Electricity Bench

Methodology

The method is fully public; only the questions are private. That is what makes “tests private” credible rather than suspicious.

How grading works

Each suite is a set of graded tasks with a fixed scoring method - programmatic checks or a pinned LLM judge - run several times per model for stability.

Capability rolls up to a letter grade per suite via published grade bands, alongside cost, latency, and reliability. Grades are comparable within a suite version only.

Suites

structured-extraction@v1

programmatic 3 repeats

structured-extraction@v2

hybrid 3 repeats