diff --git a/srcmorph/src/main/java/net/ladenthin/srcmorph/provider/LlamaCppJniAiGenerationProvider.java b/srcmorph/src/main/java/net/ladenthin/srcmorph/provider/LlamaCppJniAiGenerationProvider.java
index 541c6054..4b8479e4 100644
--- a/srcmorph/src/main/java/net/ladenthin/srcmorph/provider/LlamaCppJniAiGenerationProvider.java
+++ b/srcmorph/src/main/java/net/ladenthin/srcmorph/provider/LlamaCppJniAiGenerationProvider.java
@@ -175,7 +175,14 @@ private LlamaModel model() {
             // llama.version alone will NOT surface this: the guard keeps throwing and the knob keeps
             // looking broken, so the bump checklist has to name it. See TODO.md.
             if (config.flashAttn()) {
-                throw new IllegalArgumentException(FLASH_ATTN_UNSUPPORTED_MESSAGE);
+                // The model path is not decoration: this guard fires on the direct-API path, where
+                // there is no aiDefinition key to name (the plan-time guard in EngineSupport prefixes
+                // that one), so without it a caller running several models is told what is wrong but
+                // not which configuration to change. It also keeps the message out of
+                // WEM_WEAK_EXCEPTION_MESSAGING, which SpotBugs raises on a throw whose whole message
+                // is a compile-time constant.
+                throw new IllegalArgumentException(
+                        "model '" + config.modelPath() + "': " + FLASH_ATTN_UNSUPPORTED_MESSAGE);
             }
             // KV-cache quantization. Set independently of each other, but note that a quantized V cache
             // generally needs Flash Attention above -- llama.cpp refuses the combination otherwise.
diff --git a/srcmorph/src/test/java/net/ladenthin/srcmorph/provider/LlamaCppJniKnobSweepTest.java b/srcmorph/src/test/java/net/ladenthin/srcmorph/provider/LlamaCppJniKnobSweepTest.java
index d1787ebb..96f34c8e 100644
--- a/srcmorph/src/test/java/net/ladenthin/srcmorph/provider/LlamaCppJniKnobSweepTest.java
+++ b/srcmorph/src/test/java/net/ladenthin/srcmorph/provider/LlamaCppJniKnobSweepTest.java
@@ -3,6 +3,7 @@
 // SPDX-License-Identifier: Apache-2.0
 package net.ladenthin.srcmorph.provider;
 
+import static org.hamcrest.CoreMatchers.containsString;
 import static org.hamcrest.CoreMatchers.is;
 import static org.hamcrest.CoreMatchers.notNullValue;
 import static org.hamcrest.MatcherAssert.assertThat;
@@ -217,7 +218,12 @@ public void flashAttn_isRefusedRatherThanSilentlyDropped() {
         try (LlamaCppJniAiGenerationProvider provider = new LlamaCppJniAiGenerationProvider(config, promptSupport)) {
             final IllegalArgumentException thrown =
                     Assertions.assertThrows(IllegalArgumentException.class, () -> provider.generate(request()));
-            assertThat(thrown.getMessage(), is(LlamaCppJniAiGenerationProvider.FLASH_ATTN_UNSUPPORTED_MESSAGE));
+            // containsString, not equality: the provider prefixes the model it is refusing for, so a
+            // caller running several models can tell which configuration to change.
+            assertThat(
+                    thrown.getMessage(),
+                    containsString(LlamaCppJniAiGenerationProvider.FLASH_ATTN_UNSUPPORTED_MESSAGE));
+            assertThat(thrown.getMessage(), containsString(NativeLlamaAvailability.modelPath()));
         }
     }