Grounded in research
The thresholds come from three studies that all reach the same conclusion: review quality drops sharply past ~400 LOC, and becomes essentially impossible past ~800.
- Cohen 2006 (SmartBear): review quality drops above ~200 LOC, sharply above 400 LOC.
- Bosu et al. 2017 (Microsoft): PRs > 500 LOC have 2.3× higher post-merge defect-leak rate.
- GitHub 2023 data: median PR is 30 LOC; outliers > 1000 LOC get less time per line due to fatigue.
How the estimate works
Base estimate is 0.6 minutes per line of churn (added + deleted), floored at 5 minutes. Every PR takes at least that long to context-load.
Adjustments:
- Files changed: 1 minute per file (context-switching overhead).
- Cross-module: +20% (reviewer mentally switches between two domains).
- Tests included: –15% (the test acts as a spec; less re-reading needed to understand intent).
Total churn drives the tier: ≤ 100 lines is the sweet spot (Cohen 2006), ≤ 400 is still effective, ≤ 800 is risky, > 800 should be split before review.
Want this enforced in your workflow?
Stride's Verify module surfaces PR size warnings before merge, blocks PRs > 800 LOC by policy, and tracks per-team review cycle time as a DORA metric.