diff --git a/loopx/capabilities/explore/adaptive_replay_planner.py b/loopx/capabilities/explore/adaptive_replay_planner.py
index cefa76463c..f03100af6e 100644
--- a/loopx/capabilities/explore/adaptive_replay_planner.py
+++ b/loopx/capabilities/explore/adaptive_replay_planner.py
@@ -10,20 +10,38 @@
 from .trace_runtime import TraceLog, require_nonempty_text
 
 
-def _nonnegative(value: float, field_name: str) -> float:
+def _finite_number(value: float, field_name: str, *, allow_zero: bool) -> float:
+    """Normalize a public numeric input without admitting NaN or overflow.
+
+    ``float("nan") < 0`` and ``float("nan") <= 0`` are both false, so a
+    comparison-only guard admits NaN; a NaN then makes every later floor
+    comparison false too, which turns a bound into a pass-through. Booleans are
+    refused because ``float(True)`` is silently ``1.0``, and integers too large
+    for a float are refused so the raw ``OverflowError`` does not escape
+    validation as a non-field-specific error.
+    """
+
+    requirement = "non-negative" if allow_zero else "positive"
+    message = f"{field_name} must be a finite {requirement} number"
     if isinstance(value, bool):
-        raise ValueError(f"{field_name} must be a finite non-negative number")
+        raise ValueError(message)
     try:
         normalized = float(value)
     except (OverflowError, TypeError, ValueError) as exc:
-        raise ValueError(
-            f"{field_name} must be a finite non-negative number"
-        ) from exc
-    if not math.isfinite(normalized) or normalized < 0:
-        raise ValueError(f"{field_name} must be a finite non-negative number")
+        raise ValueError(message) from exc
+    if not math.isfinite(normalized) or normalized < 0 or (normalized == 0 and not allow_zero):
+        raise ValueError(message)
     return normalized
 
 
+def _nonnegative(value: float, field_name: str) -> float:
+    return _finite_number(value, field_name, allow_zero=True)
+
+
+def _positive(value: float, field_name: str) -> float:
+    return _finite_number(value, field_name, allow_zero=False)
+
+
 @dataclass(frozen=True)
 class ReplayPositionCost:
     """Public-safe Adapter estimate for one addressable trace boundary."""
@@ -94,10 +112,7 @@ def __post_init__(self) -> None:
         )
         if self.minimum_fidelity == ReplayFidelity.NON_REPLAYABLE:
             raise ValueError("failure observations require replayable fidelity")
-        normalized_weight = float(self.weight)
-        if normalized_weight <= 0:
-            raise ValueError("weight must be positive")
-        object.__setattr__(self, "weight", normalized_weight)
+        object.__setattr__(self, "weight", _positive(self.weight, "weight"))
 
 
 @dataclass(frozen=True)
@@ -167,9 +182,8 @@ def plan_adaptive_replay_points(
         for observation in normalized_observations
     ):
         raise TypeError("observations must contain FailureReplayObservation values")
-    if int(max_proposals) <= 0:
-        raise ValueError("max_proposals must be positive")
-    threshold = float(minimum_score_minutes)
+    proposal_limit = int(_positive(max_proposals, "max_proposals"))
+    threshold = _nonnegative(minimum_score_minutes, "minimum_score_minutes")
     for candidate in normalized_candidates:
         _validate_position(trace, candidate)
     for observation in normalized_observations:
@@ -226,4 +240,4 @@ def plan_adaptive_replay_points(
             proposal.after_event_id or "",
         )
     )
-    return tuple(proposals[: int(max_proposals)])
+    return tuple(proposals[:proposal_limit])
diff --git a/tests/test_explore_replay_runtime.py b/tests/test_explore_replay_runtime.py
index d579f489e9..96afe42ca5 100644
--- a/tests/test_explore_replay_runtime.py
+++ b/tests/test_explore_replay_runtime.py
@@ -1467,6 +1467,55 @@ def test_replay_cost_rejects_non_finite_or_non_numeric_minutes(invalid_minutes)
         )
 
 
+@pytest.mark.parametrize(
+    "invalid_weight",
+    [True, False, float("nan"), float("inf"), float("-inf"), 10**400, 0, -1.0],
+)
+def test_observation_weight_rejects_non_finite_or_non_positive_values(invalid_weight) -> None:
+    """A NaN weight used to reach a proposal score and defeat every comparison."""
+
+    with pytest.raises(ValueError, match="weight must be a finite positive number"):
+        FailureReplayObservation(
+            1,
+            3,
+            10.0,
+            ReplayFidelity.SEMANTIC_EQUIVALENT,
+            weight=invalid_weight,
+        )
+
+
+@pytest.mark.parametrize(
+    "invalid_floor",
+    [True, float("nan"), float("inf"), float("-inf"), 10**400, -1.0],
+)
+def test_planner_rejects_a_score_floor_that_cannot_bound_a_score(invalid_floor) -> None:
+    """Every comparison against a NaN floor is false, so a NaN floor admits everything."""
+
+    with pytest.raises(
+        ValueError,
+        match="minimum_score_minutes must be a finite non-negative number",
+    ):
+        plan_adaptive_replay_points(
+            _planning_trace(),
+            (ReplayPositionCost(None, 0, ReplayFidelity.EXACT, 1.0, 0.0),),
+            (FailureReplayObservation(1, 3, 10.0, ReplayFidelity.EXACT, weight=1.0),),
+            minimum_score_minutes=invalid_floor,
+        )
+
+
+@pytest.mark.parametrize("invalid_limit", [True, float("nan"), float("inf"), 0, -1])
+def test_planner_rejects_a_proposal_limit_that_is_not_a_positive_integer(
+    invalid_limit,
+) -> None:
+    with pytest.raises(ValueError, match="max_proposals must be"):
+        plan_adaptive_replay_points(
+            _planning_trace(),
+            (ReplayPositionCost(None, 0, ReplayFidelity.EXACT, 1.0, 0.0),),
+            (FailureReplayObservation(1, 3, 10.0, ReplayFidelity.EXACT, weight=1.0),),
+            max_proposals=invalid_limit,
+        )
+
+
 def test_full_scope_composes_adaptive_children_risk_and_promotion() -> None:
     trace = TraceLog("trace-full-scope", "branch-main")
     prefix = TraceEvent(