Recoleta
GitHub
Source note

Automating Low-Risk Code Review at Meta: RADAR, Risk Calibration, and Review Efficiency

Authors
Chris Adams, Arjun Singh Banga, Parveen Bansal, Souvik Bhattacharya, Rujin Cao, Pedro Canahuati, et al.
Source
arXiv
Published
Collection
Software Intelligence
Code Review AutomationSoftware Foundation ModelsCode IntelligenceAI Generated CodeDeveloper Productivity
Open arXivMarkdown source

Summary

RADAR automates low-risk code review at Meta by combining source eligibility rules, a Diff Risk Score model, LLM-based code review, and deterministic checks before landing changes.

Problem

  • Meta’s code volume is rising faster than human review capacity: significant lines of code per human-landed diff grew 105.9% year over year, and diffs per developer per month grew 51%.
  • Agentic AI accounts for more than 80% of that growth, while the share of diffs reviewed within 24 hours has declined.
  • The paper targets routine low-risk diffs so human reviewers can spend time on changes with higher production or design risk.

Approach

  • RADAR classifies each diff by authorship and source: human-authored diffs, deterministic codemods, AI-generated codemods, and RACER runbook diffs.
  • Bot diffs pass through ACE, which applies static safety gates, configurable Diff Risk Score thresholds, LLM-based Automated Code Review, and final validation before landing.
  • Human-authored diffs pass through RADAR Verification and RADAR Approval, using author eligibility, scope exclusions, content blocklists, Diff Risk Score, and LLM review.
  • Diff Risk Score thresholds control risk appetite: examples include P5 for default human-authored diffs, P20 for stricter bot sources, and P50 for allowlisted RACER runbooks.
  • RACER runbooks also use 60-day risk history checks, daily landing caps, and denylists for sources with incidents or sensitive targets.

Results

  • RADAR reviewed more than 535K diffs and landed more than 331K diffs in production at Meta.
  • The system reached more than 25K diffs per day.
  • Relaxing the Diff Risk Score threshold from the 25th percentile to the 50th percentile raised the approve rate to 60.31%.
  • RADAR-reviewed diffs had a revert rate one-third that of non-RADAR diffs.
  • RADAR-reviewed diffs had a Production Incident rate one-fiftieth that of non-RADAR diffs.
  • Compared with human-reviewed diffs, the paper states that RADAR reduces median time to close by over 330% and median diff review wall time by 35%.

Link