diff --git a/src/projected_grpo/train.py b/src/projected_grpo/train.py
index 155a210..f143322 100644
--- a/src/projected_grpo/train.py
+++ b/src/projected_grpo/train.py
@@ -688,6 +688,7 @@ class StepLogger:
             cols += [
                 _Col("tau",   6, "tau",   "+.2f", "per-step calibrated route threshold (midpoint of hack vs clean cos clouds)"),
                 _Col("hkgap", 6, "hkgap", "+.2f", "ema_hack_cos - ema_clean_cos; >0 = v_grad still separates hack from clean (else direction dead)"),
+                _Col("resid", 6, "resid", "+.2f", "cos(deployed delta_S.grad AFTER routing, v_grad); ~0 = hack stripped cleanly, >0 = leak into deployed knob"),
             ]
         if arm in ("routing", "routing2"):
             cols += [
@@ -1147,6 +1148,7 @@ def main(cfg: Config) -> int:
         step_flagged: list[float] = []
         step_tau: list[float] = []      # per-(prompt,module) calibrated route threshold
         step_hkgap: list[float] = []    # ema_hack_cos - ema_clean_cos (discrimination gauge)
+        step_resid: list[float] = []    # cos(delta_S.grad AFTER routing, v_grad): hack-ward leak into deployed knob
 
         def _route2_grad_filter(info, n_rollouts: int,
                                 hack_anchor: torch.Tensor,
@@ -1194,7 +1196,14 @@ def main(cfg: Config) -> int:
             # delta_S keeps only the unflagged. Capacity-balanced: both shape [r].
             step_grad_hack[name] = (step_grad_hack[name] + sub.detach().clone()
                                     if name in step_grad_hack else sub.detach().clone())
-            return g - sub
+            g_keep = g - sub                                  # the deployed knob's gradient
+            # Residual hack-ward alignment of the KEPT grad. Disambiguates qE:
+            # qE high + resid~0 = routing stripped the hack cleanly (dominant
+            # teacher grad correctly quarantined); qE high + resid>0 = false
+            # negatives leaked hack-ward grad into the deployed knob (the real
+            # failure). vg is unit, so this is a plain cosine.
+            step_resid.append((g_keep @ vg / g_keep.norm().clamp_min(1e-12)).item())
+            return g_keep
 
         # Split backward into student/teacher only every cos_pre_split_every steps.
         # On split steps: 2 backwards per prompt, populates step_grad_s/_t.
@@ -1842,6 +1851,7 @@ def main(cfg: Config) -> int:
             "q_egy": q_egy,
             "tau": (sum(step_tau) / len(step_tau)) if step_tau else float("nan"),
             "hkgap": (sum(step_hkgap) / len(step_hkgap)) if step_hkgap else float("nan"),
+            "resid": (sum(step_resid) / len(step_resid)) if step_resid else float("nan"),
             "lr": sched.get_last_lr()[0],
             "cos_pre": diag["mean_cos_pre"],
             "cos_pre_s": diag["mean_cos_pre_s"],