--- a/code/losses.py
+++ b/code/losses.py
@@ -36,13 +36,24 @@
     n = int(m.sum())
     if n == 0:
         return {}
-    lp, olp, a = logp[m].double(), old_lp[m].double(), adv[m].double()
+    # Match the loss dtype for decisions; use float64 only for drift summaries.
+    lp, olp, a = logp[m], old_lp[m], adv[m]
     p, mu, r = lp.exp(), olp.exp(), (lp - olp).exp()
-    dppo_masked = torch.where(a > 0, (p - mu) > delta, (p - mu) < -delta)
+    dppo_masked = ((a > 0) & ((p - mu) > delta)) | ((a < 0) & ((p - mu) < -delta))
     grpo_clipped = ((a > 0) & (r > 1 + eps)) | ((a < 0) & (r < 1 - eps))
-    absdp = (p - mu).abs()
+    drift = p.double() - mu.double()
+    absdp = drift.abs()
+    active = a != 0
     return {
         "n_loss_tokens": float(n),
+        "n_nonzero_adv_tokens": int(active.sum()),
+        "dppo_removed_gradient_tokens": int(dppo_masked.sum()),
+        "dppo_masked_nonzero_adv_fraction": float(dppo_masked[active].double().mean()) if active.any() else None,
+        "policy_drift_abs_p_q99": float(torch.quantile(absdp, 0.99)),
+        "policy_drift_abs_p_max": float(absdp.max()),
+        "policy_drift_signed_p_min": float(drift.min()),
+        "policy_drift_signed_p_max": float(drift.max()),
+        "abs_drift_above_delta_tokens": int((absdp > delta).sum()),
         "dppo_masked_token_fraction": float(dppo_masked.double().mean()),
         "grpo_clip_fraction": float(grpo_clipped.double().mean()),
         "policy_drift_abs_p_q90": float(torch.quantile(absdp.float(), 0.9)),