← All benchmarks

VulcanBench Frontier v4 (functional pass@1)

Code

The share of VulcanBench Frontier v4's 23 behavioural-reconstruction tasks that a configuration fully solves in its single run, in percent. A task counts only when its functional score from the deterministic hidden tests captured from the legacy binary is 1.0, and a regression-test failure gates the functional score to zero. No model judges this metric, so it does not depend on the Code quality panel. The denominator is the tasks scored, which is 23 unless a run has no published result.

Benchmark source: Vulcanbench

Top models (higher is better)