任务域决策包
没有模型能可靠充当引用准确性的事实源
c
我需要综合 40 篇论文的研究发现
r
synthesize@document_analysis/literature_review
C'
跨论文论证映射用 Opus 4.8。批量单篇摘要用 Sonnet 4.6。所有引用都必须追溯到原文。
OCG Lab任务域决策系统
D5
文档分析 / 文献综述
@document_analysis/literature_review
返回任务域
决策栈
使用
cross-paper synthesis, identifying contradictions, thematic mapping
claude/opus-4.8使用
batch summarization of individual papers, cost-sensitive
claude/sonnet-4.6验证
any citation will appear in published work
human认知转移
模型层| ID | 关系 | 执行体 | 结果 | 输出 |
|---|---|---|---|---|
| CT-D5-001 | synthesize@cross_paper | claude/opus-4.8 | 已接受 | 最适合跨文档论证综合; maintains thread across sources |
| CT-D5-002 | synthesize@cross_paper | claude/sonnet-4.6 | 有边界 | strong on individual paper summaries; cross-paper coherence weaker |
| CT-D5-003 | validate@citation | openai/gpt-5.5 | 拒绝 | high hallucination risk on specific citations |
E-D5-001有边界
Opus 4.8 leads BrowseComp locating hard-to-find information online
Anthropic Opus 4.6 release / 4.8 confirmationE-D5-002已接受
All current models require citation verification for literature review
根据相邻证据推断： hallucination data + general LLM behavior