diff --git a/engine/enginefuzz_test.go b/engine/enginefuzz_test.go
index a1c60d563..3557612da 100644
--- a/engine/enginefuzz_test.go
+++ b/engine/enginefuzz_test.go
@@ -56,6 +56,15 @@ func shouldValidateSamples(expr parser.Expr) bool {
 
 	parser.Inspect(expr, func(node parser.Node, path []parser.Node) error {
 		switch n := node.(type) {
+		case *parser.VectorSelector:
+			if n.Timestamp != nil || n.StartOrEnd != 0 {
+				// The Thanos engine's step invariant operator caches the result of the
+				// first evaluation and replays it for subsequent steps without re-counting
+				// samples. This leads to fewer total samples compared to Prometheus which
+				// counts samples at every step.
+				valid = false
+				return errors.New("error")
+			}
 		case *parser.Call:
 			switch n.Func.Name {
 			case "scalar":
@@ -81,6 +90,14 @@ func validateExpr(expr parser.Expr, testType testType) bool {
 
 	parser.Inspect(expr, func(node parser.Node, path []parser.Node) error {
 		switch n := node.(type) {
+		case *parser.AggregateExpr:
+			if n.Op == parser.COUNT_VALUES {
+				// count_values converts float values to string labels. Tiny floating point
+				// precision differences between engines (e.g. 61.24999999999997 vs 61.24999999999998)
+				// produce different label values, causing result mismatches.
+				valid = false
+				return errors.New("error")
+			}
 		case *parser.Call:
 			switch n.Func.Name {
 			case "sort", "sort_desc", "sort_by_label", "sort_by_label_desc":
diff --git a/engine/testdata/fuzz/FuzzNativeHistogramQuery/df60c8694f1c9759 b/engine/testdata/fuzz/FuzzNativeHistogramQuery/df60c8694f1c9759
new file mode 100644
index 000000000..e3baa6eb3
--- /dev/null
+++ b/engine/testdata/fuzz/FuzzNativeHistogramQuery/df60c8694f1c9759
@@ -0,0 +1,10 @@
+go test fuzz v1
+int64(-57)
+uint32(27)
+uint32(248)
+uint32(12)
+int8(0)
+int8(0)
+uint64(1)
+uint64(1)
+uint64(16)