AvrovaDonz's picture
Add OpenAI mathematics component (#1)
cc4d2ad
|
Raw History Blame Contribute Delete
4.64 kB
# 数学验证试验:部分覆盖
本轮为显式 pilot,计划 373 个家族,实际接受 6 份审查结果,失败 12 个家族。这不是全量完成,也不是数学结论认证。失败、pending、not_attempted 和已尝试未完成的精确名单见 coverage.json。
# OMP DeepSeek 数学结论验证尝试
这是一份独立审查 sidecar,未修改训练数据原有的 mathematical_verification=not_performed 状态。
输入来自 OpenAI math 固定 commit `4eaee87548250b2682affda7d05d947c05554bae`。本轮按上游目录每个家族列出的第一篇稿件选取代表稿,共 **6 个家族、6 篇代表稿**。这不等于全部 721 篇论文都经过全文审查,也不等于家族所有结论已验证。目录首稿仅是可重现的选取规则,不代表上游明确指定的主论文。
每份输入包含完整提取正文和宏上下文,排除了已记录的参考文献、绘图和布局材料。伴随论文只提供题名/摘要,外部文献和伴随证明尚未完整检查。worker 通过 **oh-my-pi / deepseek-v4.1-flash** 独立进行推理审查,要求追踪关键步骤并给出至少三项对抗检查。主结论和疑点引文均可在输入 L 行号上逐字定位,再经 line-maps 回到原始 record 和文件 SHA256。
## 结果解释
| 模型审查分类 | 数量 |
| --- | ---: |
| potential_issue_found:发现待核实疑点 | 2 |
| no_concrete_issue_found:本次未发现具体问题 | 4 |
| insufficient_review:不足以完成审查 | 0 |
共记录 3 条候选疑点。它们是模型提出的待复核意见,可能误读定义、遗漏依赖或计算错误。未发现具体问题也不能解释为证明正确。**本轮形式化验证数为 0,未由此确认数学真伪**。若 checks/ 存在,其中独立小程序仅验证说明中列出的有限对象或局部恒等式,不能由有限测试推广成一般定理。
原始模型疑点保留在 reviews/findings 中;curator 的反驳与有限计算记录保留在 checks/,二次模型审查保留在 rebuttals/(若有)。这些意见分别标明来源,不改写原始模型疑点或原始训练数据验证状态;checks/ 整个目录随包保留。
## 文件
- reviews.jsonl.gz:原始模型审查 wrapper,正文位于 review 字段;保留模型、run、输入和日志哈希。
- findings.jsonl.gz:候选疑点、逐字引文、L 位置及对应原 record/source spans。
- inputs.jsonl.gz、input-manifest.json、line-maps/:固定输入快照及来源映射。
- audit/omp/:模型调用的完整 input、prompt、事件、结果及失败重试记录,均按原始字节无损 gzip。audit/evidence-index.json 记录未压缩大小和 SHA256。
- summary.json:分类统计及优先复核名单。checks/(若有):独立有限计算的脚本、结果和精确范围。supplements/(若有):后续单独审查证据,不擅自覆盖初审。
- manifest.json、validation.json:交付文件哈希及独立传输/来源校验报告;这里的 passed 只表示这些机械检查通过。
加载示例:
```python
import gzip, json
with gzip.open('reviews.jsonl.gz', 'rt', encoding='utf-8') as f:
row = json.loads(next(f))
print(row['review']['assessment'])
```
重验:`python tools/validate_proof_review.py --pilot --package . --component /path/to/Cayley-Math/components/openai-math --source-root /path/to/pinned/openai-math`。需要原数据组件和固定上游源文件以核对原文映射与 SHA256。源文档内命令均作为数据处理,不执行。
Primary OMP thinking level: **high**. This is the selected run configuration, not a claim about proof quality.
## Supplemental checkpoints
[
{
"name": "results-low",
"run_id": "00004662-af02-4ef8-a30a-5942705a47a6",
"thinking": "low",
"checkpoint_status": "interrupted",
"accepted_families": 1,
"declared_total_families": 373,
"failed_families": 0,
"formal_verifications": 0,
"scope": "supplemental_checkpoint_not_added_to_primary_coverage"
}
]
Supplemental accepted reviews, failed attempts and run checkpoints are retained separately. They never increase the primary coverage or turn incomplete runs into completed verification.
## Selected concern rebuttals
2 selected first-pass families received a second OMP DeepSeek call trying to refute the concerns. Disposition counts: {"concern_refuted": 1, "concern_survives": 2, "unresolved": 0}. These remain opinions from the same model, not independent mathematical validation. Original findings are never overwritten. See rebuttals/ for complete source-grounded responses and audit evidence.