diff --git a/crates/solidcore/src/surface/surface/p1.rs b/crates/solidcore/src/surface/surface/p1.rs
index 0c95c953fd..9e09eedc43 100644
--- a/crates/solidcore/src/surface/surface/p1.rs
+++ b/crates/solidcore/src/surface/surface/p1.rs
@@ -1131,6 +1131,7 @@ impl Surface {
         highest_v_deriv: usize,
         flags: EvaluateFlags,
     ) -> Result<Vec<Vec<Vec3>>> {
+        self.check_evaluation_copy_depth(0)?;
         match self.kind() {
             SurfaceKind::BSpline(s) => s.evaluate_simple(
                 uv,
diff --git a/crates/solidcore/src/surface/surface/p2.rs b/crates/solidcore/src/surface/surface/p2.rs
index a9b25b0fd7..1fe4fe5784 100644
--- a/crates/solidcore/src/surface/surface/p2.rs
+++ b/crates/solidcore/src/surface/surface/p2.rs
@@ -4979,6 +4979,25 @@ mod kernel_dispatch_tests {
                 EvaluateFlags::default()
             )
             .is_err());
+        let mut bounded = surface.clone();
+        for _ in 0..70 {
+            bounded = Surface::from(CurveBoundedSurface::new(
+                bounded,
+                Vec::new(),
+                Vec::new(),
+                true,
+            ));
+        }
+        assert!(matches!(
+            bounded.evaluate_simple(
+                &kernel,
+                Point2::new(0.5, 0.5),
+                0,
+                0,
+                EvaluateFlags::default(),
+            ),
+            Err(SolidError::InvalidInput { .. })
+        ));
         for _ in 0..70 {
             surface = Surface::from(StepSurface::new(surface, true));
         }
diff --git a/plan/findings.md b/plan/findings.md
index 31c3fc7122..12b0ac127a 100644
--- a/plan/findings.md
+++ b/plan/findings.md
@@ -5446,3 +5446,8 @@ Apple build after their removal. The final GNU rerun is recorded below when
 complete. No #71/#100 numerical implementation is imported. Common intersection
 adapter errors now identify the remaining Surface integration rather than the
 already-partial cone numerical unit.
+
+The review also found `Surface::evaluate_simple` recursively delegates through
+curve-bounded bases without the preflight used by the other common evaluators.
+It now checks the same 64-level owned-geometry limit before dispatch; a 70-level
+curve-bounded regression verifies normal `InvalidInput` rather than recursion.