diff --git a/verifiers/v1/clients/train.py b/verifiers/v1/clients/train.py
index 6b7e7fad8e..0410b4c9e4 100644
--- a/verifiers/v1/clients/train.py
+++ b/verifiers/v1/clients/train.py
@@ -32,6 +32,7 @@
     ToolCall,
     TurnTokens,
     Usage,
+    parse_sampled_logprobs,
 )
 
 
@@ -116,6 +117,14 @@ def response_from_generate(
     ] or None
     prompt_ids = result.get("prompt_ids") or []
     completion_ids = result.get("completion_ids") or []
+    try:
+        completion_logprobs = parse_sampled_logprobs(
+            result.get("completion_logprobs", []),
+            len(completion_ids),
+            what="generate response",
+        )
+    except ValueError as e:
+        raise model_error(str(e)) from e
     # Per-message token spans (the renderer's attribution) let the trace graph store each
     # message's tokens once; carried transiently on TurnTokens and consumed by turn.commit().
     attribution = result.get("prompt_attribution")
@@ -140,12 +149,12 @@ def response_from_generate(
         usage=Usage(
             prompt_tokens=len(prompt_ids), completion_tokens=len(completion_ids)
         ),
-        # generate() returns owned, typed lists. Skip revalidation here to avoid copying
-        # million-token contexts synchronously on the event loop.
+        # The token lists are generate()-owned. Validate the small sampled-evidence list above,
+        # then bypass model copies of the potentially million-token context lists.
         tokens=TurnTokens.model_construct(
             prompt_ids=prompt_ids,
             completion_ids=completion_ids,
-            completion_logprobs=result.get("completion_logprobs") or [],
+            completion_logprobs=completion_logprobs,
             message_spans=message_spans,
             is_content=attribution.is_content if attribution is not None else None,
             multi_modal_data=result.get("multi_modal_data"),
diff --git a/verifiers/v1/trace.py b/verifiers/v1/trace.py
index ae90009e99..8757cfcb18 100644
--- a/verifiers/v1/trace.py
+++ b/verifiers/v1/trace.py
@@ -34,6 +34,7 @@
     ToolMessage,
     Usage,
     content_text,
+    parse_sampled_logprobs,
 )
 
 logger = logging.getLogger(__name__)
@@ -152,18 +153,31 @@ def logprobs(self) -> list[float]:
         """Per-token sampling logprobs aligned to `token_ids` — the node logprobs spread onto
         their sampled positions, 0.0 on every non-sampled token."""
         out: list[float] = []
-        for node in self.nodes:
+        for node_index, node in enumerate(self.nodes):
             mask = node.mask
-            sampled = sum(mask) if node.logprobs else 0
+            if len(mask) != len(node.token_ids):
+                raise ValueError(
+                    f"branch {self.index} node {node_index} has {len(mask)} mask entries "
+                    f"for {len(node.token_ids)} token ids"
+                )
+            if any(not isinstance(sampled, bool) for sampled in mask):
+                raise ValueError(
+                    f"branch {self.index} node {node_index} mask entries must be booleans"
+                )
+            sampled = sum(mask)
+            logprobs = parse_sampled_logprobs(
+                node.logprobs,
+                sampled,
+                what=f"branch {self.index} node {node_index}",
+            )
             # Bulk-fill the canonical unsampled-prefix/sampled-suffix layout.
             if not sampled or all(mask[-sampled:]):
-                out += [0.0] * (len(mask) - sampled) + node.logprobs[:sampled]
-                out += [0.0] * max(0, sampled - len(node.logprobs))
+                out += [0.0] * (len(mask) - sampled) + logprobs
                 continue
             li = 0
-            for sampled in mask:
-                if sampled:
-                    out.append(node.logprobs[li] if li < len(node.logprobs) else 0.0)
+            for is_sampled in mask:
+                if is_sampled:
+                    out.append(logprobs[li])
                     li += 1
                 else:
                     out.append(0.0)
diff --git a/verifiers/v1/types.py b/verifiers/v1/types.py
index 7ee25f8502..b1bef3f4b6 100644
--- a/verifiers/v1/types.py
+++ b/verifiers/v1/types.py
@@ -1,3 +1,4 @@
+import math
 from collections.abc import Iterable
 from dataclasses import dataclass
 from typing import Annotated, Any, Literal
@@ -32,6 +33,31 @@ class ImageUrlContentPart(StrictBaseModel):
 """Plain text or typed multimodal content parts."""
 
 
+def parse_sampled_logprobs(
+    values: object, sampled_count: int, *, what: str
+) -> list[float]:
+    """Parse one finite real logprob per sampled token into built-in floats."""
+    if not isinstance(values, list):
+        raise ValueError(f"{what} logprobs must be a list")
+    if len(values) != sampled_count:
+        raise ValueError(
+            f"{what} has {len(values)} logprobs for {sampled_count} sampled tokens"
+        )
+
+    parsed: list[float] = []
+    for value in values:
+        if isinstance(value, bool) or not isinstance(value, (int, float)):
+            raise ValueError(f"{what} logprobs must be finite real numbers")
+        try:
+            value = float(value)
+        except OverflowError as exc:
+            raise ValueError(f"{what} logprobs must be finite real numbers") from exc
+        if not math.isfinite(value):
+            raise ValueError(f"{what} logprobs must be finite real numbers")
+        parsed.append(value)
+    return parsed
+
+
 def content_to_parts(content) -> MessageContent:
     """Type OpenAI content parts, dropping unsupported part types."""
     if not isinstance(content, list):