Research Paper 20 pages · PDF

Predicting Defects
Before They Happen

A validation study of Git-history risk signals across four open-source projects in three languages: MongoDB (C++), uv (Rust), Qt Base (C++), and Apache Flink (Java). Churn patterns from commit history scored at a fixed cutoff date — then checked months later against actual bug-fix activity. The signal held up in every codebase.

Qt Base · C++
72%
of bug-affected files in the top quartile of a churn-based risk ranking (26% of codebase) — AUC 0.817
uv · Rust
65.8%
of bug-affected files in the top quartile of a churn-based risk ranking (26% of codebase) — AUC 0.789
MongoDB Server · C++
60.9%
of bug-affected files in the top quartile of a churn-based risk ranking (26% of codebase) — AUC 0.751
Apache Flink · Java
50%
of bug-affected files in the top quartile of a churn-based risk ranking (25% of codebase) — AUC 0.682

Contents

  • 1The Problem — Why code review and coverage metrics miss defect concentration
  • 2Methodology — The time-lag design, how defects were labelled, and how accuracy was scored
  • 3The Ownership Paradox — Concentration vs. Silo Risk as orthogonal dimensions
  • 4Results Across Four Codebases — MongoDB, uv, Qt Base, Apache Flink
  • 5Implications for Engineering Leaders — When to trust each signal
  • AAppendices — Dataset details, label extraction, predictor computation, AUC interpretation, limitations

Download the full whitepaper

We'll send the PDF link directly — no newsletter, no follow-up unless you want one.

No spam. One email max.
✓

Thanks — here's the whitepaper.

You can also bookmark this link directly for future reference.

Download PDF (20 pages)

Calyntro · Predicting Defects Before They Happen · 2026