Case Studies

The useful question is not whether coding agents ever make mistakes. It is which mistakes escape, which gates catch them, and whether the delivery system improves after each miss.

Published studies

We audited 30 commits from an AI-heavy workflow

An adversarial audit of Shakapacker found real defects, weak tests, incomplete provenance, and a larger problem hiding in plain sight: generated process noise made the engineering signal harder to see.

The study explains the remediation and the two-control model we now use: improve what gets generated, then fail closed when integration evidence is incomplete.