Update app.py

app.py

@@ -413,10 +413,12 @@ def compute_abuse_score(matched_scores, sentiment):
     return min(adjusted_score, 100)
     
     
+
 def analyze_single_message(text, thresholds):
+    print("⚡ ENTERED analyze_single_message")
     stage = 1
     motif_hits, matched_phrases = detect_motifs(text)
-    
+
     # Get emotion profile
     emotion_profile = get_emotion_profile(text)
     sentiment_score = emotion_profile.get("anger", 0) + emotion_profile.get("disgust", 0)
@@ -427,7 +429,7 @@ def analyze_single_message(text, thresholds):
         outputs = model(**inputs)
     scores = torch.sigmoid(outputs.logits.squeeze(0)).numpy()
 
-    # Sentiment override if neutral is high while critical thresholds are passed
+    # Sentiment override
     if emotion_profile.get("neutral", 0) > 0.85 and any(
         scores[LABELS.index(l)] > thresholds[l]
         for l in ["control", "blame shifting"]
@@ -442,85 +444,79 @@ def analyze_single_message(text, thresholds):
         k: v + 0.05 if sentiment == "supportive" else v
         for k, v in thresholds.items()
     }
+
     darvo_score = predict_darvo_score(text)
 
     threshold_labels = [
         label for label, score in zip(LABELS, scores)
         if score > adjusted_thresholds[label]
     ]
-    # Suppress all outputs if no abuse labels pass threshold
+
+    # Early exit if nothing passed
     if not threshold_labels:
-        abuse_score = 0.0
-        darvo_score = 0.0
-        top_patterns = []
-        tone_tag = "supportive"
-        return abuse_score, threshold_labels, top_patterns, {"label": sentiment}, stage, darvo_score, tone_tag
+        return 0.0, [], [], {"label": sentiment}, 1, 0.0, "supportive"
 
     top_patterns = sorted(
         [(label, score) for label, score in zip(LABELS, scores)],
         key=lambda x: x[1],
         reverse=True
     )[:2]
-    # Post-threshold validation: strip recovery if it occurs with undermining sentiment
-    if "recovery" in threshold_labels and tone_tag == "forced accountability flip":
-        threshold_labels.remove("recovery")
-        top_patterns = [p for p in top_patterns if p[0] != "recovery"]
-        print("⚠️ Removing 'recovery' due to undermining sentiment (not genuine repair)")
 
     matched_scores = [
         (label, score, PATTERN_WEIGHTS.get(label, 1.0))
         for label, score in zip(LABELS, scores)
         if score > adjusted_thresholds[label]
     ]
-    # Classify insult pattern more precisely for UI
+
+    # Determine insult subtype
     insults_score = next((score for label, score, _ in matched_scores if label == "insults"), 0)
     insult_label_display = None
     if insults_score > 0.9 and (emotion_profile.get("anger", 0) > 0.1 or emotion_profile.get("disgust", 0) > 0.1):
         insult_label_display = "Direct Insult"
     elif 0.5 < insults_score <= 0.9 and emotion_profile.get("neutral", 0) > 0.85:
         insult_label_display = "Subtle Undermining"
+
+    # Abuse score
     abuse_score_raw = compute_abuse_score(matched_scores, sentiment)
-    abuse_score = abuse_score_raw
 
-    # Risk stage logic
-    stage = get_risk_stage(threshold_labels, sentiment) if threshold_labels else 1
-    if weapon_flag and stage < 2:
-        stage = 2
+    # Weapon adjustment
     if weapon_flag:
         abuse_score_raw = min(abuse_score_raw + 25, 100)
-
-
+        if stage < 2:
+            stage = 2
 
     abuse_score = min(abuse_score_raw, 100 if "control" in threshold_labels else 95)
 
-# Tag must happen after abuse score is finalized
+    # Tone tag
     tone_tag = get_emotional_tone_tag(emotion_profile, sentiment, threshold_labels, abuse_score)
-     # Final display tweak: swap 'insults' with its refined label
-    if insult_label_display and "insults" in threshold_labels:
-        threshold_labels = [
-            insult_label_display if label == "insults" else label
-            for label in threshold_labels
-    ]
-    return abuse_score, threshold_labels, top_patterns, {"label": sentiment}, stage, darvo_score, tone_tag 
-# ---- Profanity + Anger Override Logic ----
+
+    # Remove recovery tag if tone is fake
+    if "recovery" in threshold_labels and tone_tag == "forced accountability flip":
+        threshold_labels.remove("recovery")
+        top_patterns = [p for p in top_patterns if p[0] != "recovery"]
+        print("⚠️ Removing 'recovery' due to undermining sentiment (not genuine repair)")
+
+    # Override profanity/anger for short texts
     profane_words = {"fuck", "fucking", "bitch", "shit", "cunt", "ho", "asshole", "dick", "whore", "slut"}
     tokens = set(text.lower().split())
     has_profane = any(word in tokens for word in profane_words)
-
-    anger_score = emotion_profile.get("Anger", 0)
     short_text = len(tokens) <= 10
-    insult_score = next((s for l, s in top_patterns if l == "insults"), 0)
-
+    anger_score = emotion_profile.get("anger", 0)
     if has_profane and anger_score > 0.75 and short_text:
         print("⚠️ Profanity + Anger Override Triggered")
-        top_patterns = sorted(top_patterns, key=lambda x: x[1], reverse=True)
-        if top_patterns[0][0] != "insults":
-            top_patterns.insert(0, ("insults", insult_score))
+        insult_score = next((s for l, s in top_patterns if l == "insults"), 0)
+        if ("insults", insult_score) not in top_patterns:
+            top_patterns = [("insults", insult_score)] + top_patterns
         if "insults" not in threshold_labels:
             threshold_labels.append("insults")
-        top_patterns = [("insults", insult_score)] + [p for p in top_patterns if p[0] != "insults"]
-# Debug
-    print(f"Emotional Tone Tag: {tone_tag}")
+
+    # Replace 'insults' with descriptive label in output
+    if insult_label_display and "insults" in threshold_labels:
+        threshold_labels = [
+            insult_label_display if label == "insults" else label
+            for label in threshold_labels
+        ]
+
     # Debug
     print(f"Emotional Tone Tag: {tone_tag}")
     print("Emotion Profile:")
@@ -538,6 +534,7 @@ def analyze_single_message(text, thresholds):
     print("------------------\n")
 
     return abuse_score, threshold_labels, top_patterns, {"label": sentiment}, stage, darvo_score, tone_tag
+
 import spaces
 
 @spaces.GPU