Blog
Notes on code review and security
Benchmarks, practices, and what we learned building an AI reviewer.
Precision Beats Recall: Why the Cheaper Model Wrote the Better Code Review
GPT-6.1 Terra costs about 8x more than GPT-6 Luna and found the same bugs. It also raised nearly twice as many issues, and once an auto-fix agent reads the review, every extra comment turns into code.
6 min read
4 min read
Patchlight vs Cursor: Whole-Repo Security Scans, Benchmarked
Can the coding agent you already pay for replace a security scanner? We ran Cursor with Grok 4.5 against Patchlight on our internal security benchmark. Patchlight scored an F3 of 74.1 to Cursor's 51.5, at under 5 cents a scan.
8 min read
Security in the Age of AI-Written Code
AI assistants made writing code cheap. They did not make reviewing it cheap. What changes about application security when most of your diff was generated, and what a team should actually do about it.
9 min read
Shipping Fewer Bugs Without Shipping Slower
For teams deploying several times a day, most escaped defects don't come from moving fast. They come from checks that quietly became optional. Five practices that hold up under deadline pressure.
9 min read
Benchmarking Patchlight: Our Scores on Code Review Bench
We ran Patchlight through Martian's open-source Code Review Bench: 50 real PRs, 137 human-verified findings, one judge for every tool. Here are the numbers, the methodology, and where we're strong.