diff --git a/backend/src/main/java/org/raddatz/familienarchiv/service/OcrTrainingService.java b/backend/src/main/java/org/raddatz/familienarchiv/service/OcrTrainingService.java
index 4b688119..e241aaa3 100644
--- a/backend/src/main/java/org/raddatz/familienarchiv/service/OcrTrainingService.java
+++ b/backend/src/main/java/org/raddatz/familienarchiv/service/OcrTrainingService.java
@@ -42,6 +42,10 @@ public class OcrTrainingService {
             List<OcrTrainingRun> runs
     ) {}
 
+    // Not safe for horizontal scaling: training reloads the Kraken model in-process on the
+    // Python OCR service after each run. The DB-level RUNNING constraint (V30 partial unique
+    // index) prevents concurrent training API calls, but cannot prevent two OCR service replicas
+    // from diverging on model state. Deploy as a single instance only. See ADR-001.
     @Transactional
     public OcrTrainingRun triggerTraining(UUID triggeredBy) {
         if (trainingRunRepository.findFirstByStatus(TrainingStatus.RUNNING).isPresent()) {
diff --git a/docker-compose.yml b/docker-compose.yml
index 5e06094b..bf57501a 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -72,6 +72,9 @@ services:
       - archive-net
 
   # --- OCR: Python microservice (Surya + Kraken) ---
+  # Single-node only: OCR training reloads the model in-process after each run.
+  # Running multiple replicas would cause training conflicts and model-state divergence.
+  # See ADR-001 for the architectural rationale.
   ocr-service:
     build:
       context: ./ocr-service