diff --git a/src/lib/intake-extract.ts b/src/lib/intake-extract.ts
index 189b752..98919a3 100644
--- a/src/lib/intake-extract.ts
+++ b/src/lib/intake-extract.ts
@@ -185,9 +185,19 @@ export function liteLLMExtractor(opts: {
     });
     if (!response.ok) throw new ExtractionError(`LiteLLM returned ${response.status}`);
     const raw = (await response.json()) as { choices?: { message?: { content?: string }; finish_reason?: string }[] };
-    // A truncated reply can still parse once the JSON is closed, which quietly
-    // drops the tail of a long receipt. Losing lines silently is worse than failing.
-    if (raw.choices?.[0]?.finish_reason === "length") throw new ExtractionError("The receipt was too long to read in one pass.");
+    // Only `stop` means the strict schema was satisfied. A truncated reply can
+    // still parse once the JSON is closed, which quietly drops the tail of a
+    // long receipt; a `content_filter` or `tool_calls` stop is a different reply
+    // shape entirely. Accepting anything but `stop` accepts a reply the schema
+    // never constrained — losing lines silently is worse than failing.
+    const finish = raw.choices?.[0]?.finish_reason;
+    if (finish && finish !== "stop") {
+      throw new ExtractionError(
+        finish === "length"
+          ? "The receipt was too long to read in one pass."
+          : `The model stopped before it finished (${finish}).`
+      );
+    }
     const content = raw.choices?.[0]?.message?.content;
     if (!content) throw new ExtractionError("LiteLLM returned no content");
     try {