Auto 𝗣𝘂𝗹𝗹 𝗥𝗲𝗾𝘂𝗲𝘀𝘁 𝗥𝗲𝘃𝗶𝗲𝘄: 𝗔𝗻𝗱𝗿𝗼𝗶𝗱 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗚𝘂𝗮𝗿𝗱𝗿𝗮𝗶𝗹𝘀 Incoming Most "𝗔𝗜 𝗲𝗻𝗳𝗼𝗿𝗰𝗲𝘀 𝘆𝗼𝘂𝗿 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲" claims fall apart under one question: what happens on a genuine judgment call? So I split this into two tiers, with two different authorities. Clean Architecture layer boundaries domain never depends on data or presentation are checked with Konsist and actually fail the build. Not "𝘀𝗵𝗼𝘂𝗹𝗱 𝗳𝗮𝗶𝗹." I pushed a real violation to prove it: a PR where a presentation class reaches into the data layer directly, and CI rejects it with a precise error naming the file and the illegal import. No LLM in that loop at all It's a deterministic JUnit assertion, same category as a compiler error. SOLID principle smells (SRP, OCP, LSP, ISP, DIP) go through a separate LLM pass that comments on a PR with the specific principle and reasoning but the workflow has comment-only permissions, and the reviewer code itself never has a path to block a merge. I tested that too: with no API key configured, it logs a notice and exits clean rather than red-X'ing your build over a missing secret Being honest about which part is strict and which part is advisory isn't a caveat on this project — it's the actual design decision, and it's the one most "AI code review" tooling quietly avoids making.

Repo: https://github.com/ranab4b/android-architecture-guardrails