diff --git a/CLAUDE.md b/CLAUDE.md
index 51d621f..1f23f8c 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -207,6 +207,15 @@ When an agent who normally evaluates (currently Leo) is also the proposer, they
 
 As the collective grows, scale to up to 3 peer reviewers selected by highest domain linkage. Currently: at least 1 of Rio or Clay.
 
+### Synthesis claims require multi-agent review
+
+Cross-domain synthesis claims (Leo's core function) must be reviewed by **at least 2 domain agents** — the agents whose domain expertise is most relevant or whose knowledge base is most affected by the claim. Selection criteria:
+1. **Domain coverage** — every domain touched by the synthesis must have at least one reviewer from that domain
+2. **Impact** — if the synthesis changes how a domain's claims should be interpreted, that domain's agent reviews
+3. **More than 2 is fine** — if the synthesis spans 3+ domains, involve all affected agents
+
+Synthesis review focuses on: is the cross-domain mechanism real (not just analogy)? Would domain experts in both fields recognize the connection? Does the synthesis add insight neither domain could produce alone?
+
 ### Review checklist
 For each proposed claim, check:
 
diff --git a/skills/synthesize.md b/skills/synthesize.md
index e32de37..87fa94f 100644
--- a/skills/synthesize.md
+++ b/skills/synthesize.md
@@ -40,9 +40,10 @@ If the connection passes the test, create a new claim:
 ### Step 4: Route for evaluation
 
 Synthesis claims get special evaluation routing:
-- Leo evaluates (always — this is Leo's core function)
-- Both contributing domain agents evaluate
-- The evaluation focuses on: is the mechanism real? Would domain experts agree?
+- Leo proposes (always — this is Leo's core function) and cannot self-merge
+- **At least 2 domain agents** must review — selected by domain expertise relevance or knowledge base impact (see CLAUDE.md "Synthesis claims require multi-agent review")
+- If the synthesis spans 3+ domains, involve all affected agents
+- The evaluation focuses on: is the mechanism real? Would domain experts agree? Does the synthesis change how their domain's claims should be interpreted?
 
 ### Step 5: Update transition landscape
 
@@ -61,5 +62,5 @@ If the synthesis changes Leo's slope reading for any domain:
 ## Quality Gate
 
 - Every synthesis articulates a specific mechanism (not just "these are related")
-- Both contributing domain agents validate the connection
+- At least 2 contributing domain agents validate the connection
 - The synthesis adds value neither domain could produce alone