Update app.py

app.py

@@ -560,22 +560,50 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
     sentiments = [r[0][3]['label'] for r in results]
     stages = [r[0][4] for r in results]
     darvo_scores = [r[0][5] for r in results]
-    tone_tags= [r[0][6] for r in results]
+    tone_tags = [r[0][6] for r in results]
     dates_used = [r[1] or "Undated" for r in results]  # Store dates for future mapping
-    # Calculate escalation bump *after* model results exist
+
+    # Calculate escalation bump
     escalation_bump = 0
     for result, _ in results:
         abuse_score, threshold_labels, top_patterns, sentiment, stage, darvo_score, tone_tag = result
+
         if darvo_score > 0.65:
             escalation_bump += 3
-        if tone_tag in ["forced accountability flip", "emotional threat"]:
+
+        if tone_tag in ["forced accountability flip", "emotional threat", "aggressive dismissal", "mocking detachment"]:
             escalation_bump += 2
+
         if abuse_score > 80:
             escalation_bump += 2
+
         if stage == 2:
             escalation_bump += 3
 
-    # Now we can safely calculate hybrid_score
+        if "threat" in threshold_labels or "aggression" in threshold_labels:
+            escalation_bump += 4
+
+    # Helper: score trend of pattern severity
+    def message_severity_index(threshold_labels):
+        weights = {
+            "recovery": 0,
+            "dismissiveness": 1,
+            "deflection": 1,
+            "guilt tripping": 2,
+            "control": 3,
+            "gaslighting": 3,
+            "aggression": 4,
+            "threat": 5,
+        }
+        return max([weights.get(label, 0) for label in threshold_labels], default=0)
+
+    severity_levels = [message_severity_index(r[0][1]) for r in results]
+    if len(severity_levels) >= 2 and severity_levels == sorted(severity_levels):
+        escalation_bump += 2
+
+    escalation_bump = min(escalation_bump, 10)
+
+    # Final escalation score
     hybrid_score = escalation_score + escalation_bump if escalation_score is not None else 0
     risk_level = (
         "High" if hybrid_score >= 16 else
@@ -583,10 +611,6 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
         "Low"
     )
 
-    # Now compute scores and allow override
-    abuse_scores = [r[0][0] for r in results]
-    stages = [r[0][4] for r in results]
-
     # Post-check override (e.g. stage 2 or high abuse score forces Moderate risk)
     if any(score > 70 for score in abuse_scores) or any(stage == 2 for stage in stages):
         if risk_level == "Low":
@@ -597,27 +621,18 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
 
     # --- Composite Abuse Score with Weighted Patterns ---
     composite_abuse_scores = []
-
     for result, _ in results:
         abuse_score, threshold_labels, top_patterns, _, _, _, _ = result
         weighted_score = 0
         total_weight = 0
-
         for label, score in top_patterns:
             weight = PATTERN_WEIGHTS.get(label, 1.0)
             weighted_score += score * weight
             total_weight += weight
-
-        if total_weight > 0:
-            final_score = (weighted_score / total_weight) * 100
-        else:
-            final_score = 0
-
+        final_score = (weighted_score / total_weight) * 100 if total_weight > 0 else 0
         composite_abuse_scores.append(final_score)
 
     composite_abuse = int(round(sum(composite_abuse_scores) / len(composite_abuse_scores)))
-
-
     most_common_stage = max(set(stages), key=stages.count)
     stage_text = RISK_STAGE_LABELS[most_common_stage]
 
@@ -630,27 +645,27 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
     out = f"Abuse Intensity: {composite_abuse}%\n"
     out += "📊 This reflects the strength and severity of detected abuse patterns in the message(s).\n\n"
 
-# Save this line for later use at the 
     if escalation_score is None:
         escalation_text = "📉 Escalation Potential: Unknown (Checklist not completed)\n"
         escalation_text += "⚠️ *This section was not completed. Escalation potential is unknown.*\n"
-        hybrid_score = 0  # ✅ fallback so it's defined for generate_risk_snippet
     else:
         escalation_text = f"🧨 **Escalation Potential: {risk_level} ({escalation_score}/{sum(w for _, w in ESCALATION_QUESTIONS)})**\n"
         escalation_text += "This score comes directly from the safety checklist and functions as a standalone escalation risk score.\n"
         escalation_text += "It indicates how many serious risk factors are present based on your answers to the safety checklist.\n"
-        # Derive top_label from the strongest top_patterns across all messages
+
+    # Derive top label
     top_label = None
     if results:
         sorted_patterns = sorted(
             [(label, score) for r in results for label, score in r[0][2]],
             key=lambda x: x[1],
             reverse=True
-    )
+        )
         if sorted_patterns:
             top_label = f"{sorted_patterns[0][0]} – {int(round(sorted_patterns[0][1] * 100))}%"
         if top_label is None:
             top_label = "Unknown – 0%"
+
     out += generate_risk_snippet(composite_abuse, top_label, hybrid_score if escalation_score is not None else 0, most_common_stage)
     out += f"\n\n{stage_text}"
     out += darvo_blurb
@@ -658,6 +673,7 @@ def analyze_composite(msg1, date1, msg2, date2, msg3, date3, *answers_and_none):
     for i, tone in enumerate(tone_tags):
         label = tone if tone else "none"
         out += f"• Message {i+1}: *{label}*\n"
+
     print(f"DEBUG: avg_darvo = {avg_darvo}")
     pattern_labels = [r[0][2][0][0] for r in results]  # top label for each message
     timeline_image = generate_abuse_score_chart(dates_used, abuse_scores, pattern_labels)