diff --git a/deps/verifiers b/deps/verifiers
index 576506d66f..dd2950909b 160000
--- a/deps/verifiers
+++ b/deps/verifiers
@@ -1 +1 @@
-Subproject commit 576506d66f37a2d6d996a85b8eb2ef53bbfddd9f
+Subproject commit dd2950909bc742a60505a355cc779cf2f39a03b0
diff --git a/src/prime_rl/orchestrator/metrics.py b/src/prime_rl/orchestrator/metrics.py
index c877b7edf8..4835f10a81 100644
--- a/src/prime_rl/orchestrator/metrics.py
+++ b/src/prime_rl/orchestrator/metrics.py
@@ -134,8 +134,11 @@ def stats(self) -> dict[str, Stat]:
 
 class CustomMetrics(StatGroup):
     """Per-key ``Stat``s over a dynamic per-rollout dict attribute (env ``@metric``s or reward
-    components), each averaged over the rollouts that report the key. ``value`` extracts the
-    float from each entry (rewards are ``vf.Reward`` records; metrics are plain floats)."""
+    components), each over the rollouts that carry the key. Scoring seeds every expected key
+    with ``None`` before invoking it, so a ``None`` value means the signal never produced a
+    score and counts as 0.0 — the ``effective`` subset excludes errored rollouts and gives the
+    clean means. ``value`` extracts the float from each scored entry (rewards are ``vf.Reward``
+    records; metrics are plain floats)."""
 
     def __init__(self, rollouts: list[Rollout], attr: str, value: Callable[[Any], float] = float) -> None:
         super().__init__(rollouts)
@@ -146,7 +149,11 @@ def stats(self) -> dict[str, Stat]:
         names = sorted({name for r in self.rollouts for name in getattr(r, self.attr)})
         return {
             name: Stat(
-                [self.value(getattr(r, self.attr)[name]) for r in self.rollouts if name in getattr(r, self.attr)]
+                [
+                    self.value(scores[name]) if scores[name] is not None else 0.0
+                    for r in self.rollouts
+                    if name in (scores := getattr(r, self.attr))
+                ]
             )
             for name in names
         }
diff --git a/tests/unit/orchestrator/test_metrics.py b/tests/unit/orchestrator/test_metrics.py
index 3087721de5..579a1b79dc 100644
--- a/tests/unit/orchestrator/test_metrics.py
+++ b/tests/unit/orchestrator/test_metrics.py
@@ -144,13 +144,25 @@ def test_nested_metrics_and_rewards():
     rollouts = [
         mk(metrics={"acc": 1.0}, rewards={"correct": vf.Reward(score=1.0), "format": vf.Reward(score=0.0)}),
         mk(metrics={"acc": 3.0, "fmt": 5.0}, rewards={"correct": vf.Reward(score=0.0), "format": vf.Reward(score=1.0)}),
+        # scoring failed after seeding: unscored (None) entries count as 0.0 on `all`
+        mk(has_error=True, metrics={"acc": None}, rewards={"correct": None, "format": None}),
     ]
-    m = TrainRollouts(rollouts).metrics
-    assert m.metrics["acc"].mean() == 2.0 and m.rewards["correct"].mean() == 0.5  # nested group access
+    rc = TrainRollouts(rollouts)
+    m = rc.metrics
+    assert m.metrics["acc"].mean() == pytest.approx(4 / 3) and m.rewards["correct"].mean() == pytest.approx(1 / 3)
     out = m.to_wandb(prefix="train/agg", subset="all")
-    assert out["train/agg/all/metrics/acc/mean"] == 2.0  # averaged over reporters
+    assert out["train/agg/all/metrics/acc/mean"] == pytest.approx(4 / 3)
     assert out["train/agg/all/metrics/fmt/mean"] == 5.0  # single reporter
-    assert out["train/agg/all/rewards/format/mean"] == 0.5
+    assert out["train/agg/all/rewards/format/mean"] == pytest.approx(1 / 3)
+    # effective drops the errored rollout, so its seeds don't dilute the effective means
+    eff = rc.effective.metrics.to_wandb(prefix="train/agg", subset="effective")
+    assert eff["train/agg/effective/metrics/acc/mean"] == 2.0
+    assert eff["train/agg/effective/rewards/format/mean"] == 0.5
+    # cross-env agg: another env's unscored trace carries different keys, so it can't dilute these
+    other = mk(env_name="other", has_error=True, rewards={"solved": None})
+    agg = TrainRollouts(rollouts + [other]).metrics.to_wandb(prefix="train/agg", subset="all")
+    assert agg["train/agg/all/rewards/format/mean"] == pytest.approx(1 / 3)
+    assert agg["train/agg/all/rewards/solved/mean"] == 0.0
 
 
 def test_nested_timing():