@structured_data/financial_table_extraction
numerical verification against source is mandatory regardless of model
I need to extract revenue and margin data from 20 quarterly reports
reconcile@structured_data/financial_table_extraction
Opus 4.8 for cross-report reconciliation and audit-quality work. Sonnet 4.6 for structured batch extraction. All numbers must be verified against source; no model is the final gate.
tables are structured, headers explicit, batch volume high
claude/sonnet-4.6cross-report reconciliation, implicit values, audit-quality required
claude/opus-4.8numbers feed into any automated financial calculation
openai/gpt-5.5any financial extraction for reporting or audit
human| ID | Relation | Actor | Result | Output |
|---|---|---|---|---|
| CT-D10-001 | reconcile@numerical_consistency | claude/opus-4.8 | accepted | appropriate for cross-report reconciliation and financial citation precision |
| CT-D10-002 | locate@cell_value | claude/sonnet-4.6 | bounded | matches Opus on structured tables with headers; weaker on cross-table implicit values |
| CT-D10-003 | reconcile@numerical_consistency | openai/gpt-5.5 | rejected | high hallucination rate makes numerical outputs unreliable for financial use without full verification |
Opus 4.8 Finance Agent v2: highest score among tested models
E-GLOBAL-006Opus 4.8 delivers better citation precision and token efficiency on financial document workflows
Anthropic Opus 4.8 launch, May 2026Sonnet 4.6 matches Opus on OfficeQA which includes table reading
E-GLOBAL-005GPT-5.5 hallucination rate 86% makes it unreliable for numerical extraction without verification
E-GLOBAL-003