1 / 6
ABIGAIL
PatentBench
We Benchmarked Patent AI.
Here's What We Found.
298
Tests
82
Real Office
Actions
8
Technology
Centers
ABIGAIL abigail.app/patentbench
2 / 6
$100M+ in funding.
Zero published benchmarks.
Patent AI Vendor A
$55M raised
None
Patent AI Vendor B
$21M raised
None
Patent AI Vendor C
$35M raised
None
Published accuracy data: zero across the board.
ABIGAIL abigail.app/patentbench
3 / 6
PatentBench Results
Action Classification
100%
Timeline Analysis
100%
Fee Computation
100%
Deadline Calculation
81.1%*
*date arithmetic edge cases in test harness
Overall
92.3%
ABIGAIL abigail.app/patentbench
4 / 6
Technology Center Coverage
TC2800 Electrical (19)
100%
TC3700 Mechanical (14)
100%
TC3600 Business (11)
100%
TC2400 Networking (10)
100%
TC1700 Chemical (10)
100%
TC2600 Comms (9)
100%
TC1600 Biotech (6)
100%
TC2100 Software (3)
100%
Numbers in parentheses = Office Actions tested per Technology Center
ABIGAIL abigail.app/patentbench
5 / 6
The Glass Box Standard
Radical transparency for patent AI evaluation
1
Test Set
Publication
2
Rubric
Transparency
3
Output
Availability
4
Failure Mode
Analysis
5
Continuous
Reporting
"Every number. Every failure. Every limitation."
ABIGAIL abigail.app/patentbench
6 / 6
Open Source. Open Data.
Open Challenge.
Full Results
abigail.app/patentbench
GitHub
github.com/rhahn28/patentbench
We challenge every patent AI vendor
to publish their benchmarks.
ABIGAIL abigail.app/patentbench