{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/bebda7f7cd114626a5593455e924ee86\" frameborder=\"0\" width=\"1104\" height=\"828\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":828,"width":1104,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":828,"thumbnail_width":1104,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/bebda7f7cd114626a5593455e924ee86-810b9473ff04c56f.gif","duration":209.256,"title":"Canary Brain Prevents Costly Production Regressions","description":"This Loom explains Canary Brain, an open source system to reduce risky code changes in high-stakes Ads infrastructure. Dave notes deployments happen continuously, with changes potentially costing thousands to millions if they break production, and that simple diffs can look fine through unit tests or quick reviews but still cause incidents. Canary Brain uses layers including Diff Risk Score from past incidents and an offline RPC Replay “Service Lab” that runs the change against the previous 14 days of production data to predict impacts on pacing. He also mentions Bicep to find quality-failure root causes, canary rollout to low value regions first, and that Difference Score reportedly catches 84.6% of CEVs at 50% gating (run locally with Claude 2.5)."}