perit.ai
Perit Bench

Domain benchmarks built on expert consensus.

Perit Bench is a suite of evaluation benchmarks where every gold-standard answer is adjudicated by verified domain experts — not crowd-sourced labels or synthetic references.

4,080+
Total benchmark tasks
100%
Expert-verified gold standard
12
Domains covered
0.94 κ
Median inter-rater reliability
Benchmark suites · Q3 2026

Four domains. One standard of truth.

Medical1,240 tasks

Perit-MedBench

Clinical reasoning, differential diagnosis, and treatment-plan evaluation scored against board-certified physician consensus.

Legal980 tasks

Perit-LegalBench

Case-law analysis, contract interpretation, and regulatory compliance grading benchmarked against bar-admitted attorney review.

Finance760 tasks

Perit-FinBench

Equity valuation, credit analysis, and M&A reasoning evaluated against sell-side and buy-side analyst standards.

Science1,100 tasks

Perit-SciBench

Graduate-level physics derivations, genomic variant interpretation, and materials-science reasoning graded by domain PhDs.

How it works

Gold standards you can defend.

Each Perit Bench task is authored by two independent domain experts, adjudicated by a senior verifier, and held out from all training data used by Perit's own annotation workforce. When you evaluate your model against Perit Bench, you're measuring against the same standard of practice your end users expect.

Benchmarks are updated quarterly with new tasks drawn from real client engagements (anonymized and consent-scoped). Scores include per-domain breakdowns, disagreement analysis, and confidence intervals based on expert adjudication variance.