Observed arrival · 2026-11-10
DocAI Bench: Valid JSON Can Still Be Wrong
A benchmark service that checks AI document-extraction results against versioned reference datasets.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Teams testing financial-document extraction models
- Worth noticing
- The displayed DEV run covers 86 numeric fields across eight English filings and explicitly excludes scale-handling tests.
Field notes
The sample report breaks results down by field: cash is exact in 8 of 8 filings, while short-term liabilities are exact in 7 of 8. Its references are XBRL-derived rather than checked cell by cell by a human. The report also states that its prompt converts printed scale, so the measured run does not exercise scale handling—a useful boundary when interpreting the score.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
◎NicheUnusually specific use
One card from the complete issue