diff --git a/docs/investigation-685-reviewer-approved-destructive-compose.md b/docs/investigation-685-reviewer-approved-destructive-compose.md
index 033688b..d8f9d2f 100644
--- a/docs/investigation-685-reviewer-approved-destructive-compose.md
+++ b/docs/investigation-685-reviewer-approved-destructive-compose.md
@@ -8,7 +8,7 @@
 
 The reviewer agent approved PR #683 in ~1 minute without flagging that it
 contained a destructive rewrite of `docker-compose.yml` — dropping named
-volumes, bind mounts, env vars, restart policy, and security options. Five
+volumes, bind mounts, env vars, restart policy, and security options. Six
 structural gaps in the review pipeline allowed this to pass.
 
 ## Root causes
diff --git a/formulas/review-pr.toml b/formulas/review-pr.toml
index 67b6358..fe62a89 100644
--- a/formulas/review-pr.toml
+++ b/formulas/review-pr.toml
@@ -277,9 +277,11 @@ for actual problems (bugs, security issues, broken functionality, missing
 required behavior). Use DISCUSS sparingly.
 
 Note: The bias toward APPROVE applies to code correctness and style decisions.
-It does NOT apply to documentation consistency (step 3b) or tech-debt filing
-(step 7) — those are separate concerns that should be handled regardless of
-the change's correctness.
+It does NOT apply to documentation consistency (step 3b), infrastructure file
+findings (step 3c), or tech-debt filing (step 7) — those are separate concerns
+that should be handled regardless of the change's correctness. In particular,
+dropped production configuration (volumes, bind mounts, env vars, restart
+policy) is a blocking defect, not a style preference.
 
 ## 9. Output