diff --git a/DECISIONS.md b/DECISIONS.md
new file mode 100644
index 0000000..486af8d
--- /dev/null
+++ b/DECISIONS.md
@@ -0,0 +1,21 @@
+# Decision Log
+
+_Auto-generated by the Custodian State Hub._
+
+## Restore the previously-live seven-action tenant-engine image rather than rebuild
+
+**Date:** 2026-08-16  
+**Decided by:** grok  
+
+The four-action regression is a pin problem, not a policy-authoring problem. Image sha256:9320df39 was CI-built from e9911eb, previously live 2026-08-11, and confirmed by tenant-engine on 2026-08-13. The tenant-engine policy has not changed since that commit. A new image would re-bake unrelated later packages and would not be the artifact already verified. TEN-WP-0006 guardrail actions stay out of this restore.
+
+---
+
+## Use the guardrail read/write split now: flex-auth may read, may not set
+
+**Date:** 2026-08-16  
+**Decided by:** grok  
+
+TEN-WP-0006 split the actions so a PDP can read ceilings without changing them, and their tests call the read as actor=flex-auth. Leaving both actions on the single tenant-engine subject would keep the seam unused and leave the intended reader failing unknown_subject. This is a real difference: flex-auth set is denied action_not_granted. An ops write subject is out of scope; writes stay on the existing tenant-engine identity.
+
+---