Case studies
Documented runs, misses included: what the agents wrote, what my review caught, what it let through, and the number each run moved.
3
shell-injection paths from request input caught in review. The study lists the 5 security bugs it missed.
30 Agent PRs, One Reviewer: What Review Caught, and the 5 Security Bugs It Missed
Mar 2026 · Own project · under 3 days · Claude Code, Next.js, git worktrees · re-audited 12 Oct 2026
March 2026: agents opened 31 PRs on my own dashboard and I merged 30 after review, with no tests or CI. Review caught 3 shell injections and missed 5 bugs.
Read the case study →0
the true count behind a landing-page line labelled "real number". The gate keeps lines like it off the page.
The Honesty Gate: Stopping Agents From Publishing Invented Facts
2026 · Client work (NDA) · Role: reviewer, author of the gates
An agent-built landing page labelled a launch count "real number". The true count was 0. The fix makes every rendered string prove where it came from.
Read the case study →57%
lower cost per merged PR in one month, then it rose 42% by early October
Cost per Merged PR: What 3 Months of My Claude Code Ledger Show
Jul–Oct 2026 · Own setup · Claude Code
My Claude Code cost per merged PR fell from $73 to $32 in a month, then rose 42% by early October. Three months of ledger data, and my counting mistake.
Read the case study →