `futile_collect_stop_replay`(经 `guided_collect_holdout_replay.load_cases`)的模块常量指到指定数据集与独立输出文件,只跑 `baseline` 变体;不改这三个脚本。
|
||||
- **硬红线 5**:`--dataset v4` 复跑,三个脚本共 7+12+6 组数字与 `minute_resolution_2026_09_14.json`、`cluster_width_2026_09_14.json`、`futile_collect_stop_replay_2026_09_29.json` 逐格一致(表在 `holdout_v5_build_2026_09_29.md` §3;JSON `holdout_v5_baseline_v4_all_2026_09_29.json`)。
|
||||
- v5 全集:`holdout_v5_baseline_v5_all_2026_09_29.json`,数字见构建说明 §4。
|
||||
- 发现并处理:`cluster_width_probe.run` 会顺手重写 `minute_resolution_2026_09_14.md` 里的勘误表(数值相同、格式 `15.0`→`15`)。已 `git checkout` 还原该文件,并在 `holdout_v5_baseline.py` 里把 `append_previous_errata` 置空,保证复跑不碰已发布文档。
|
||||
|
||||
| 半径 | 引擎先验头名(v4) | 六题回放头名(v4) | 六题后真值在区间(v4) | 六题后宽度中位(v4) | D1 六格真值在区间 truth / opposite |
|
||||
| --- | --- | --- | --- | --- | --- |
|
||||
|
||||
Reference in New Issue
Block a user