DocsReference
Sources
Every figure, dataset and paper these docs rely on, and what each is used for.
Every figure, dataset and design reference these docs rely on, with what it is used for. Each link was checked to resolve.
Industry research
The case for Riffle: review, not authoring, is now the bottleneck.
| Source | Used for |
|---|---|
| Faros AI, The AI Engineering Report 2026: The Acceleration Whiplash | Median review time up 441.5%; bugs per developer up 54%; 31% more PRs merged with no review |
| LinearB, 2026 Software Engineering Benchmarks | 8.1M PRs: AI PRs wait 4.6× longer for review; 32.7% vs 84.4% acceptance |
| GitHub, Agent pull requests are everywhere | More than 1 in 5 code reviews now involve an agent |
Research: defect prediction
The literature the model, labels and planning targets rest on.
| Source | Used for |
|---|---|
| Kamei et al., A large-scale empirical study of just-in-time quality assurance, TSE 2013 | Within-project ROC-AUC and effort-aware recall targets |
| Kamei et al., Studying just-in-time defect prediction using cross-project models, EMSE 2016 | The 0.65+ ROC-AUC target on repositories the model never trained on |
| Yan et al., Characterizing and identifying reverted commits, EMSE 2019 | Lift, recall at the top 20% of effort, and the revert label |
| Shimagaki et al., Why are commits being reverted?, ICSME 2016 | Revert rate in open source: 1–5% of commits |
| Zeng et al., Deep just-in-time defect prediction: how far are we?, ISSTA 2021 | Why full diffs are not kept: diff-reading deep models did not reliably beat metric-based ones |
Datasets
| Source | Used for |
|---|---|
| Keshavarz and Nagappan, ApacheJIT, MSR 2022 | 106,674 commits, 28,239 labelled bug-inducing; the ~26% SZZ rate |
| Zhang, Rastogi and Yu, On the Shoulders of Giants, MSR 2020 | Engineered features for pull request outcome prediction |
| Gousios and Zaidman, A dataset for pull request research, MSR 2014 | Pull request dataset methodology |
| AIDev | Agent-authored pull requests with merge outcomes, CI results and reviewer interactions |
Platform documentation
Constraints the data pipeline is designed around.
| Source | Used for |
|---|---|
| GitHub Docs, Rate limits for the REST API | 5,000 requests per hour, secondary limits, and the extraction timeline |
| GitHub Changelog, Actions retention will cover checks, workflow runs and statuses | From 1 October 2026, check runs and statuses on public repositories are kept for at most 90 days |
| Grunert, My exciting journey into Kubernetes' history, Kubernetes Blog 2020 | A size anchor for structured pull request history |