diff --git a/dd-java-agent/agent-llmobs/src/main/java/datadog/trace/llmobs/domain/LLMObsEval.java b/dd-java-agent/agent-llmobs/src/main/java/datadog/trace/llmobs/domain/LLMObsEval.java
index f38f00f4c4c..d641212b539 100644
--- a/dd-java-agent/agent-llmobs/src/main/java/datadog/trace/llmobs/domain/LLMObsEval.java
+++ b/dd-java-agent/agent-llmobs/src/main/java/datadog/trace/llmobs/domain/LLMObsEval.java
@@ -5,6 +5,7 @@
 import com.squareup.moshi.JsonReader;
 import com.squareup.moshi.JsonWriter;
 import com.squareup.moshi.Moshi;
+import datadog.trace.api.DDTraceApiInfo;
 import datadog.trace.llmobs.LLMObsIntakeWorker;
 import java.io.IOException;
 import java.util.List;
@@ -45,7 +46,10 @@ public LLMObsEval(
     this.ml_app = mlApp;
     this.metric_type = metricType;
     this.label = label;
-    this.tags = tags == null ? null : IntakeTags.flatten(tags);
+    // Every submission carries the tracer version and the ML app, as dd-trace-py and dd-trace-js
+    // do, so that evals can be filtered by tag and not only by the top level ml_app field.
+    this.tags =
+        IntakeTags.flatten(tags, "ddtrace.version:" + DDTraceApiInfo.VERSION, "ml_app:" + mlApp);
   }
 
   /**
diff --git a/dd-java-agent/agent-llmobs/src/test/java/datadog/trace/llmobs/domain/LLMObsEvalTest.java b/dd-java-agent/agent-llmobs/src/test/java/datadog/trace/llmobs/domain/LLMObsEvalTest.java
index c293537a3f9..381ef849824 100644
--- a/dd-java-agent/agent-llmobs/src/test/java/datadog/trace/llmobs/domain/LLMObsEvalTest.java
+++ b/dd-java-agent/agent-llmobs/src/test/java/datadog/trace/llmobs/domain/LLMObsEvalTest.java
@@ -6,6 +6,7 @@
 import com.squareup.moshi.JsonAdapter;
 import com.squareup.moshi.Moshi;
 import com.squareup.moshi.Types;
+import datadog.trace.api.DDTraceApiInfo;
 import java.io.IOException;
 import java.util.Arrays;
 import java.util.Collections;
@@ -21,8 +22,9 @@
  * an adapter held by the worker. These tests exist to prove that indirection changed nothing on the
  * wire: the same key set as before, no {@code submitter}, no feedback-only target.
  *
- * <p>The single intentional addition is {@code event_kind:"evaluation"}, which discriminates
- * evaluations from feedback the way dd-trace-py and dd-trace-js do.
+ * <p>The intentional additions are {@code event_kind:"evaluation"}, which discriminates evaluations
+ * from feedback, and the {@code ddtrace.version} and {@code ml_app} tags every submission carries.
+ * Both align the payload with dd-trace-py and dd-trace-js.
  */
 class LLMObsEvalTest {
 
@@ -67,9 +69,12 @@ void testScoreEvalCarriesTheV1KeySetAndNothingElse() throws IOException {
     assertEquals("score", metric.get("metric_type"));
     assertEquals("sentiment", metric.get("label"));
     assertEquals(0.75, metric.get("score_value"));
-    assertEquals(Collections.singletonList("source:web-ui"), metric.get("tags"));
+    assertEquals(
+        Arrays.asList(
+            "ddtrace.version:" + DDTraceApiInfo.VERSION, "ml_app:my-app", "source:web-ui"),
+        metric.get("tags"));
 
-    // The only addition to the v1 payload.
+    // The additions to the v1 payload.
     assertEquals("evaluation", metric.get("event_kind"));
 
     // Feedback-only keys must never leak into the v1 payload.
@@ -92,8 +97,10 @@ void testCategoricalEvalCarriesTheV1KeySet() throws IOException {
     assertEquals("categorical", metric.get("metric_type"));
     assertEquals("positive", metric.get("categorical_value"));
     assertFalse(metric.containsKey("score_value"), metric.toString());
-    // A null tag map is omitted rather than serialized as an empty list.
-    assertFalse(metric.containsKey("tags"), metric.toString());
+    // A null user tag map still yields the tags every submission carries.
+    assertEquals(
+        Arrays.asList("ddtrace.version:" + DDTraceApiInfo.VERSION, "ml_app:my-app"),
+        metric.get("tags"));
     assertEquals("evaluation", metric.get("event_kind"));
   }