Generated-code benchmarks

Benchmarks

The uv-benchmarks harness measures how AI models perform when writing Ultraviolet on fixed tasks: whether the result builds, passes tests and scanner checks, stays organized, and what it costs in tokens, dollars, and turns.

Benchmark Explorer

Fallback Methodology
By model generation / csv_invoice_transform

latest summary

gpt-5.5 / uv
Score
21%
Tokens
978k
Cost
$5.22
Build: blockedScanner: cleanTests: blockedLayout: valid
Tasks
Task gpt-5.5 / uv

Methodology and reproducibility

Each run gives a model a fixed task specification and records the outcome of the generated Ultraviolet project. Scores weight project organization, build and functional checks, code quality, performance, and bug or security findings; token, cost, and turn counts come from the model provider's usage reporting for the run.

These results are snapshots, not independently reproducible claims yet. The harness repository is being prepared for public release; until it is published, treat the numbers above as indicative. The status pill on the explorer shows whether you are viewing live harness output or the bundled snapshot.

If you want to evaluate generated Ultraviolet yourself today, the AI-written code review model describes what reviewers should check, and the handbook defines the exact rules the output must satisfy.