diff --git a/crates/specialized/rtx-fsi/tests/fsi2_harness/overset_march.rs b/crates/specialized/rtx-fsi/tests/fsi2_harness/overset_march.rs
index 494d522..111a8ec 100644
--- a/crates/specialized/rtx-fsi/tests/fsi2_harness/overset_march.rs
+++ b/crates/specialized/rtx-fsi/tests/fsi2_harness/overset_march.rs
@@ -199,6 +199,7 @@ pub fn run_march_overset(case: BenchmarkCase, config: &OversetMarchConfig) -> Ov
     // of the flag's high-frequency modes (the thickness breathing the
     // body-fitted wall couples to at s = 1); the benchmark's average
     // acceleration (γ = ½) has none.
+    let newmark_beta: f64;
     let analysis = match std::env::var("RTX_FSI2O_NEWMARK_GAMMA")
         .ok()
         .and_then(|v| v.parse::<f64>().ok())
@@ -206,9 +207,13 @@ pub fn run_march_overset(case: BenchmarkCase, config: &OversetMarchConfig) -> Ov
         Some(g) => {
             let b = (g + 0.5).powi(2) / 4.0;
             println!("  Newmark γ = {g}, β = {b:.4}");
+            newmark_beta = b;
             analysis.with_newmark_parameters(g, b)
         }
-        None => analysis,
+        None => {
+            newmark_beta = 0.25;
+            analysis
+        }
     };
     let flag = RefCell::new(analysis.stepper().unwrap());
     let wetted_dofs: Vec<[usize; 2]> = fluid
@@ -269,6 +274,27 @@ pub fn run_march_overset(case: BenchmarkCase, config: &OversetMarchConfig) -> Ov
             fict_per_node * wetted_dofs.len() as f64
         );
     }
+    // The Newmark acceleration a CANDIDATE displacement implies at the wetted
+    // DoFs, a(d) = (d − u_pred) / (β Δt²) with u_pred from the committed
+    // state — the acceleration the structure itself would report at d, so
+    // the compensating load M_f a(d) is consistent with the relaxed candidate
+    // the fluid was evaluated at, and cancels exactly at the fixed point
+    // (a lagged structure-output acceleration does not: 09-16's first try
+    // made the loop SLOWER, 1.8 → 5.2 subit/step at α 0 → 2).
+    let accel_of = |d_candidate: &[f64], committed: &DynamicState| -> Vec<f64> {
+        let inv = 1.0 / (newmark_beta * dt * dt);
+        let mut a = vec![0.0; 2 * wetted_dofs.len()];
+        for (k, dofs) in wetted_dofs.iter().enumerate() {
+            for c in 0..2 {
+                let dof = dofs[c];
+                let u_pred = committed.displacement[dof]
+                    + dt * committed.velocity[dof]
+                    + dt * dt * (0.5 - newmark_beta) * committed.acceleration[dof];
+                a[2 * k + c] = inv * (d_candidate[2 * k + c] - u_pred);
+            }
+        }
+        a
+    };
     // The load with the compensating term for a given previous-subiterate acceleration.
     let with_fict =
         |nodal: &[(NodeId, Vector3<f64>)], accel: &[f64]| -> Vec<(NodeId, Vector3<f64>)> {
@@ -345,11 +371,7 @@ pub fn run_march_overset(case: BenchmarkCase, config: &OversetMarchConfig) -> Ov
         .unwrap_or(0);
     let phase_start = std::time::Instant::now();
 
-    // The previous subiterate's interface acceleration (the fictitious mass's
-    // compensating load): the committed state's at each step's start.
-    let last_accel: RefCell<Vec<f64>> = RefCell::new(extract_accel(&flag_state));
     for step in 0..coupled_steps {
-        last_accel.replace(extract_accel(&flag_state));
         let d_n = extract(&flag_state);
         let v_n: Option<Vec<f64>> = cfg.c1_interface.then(|| extract_velocity(&flag_state));
         let d_predicted = if cfg.predictor == "kinematic" {
@@ -359,7 +381,6 @@ pub fn run_march_overset(case: BenchmarkCase, config: &OversetMarchConfig) -> Ov
             flag.borrow_mut()
                 .set_nodal_forces(&with_fict(&committed_nodal, &extract_accel(&flag_state)));
             let (predicted, _) = flag.borrow_mut().step(&flag_state).unwrap();
-            last_accel.replace(extract_accel(&predicted));
             extract(&predicted)
         };
         let saved = fluid.borrow().snapshot();
@@ -375,9 +396,8 @@ pub fn run_march_overset(case: BenchmarkCase, config: &OversetMarchConfig) -> Ov
             t_fluid.set(t_fluid.get() + fs.elapsed().as_secs_f64());
             let ss = std::time::Instant::now();
             let mut flag_ref = flag.borrow_mut();
-            flag_ref.set_nodal_forces(&with_fict(&nodal, &last_accel.borrow()));
+            flag_ref.set_nodal_forces(&with_fict(&nodal, &accel_of(d_candidate, &flag_state)));
             let (candidate_state, _) = flag_ref.step(&flag_state).unwrap();
-            last_accel.replace(extract_accel(&candidate_state));
             t_structure.set(t_structure.get() + ss.elapsed().as_secs_f64());
             let d_new = extract(&candidate_state);
             if step < cfg.trace_steps {