Benchmarks
Benchmarks built
from real work
Every DataOrigin benchmark starts from work someone actually did: a repository carried to its next release, a close that balanced, a ticket that got resolved. Agents are graded with hidden verifiers on the finished result.
OriginBench 1.0
Measures whether an agent can make a real repository behave exactly like a reference: a deployed binary, a newer release, or another language's implementation. 30 tasks, 12 languages, 630 graded runs.
AP-Ledger Close
An accounting close run end to end inside a replicated back office: reconciliations, accruals, exceptions and the sign-off, graded against how the close actually balanced.
Pre-audit runs are on the environments console; the leaderboard publishes once the audit completes.
Repo-CI Operations
Repository and CI incidents inside a working forge replica: broken pipelines, flaky releases and dependency breakage, resolved the way the original engineers resolved them.
Pre-audit runs are on the environments console; the leaderboard publishes once the audit completes.
Clinic EHR
Clinical documentation and order workflows inside a replicated EHR seeded with licensed records, graded by practising physicians against specialty rubrics.
Pre-audit runs are on the environments console; the leaderboard publishes once the audit completes.
Want your model on a leaderboard?
We run every model with its vendor's own CLI agent under identical conditions, and ship the sample so you can recompute every number. Evaluation access is open to frontier labs.