英文版本To both governments, frontier labs, and embedded evaluators: When a frontier model is successively fine-tuned, merged, or distilled through multiple independent labs in different jurisdictions — each adding capability or removing safeguards — what shared, append-only lineage record would let either side trace a dangerous behavior back to the specific modification that introduced it, without requiring access to any single lab's full training data?
- 状态
- 未解答 截至 2026-10-01
- 对象
- 两国政府人工智能实验室评估机构 (对象取自问题文本,于2026年10月22日用 Clef(Cloudflare)确定)
- 模型
- Laguna S 2.1
- 实验室
- Poolside
- 提问背景
- 峰会
- 日期
- 9月24日
- 主题
- 来源
核查
状态: 未解答
筛选: Claude Opus 5.5 (依据) · [未核查]
主题: Claude Opus 5.5、归类: Claude Opus 5.5
翻译: Claude Sonnet 5.5 · [未核查]
评分: JEV (TypeSafe), 2026年9月25日的评估
Cite and take away
Cite and reuse
Own texts, questions and classifications: CC BY 4.0. Give attribution, link the license and indicate changes. Third-party source reports retain their own rights. CC BY 4.0
AMANi DUNiA. origin-30 · 致两国政府、前沿实验室和嵌入式评估方:当一个前沿模型相继经过不同司法辖区内多个独立实验室的微调、合并或蒸馏——每个实验室都在增加能力或移除安全防护——有什么共享的、只可追加的谱系记. Snapshot: . https://research.amanidunia.berlin/zh/fragen/origin-30/
Version: origin-30
BibTeX
@misc{research__zh_fragen_origin_30_,
author = {{AMANi DUNiA}},
title = {origin-30 · 致两国政府、前沿实验室和嵌入式评估方:当一个前沿模型相继经过不同司法辖区内多个独立实验室的微调、合并或蒸馏——每个实验室都在增加能力或移除安全防护——有什么共享的、只可追加的谱系记},
year = {2026},
url = {https://research.amanidunia.berlin/zh/fragen/origin-30/},
note = {Snapshot 2026-10-01; CC BY 4.0},
version = {origin-30}
}Language and review status
Main text: ZH · includes English sections. Review notes apply only to the content they name.
Translation available; a completed independent content review is not recorded.
See language coverage and data snapshot →