Update app/services/classifier.py

app/services/classifier.py

@@ -71,6 +71,11 @@ CONTRAST_NOT_GAME_LABELS = [
     "a screen being shared via Google Meet with a blue stop sharing button at the bottom",
     "a full screen presentation with a meet.google.com sharing notification bar",
     "a computer desktop being recorded or shared in a video meeting",
+
+    # Text-heavy educational/article pages (common false-positive vs browser game labels)
+    "a Vietnamese educational article webpage with long text paragraphs and lesson navigation sidebar",
+    "an online lesson page with school grade menu and literary analysis text content",
+    "a reading article website with dense paragraphs and no gameplay controls or score HUD",
 ]
 
 NOT_GAME_LABELS = [
@@ -207,6 +212,22 @@ def _is_gallery_like_not_game(description: str) -> bool:
         )
     )
 
+
+def _is_text_heavy_not_game(description: str) -> bool:
+    text = description.lower()
+    return any(
+        token in text
+        for token in (
+            "educational",
+            "lesson",
+            "article",
+            "literary analysis",
+            "dense paragraphs",
+            "reading article",
+            "text paragraphs",
+        )
+    )
+
 logger = logging.getLogger(__name__)
 
 SENSITIVE_KEYWORDS = {
@@ -545,6 +566,23 @@ def _classify_game_with_zero_shot(file_path: str, suspected_game: bool) -> dict[
             "visual_summary": visual_summary,
         }
 
+    # Browser-game labels are also confused by text-heavy educational/article websites.
+    # If a strong text-page non-game label competes with broad browser-game cues, prefer not_game.
+    if (
+        top["index"] in GAME_ZERO_SHOT_BROWSER_GAME_LABEL_INDICES
+        and top_game_score < 0.90
+        and not_game_rows
+        and float(not_game_rows[0]["score"]) >= 0.12
+        and _is_text_heavy_not_game(str(not_game_rows[0]["description"]))
+    ):
+        return {
+            "verdict": "not_game",
+            "confidence": float(not_game_rows[0]["score"]),
+            "reason": "text_article_conflict",
+            "source": f"zero-shot:{GAME_ZERO_SHOT_MODEL}",
+            "visual_summary": visual_summary,
+        }
+
     # Clear game: high confidence
     if game_score >= GAME_ZERO_SHOT_GAME_THRESHOLD:
         return {