Update app/services/classifier.py

app/services/classifier.py

@@ -57,24 +57,50 @@ SENSITIVE_KEYWORDS = {
 
 URL_REGEX = re.compile(r"(?:https?://)?(?:www\.)?[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}(?:/[^\s]*)?")
 
+ALLOWED_REASON_CODES = {
+        "game_domain_match",
+        "game_ui_terms",
+        "signal_plus_ocr",
+        "weak_generic_terms",
+        "conflicting_signals",
+        "no_game_evidence",
+}
+
 GAME_PROMPT = """
-You are a strict classifier for parental-control screenshots.
+You are a high-precision parental-control classifier.
+Primary objective: minimize false positives.
+
 Input fields:
 - suspected_game_signal: boolean
 - extracted_urls: list of urls/domains from OCR
 - ocr_text: raw OCR text from screenshot
-Task:
-- Decide if screenshot likely indicates gaming/web-game activity.
-- Return only compact JSON with this schema:
+
+Decision policy:
+1) Return "game" only if at least one strong signal exists:
+     - known game domain/platform in extracted_urls, OR
+     - explicit in-game UI terms in ocr_text (ranked, lobby, battle pass, matchmaking, etc.), OR
+     - suspected_game_signal=true AND at least one medium OCR signal.
+2) Return "uncertain" only when signals are conflicting and neither side is clearly dominant.
+3) Return "not_game" for weak/generic words (game, games, play) without concrete game context.
+4) Prioritize precision over recall.
+
+Output:
+Return EXACT JSON only using this schema:
 {
   "verdict": "game" | "not_game" | "uncertain",
   "confidence": 0.0-1.0,
-  "reason": "short reason"
+    "reason": "reason_code"
 }
-Rules:
-- If clear game domain or game UI terms appear, lean game.
-- If evidence is weak or contradictory, return uncertain.
-- Never output markdown, prose, or extra keys.
+
+Allowed reason codes:
+- game_domain_match
+- game_ui_terms
+- signal_plus_ocr
+- weak_generic_terms
+- conflicting_signals
+- no_game_evidence
+
+Do not output markdown, prose, or extra keys.
 """.strip()
 
 
@@ -260,11 +286,23 @@ def _parse_llm_json(content: str) -> dict[str, Any] | None:
         confidence = 0.5
 
     confidence = max(0.0, min(1.0, confidence))
-    reason = str(data.get("reason", "llm-decision"))[:200]
+    raw_reason = str(data.get("reason", "")).strip().lower()
+    reason = _normalize_reason_code(verdict, raw_reason)
 
     return {"verdict": verdict, "confidence": confidence, "reason": reason}
 
 
+def _normalize_reason_code(verdict: str, reason: str) -> str:
+    if reason in ALLOWED_REASON_CODES:
+        return reason
+
+    if verdict == "game":
+        return "game_ui_terms"
+    if verdict == "uncertain":
+        return "conflicting_signals"
+    return "no_game_evidence"
+
+
 def _domain_from_url(url: str) -> str:
     parsed = urlparse(url)
     host = parsed.netloc or parsed.path