diff --git a/build.gradle.kts b/build.gradle.kts
index 0e0df9dbc6..2ea2561289 100644
--- a/build.gradle.kts
+++ b/build.gradle.kts
@@ -156,6 +156,17 @@ allprojects {
         if (isScreenshotTest) {
             // Increase heap size for screenshot tests
             maxHeapSize = "2g"
+            // Give every screenshot test class its own JVM.
+            //
+            // There are 5 screenshot classes (PreviewShard1..4Test + PreviewA11yTest) and each
+            // shard renders ~770 previews in a single process. Gradle reuses a forked JVM for the
+            // next class once maxParallelForks is exhausted, so on a 6-core runner (forks = 3) two
+            // classes run second, in a process that has already rendered a full shard — and they
+            // render differently from how the goldens were recorded. That is exactly the observed
+            // failure: shards 2 and 4 fail 100% while 1 and 3 pass 100%, deterministically, for
+            // byte-identical states. The 8-core runner (forks = 4) reused only one JVM and was green
+            // on the same tree. Restarting per class costs 5 JVM startups and no extra peak memory.
+            forkEvery = 1L
             // Record all the languages?
             if (project.hasProperty("allLanguagesNoEnglish")) {
                 // Do not record English language