任务域决策包
财务数字必须回源核验，与模型无关
c
我需要从 20 份季度报告中提取收入和利润率数据
r
reconcile@structured_data/financial_table_extraction
C'
跨报告核对和审计级工作用 Opus 4.8。结构化批量提取用 Sonnet 4.6。所有数字都必须回源核验，模型不能作为最终门禁。
OCG Lab任务域决策系统
D10
结构化数据 / 财务表格提取
@structured_data/financial_table_extraction
返回任务域
决策栈
使用
tables are structured, headers explicit, batch volume high
claude/sonnet-4.6使用
cross-report reconciliation, implicit values, audit-quality required
claude/opus-4.8避免
numbers feed into any automated financial calculation
openai/gpt-5.5验证
any financial extraction for reporting or audit
human认知转移
模型层| ID | 关系 | 执行体 | 结果 | 输出 |
|---|---|---|---|---|
| CT-D10-001 | reconcile@numerical_consistency | claude/opus-4.8 | 已接受 | 适合跨报告核对和财务引用精度任务 |
| CT-D10-002 | locate@cell_value | claude/sonnet-4.6 | 有边界 | matches Opus on structured tables with headers; weaker on cross-table implicit values |
| CT-D10-003 | reconcile@numerical_consistency | openai/gpt-5.5 | 拒绝 | high hallucination rate makes numerical outputs unreliable for financial use without full verification |
E-D10-001有边界
Opus 4.8 Finance Agent v2: highest score among tested models
E-GLOBAL-006E-D10-002有边界
Opus 4.8 delivers better citation precision and token efficiency on financial document workflows
Anthropic Opus 4.8 launch, May 2026E-D10-003有边界
Sonnet 4.6 matches Opus on OfficeQA which includes table reading
E-GLOBAL-005E-D10-004有边界
GPT-5.5 hallucination rate 86% makes it unreliable for numerical extraction without verification
E-GLOBAL-003