diff --git a/services/discord-gateway/src/modules/ai-moderation/aiAnalyzer.ts b/services/discord-gateway/src/modules/ai-moderation/aiAnalyzer.ts
index b8bfb48..e90fc9d 100644
--- a/services/discord-gateway/src/modules/ai-moderation/aiAnalyzer.ts
+++ b/services/discord-gateway/src/modules/ai-moderation/aiAnalyzer.ts
@@ -158,9 +158,9 @@ let lastError: string | null = null;
 let moderationClient: Client | undefined;
 
 // Batch circuit breaker
-let consecutiveErrors = 0;
+const conversationConsecutiveErrors = new Map<string, number>();
 const MAX_CONSECUTIVE_ERRORS = 5;
-let globalCooldownUntil = 0;
+const CONVERSATION_CB_COOLDOWN_MS = 60000;
 
 // ---------------------------------------------------------------------------
 // Individual fallback queue — runs PARALLEL to the batch pipeline.
diff --git a/services/discord-gateway/src/shared/config/config.ts b/services/discord-gateway/src/shared/config/config.ts
index b6a639e..76274e6 100644
--- a/services/discord-gateway/src/shared/config/config.ts
+++ b/services/discord-gateway/src/shared/config/config.ts
@@ -116,12 +116,12 @@ const configSchema = z
       .number()
       .positive()
       .default(120000),
-    /** Max concurrent individual-fallback LLM calls (effectively unlimited). */
+    /** Max concurrent individual-fallback jobs admitted by the main event loop. */
     AI_ANALYSIS_INDIVIDUAL_MAX_CONCURRENT: z.coerce
       .number()
       .int()
       .positive()
-      .default(1000),
+      .default(50),
     /**
      * How many consecutive individual-fallback errors trigger the individual
      * circuit breaker (separate from the batch circuit breaker).