93% of tests still pass after a rewrite in another language
- Best published result
- 48%
- Same AI model, no Bosun
- 74%
- Bosun
- 93%
RepoMod-Bench: 21 open-source codebases, each rewritten in another language and graded by a hidden test suite. We ran all 21 three times, unattended.
8x
faster legacy migration or more
4x faster than standard agentic coding tools.
10+
languages: 4GL family, C, C++, C#/.NET, COBOL, Go, Java, JavaScript, TypeScript, Perl, PHP, Python, Rust
If it builds on Linux, we can port it.
93%
of tests still pass after rewriting 21 codebases in another language. Best published result: 48%.
Half of the differences our proofs caught would have slipped past the tests.
Measured on RepoMod-Bench, a public benchmark of 21 codebases.
The twelve large cases, by size
| Repository | Lines | Bosun | Same model, no Bosun |
|---|---|---|---|
| ledger C++ → Go | 50K | 97% | 62% |
| wabt C++ → Rust | 55K | 100% | 89% |
| taskwarrior C++ → Rust | 55K | 100% | 69% |
| lightningcss Rust → Go | 62K | 93% | 77% |
| bc C → Rust | 117K | 100% | 100% |
| hugo Go → Rust | 122K | 76% | 70% |
| jq-gojq C → Go | 147K | 100% | 100% |
| pdfcpu Go → Rust | 160K | 82% | 6% |
| uncrustify C++ → Rust | 162K | 48% | 0% |
| prettier JavaScript → Rust | 175K | 55% | 21% |
| verible C++ → Rust | 191K | 100% | 0% |
| qalculate C++ → Go | 211K | 100% | 64% |
Method, cost and the data for all 63 runs: read the full write-up