Update app.py

app.py

@@ -621,226 +621,155 @@ def split_content_in_half(text: str) -> tuple:
 
 def analyze_content_halves(model_manager, text: str, overall_result: Dict = None) -> Dict:
     """
-    Analyze text by splitting it into two halves after cleaning
+    Analyze text by splitting it into two halves after cleaning.
     Uses BOTH models for ensemble predictions on each half for improved accuracy
-    PLUS advanced linguistic analysis for 100% accuracy confidence
-    
-    Args:
-        model_manager: The ModelManager instance
-        text: Original text to analyze
-        overall_result: Overall classification result for variance calculation
-    
-    Returns:
-        Dictionary with analysis of both halves, linguistic features, and final decision
+    PLUS advanced linguistic analysis for enhanced confidence.
     """
+
     try:
-        # 🆕 STEP 1: Advanced Linguistic Analysis on full text
         logger.info("🔬 Running advanced linguistic analysis...")
         linguistic_analysis = advanced_linguistic_analysis(text)
-        
-        # Clean the content first
+
         cleaned_text = clean_content_for_analysis(text)
-        
         if not cleaned_text or len(cleaned_text.split()) < 10:
             return {
                 "halves_analysis_available": False,
                 "reason": "Content too short after cleaning",
                 "linguistic_analysis": linguistic_analysis
             }
-        
-        # Split into halves
+
+        # Split text into halves
         first_half, second_half = split_content_in_half(cleaned_text)
-        
-        # 🆕 STEP 2: Linguistic analysis on each half
-        logger.info("🔬 Analyzing first half linguistics...")
+
+        # Linguistic analysis for each half
         first_half_linguistic = advanced_linguistic_analysis(first_half)
-        
-        logger.info("🔬 Analyzing second half linguistics...")
         second_half_linguistic = advanced_linguistic_analysis(second_half)
-        
-        # Analyze first half using BOTH models (ensemble prediction)
-        logger.info("🔍 Analyzing first half with both models...")
+
+        # Ensemble model predictions
         first_half_result = model_manager.classify_text(first_half)
-        first_half_words = len(first_half.split())
-        
-        # Analyze second half using BOTH models (ensemble prediction)
-        logger.info("🔍 Analyzing second half with both models...")
         second_half_result = model_manager.classify_text(second_half)
-        second_half_words = len(second_half.split())
-        
-        # Extract key metrics
+
         first_ai = first_half_result["ai_percentage"]
         second_ai = second_half_result["ai_percentage"]
         first_model = first_half_result["predicted_model"]
         second_model = second_half_result["predicted_model"]
-        
-        # Get top predictions from both halves for comparison
+
         first_top5 = first_half_result.get("top_5_predictions", [])
         second_top5 = second_half_result.get("top_5_predictions", [])
-        
-        # Calculate average AI score from both halves
+        first_half_words = len(first_half.split())
+        second_half_words = len(second_half.split())
+
+        # Stats
         avg_halves_ai_score = (first_ai + second_ai) / 2
-        
-        # Calculate variance between halves
         variance_between_halves = abs(first_ai - second_ai)
-        
-        # Overall AI probability (use overall_result if provided, otherwise calculate)
-        overall_ai_prob = overall_result["ai_percentage"] / 100 if overall_result else avg_halves_ai_score / 100
-        
-        # Check model consistency across halves
+        overall_ai_prob = (
+            overall_result["ai_percentage"] / 100
+            if overall_result
+            else avg_halves_ai_score / 100
+        )
+
         models_agree = first_model == second_model
-        
-        # Calculate confidence boost from using both models
         models_used = first_half_result.get("models_used", 1)
         ensemble_confidence_boost = "High" if models_used > 1 else "Low"
-        
-        logger.info(f"✅ First half: {first_ai}% AI ({first_model}) | Second half: {second_ai}% AI ({second_model})")
-        logger.info(f"📊 Models used per half: {models_used} | Agreement: {models_agree}")
-        
-        # ===== FINAL DECISION LOGIC WITH LINGUISTIC ENHANCEMENT =====
-        verdict = None
-        confidence = None
-        reasoning = None
-        accuracy_percentage = 0.0
-        
-        # Get linguistic scores
-        ling_ai = linguistic_analysis["linguistic_ai_score"]
-        ling_human = linguistic_analysis["linguistic_human_score"]
-        
-        # 🆕 Combine Model predictions + Linguistic analysis
-        # Weighted average: 70% model predictions, 30% linguistic analysis
-        combined_first_ai = (first_ai * 0.7) + (ling_ai * 0.3)
-        combined_second_ai = (second_ai * 0.7) + (ling_ai * 0.3)
+
+        # Linguistic AI/Human scores
+        ling_ai = linguistic_analysis.get("linguistic_ai_score", 50)
+        ling_human = linguistic_analysis.get("linguistic_human_score", 50)
+
+        # Some fallback linguistic details
+        burstiness = linguistic_analysis.get("burstiness", 0.5)
+        formality_score = linguistic_analysis.get("formality_score", 0.5)
+        human_error_score = linguistic_analysis.get("human_error_score", 0.5)
+        emotion_markers = linguistic_analysis.get("emotion_markers", 0)
+
+        # Weighted average between model and linguistic results
         combined_avg_ai = (avg_halves_ai_score * 0.7) + (ling_ai * 0.3)
-        
-        # Calculate agreement between models and linguistic analysis
-        model_ling_agreement = abs(avg_halves_ai_score - ling_ai) < 20  # Within 20%
-        
-        # 🎯 Enhanced Decision Logic
-        # Condition 1: Both halves < 50% AI AND second_half predicted_model is "human"
+        model_ling_agreement = abs(avg_halves_ai_score - ling_ai) < 20
+
+        # ----- Final Decision Logic -----
+        verdict = "UNCERTAIN"
+        confidence = "Low"
+        accuracy_percentage = 60
+        reasoning = ""
+
+        # HUMAN
         if first_ai < 50 and second_ai < 50 and second_model.lower() == "human":
             verdict = "HUMAN"
-            
-            # Boost confidence if linguistic analysis agrees
             if ling_human > ling_ai:
                 confidence = "Very High"
-                accuracy_percentage = 95 + min(5, human_errors["human_error_score"] * 2)
+                accuracy_percentage = 95
             elif variance_between_halves < 15:
                 confidence = "High"
-                accuracy_percentage = 85 + min(10, human_errors["human_error_score"])
+                accuracy_percentage = 85
             else:
                 confidence = "Medium"
                 accuracy_percentage = 75
-            
+
             reasoning = (
-                f"Both halves scored below 50% AI probability (First: {first_ai}%, Second: {second_ai}%) "
-                f"using ensemble prediction from {models_used} model(s). "
+                f"Both halves scored below 50% AI probability (First: {first_ai}%, Second: {second_ai}%). "
                 f"Linguistic analysis confirms with {ling_human:.1f}% human indicators. "
-                f"The second half was classified as human-written. "
-                f"Detected {human_errors['emotion_markers']} emotional markers and "
-                f"{human_errors['human_error_score']:.1f} human error patterns. "
-                f"Variance between halves is {variance_between_halves:.2f}%, indicating "
-                f"{'consistent human patterns' if variance_between_halves < 15 else 'some variation but still human-like'}. "
-                f"Model predictions {'agree' if models_agree else 'differ'} across halves."
+                f"The text shows {emotion_markers} emotional markers and a human error score of {human_error_score:.2f}. "
+                f"Variance between halves is {variance_between_halves:.2f}%, indicating consistent human patterns. "
             )
-        
-        # Condition 2: Both halves > 50% AI AND second_half predicted_model is NOT "human"
+
+        # AI
         elif first_ai > 50 and second_ai > 50 and second_model.lower() != "human":
             verdict = "AI"
-            
-            # Determine confidence based on scores and linguistic agreement
             if first_ai > 80 and second_ai > 80 and model_ling_agreement:
                 confidence = "Very High"
-                accuracy_percentage = 95 + min(5, (first_ai + second_ai) / 40)
-            elif first_ai > 70 and second_ai > 70 and model_ling_agreement:
-                confidence = "High"
-                accuracy_percentage = 85 + min(10, (first_ai + second_ai) / 50)
+                accuracy_percentage = 98
             elif first_ai > 70 and second_ai > 70:
                 confidence = "High"
-                accuracy_percentage = 80
+                accuracy_percentage = 90
             else:
                 confidence = "Medium"
-                accuracy_percentage = 70
-            
-            # Boost confidence if models agree and linguistic analysis confirms
-            if models_agree and model_ling_agreement:
-                if confidence == "High":
-                    confidence = "Very High"
-                    accuracy_percentage = min(99, accuracy_percentage + 10)
-                elif confidence == "Medium":
-                    confidence = "High"
-                    accuracy_percentage = min(95, accuracy_percentage + 10)
-            
+                accuracy_percentage = 80
+
             reasoning = (
-                f"Both halves scored above 50% AI probability (First: {first_ai}%, Second: {second_ai}%) "
-                f"using ensemble prediction from {models_used} model(s). "
+                f"Both halves scored above 50% AI probability (First: {first_ai}%, Second: {second_ai}%). "
                 f"Linguistic analysis confirms with {ling_ai:.1f}% AI indicators. "
-                f"Detected high formality score ({vocabulary['formality_score']:.2f}) and "
-                f"low burstiness ({burstiness:.2f}), typical of AI generation. "
-                f"Pattern matches {second_model} outputs. "
-                f"First half suggests {first_model} while second half suggests {second_model}. "
-                f"Variance between halves is {variance_between_halves:.2f}%, "
-                f"{'showing consistent AI patterns throughout' if variance_between_halves < 20 else 'with some variation in AI generation style'}. "
-                f"{'Both halves agree on the AI model type, strengthening confidence' if models_agree else 'Different AI models detected in each half'}. "
-                f"Model-linguistic agreement: {'Yes' if model_ling_agreement else 'Partial'}."
+                f"Detected high formality score ({formality_score:.2f}) and low burstiness ({burstiness:.2f}), typical of AI generation. "
+                f"Variance between halves: {variance_between_halves:.2f}%. "
+                f"Models {'agree' if models_agree else 'disagree'} across halves."
             )
-        
-        # Condition 3: Mixed results - one half AI, one half human
+
+        # MIXED
         elif (first_ai > 50 and second_ai < 50) or (first_ai < 50 and second_ai > 50):
             verdict = "MIXED"
-            confidence = "Medium" if abs(first_ai - second_ai) > 30 else "Low"
-            accuracy_percentage = 70 + min(15, variance_between_halves / 3)
-            
+            confidence = "Medium" if variance_between_halves > 30 else "Low"
+            accuracy_percentage = 75
             reasoning = (
-                f"Mixed signals detected using {models_used} model(s) for ensemble prediction. "
-                f"First half: {first_ai}% AI ({first_model}), Second half: {second_ai}% AI ({second_model}). "
-                f"Linguistic analysis shows {ling_ai:.1f}% AI indicators overall. "
-                f"One portion appears AI-generated while the other seems human-written. "
-                f"This could indicate: partial AI assistance, human editing of AI content, "
-                f"or AI completion of human-started text. High variance of {variance_between_halves:.2f}% supports mixed authorship. "
-                f"Burstiness score of {burstiness:.2f} suggests irregular patterns."
+                f"Mixed signals detected. First half: {first_ai}% AI ({first_model}), "
+                f"Second half: {second_ai}% AI ({second_model}). "
+                f"Linguistic AI score: {ling_ai:.1f}%. "
+                f"Variance between halves ({variance_between_halves:.2f}%) supports mixed authorship."
             )
-        
-        # Condition 4: Both around 50% - uncertain
+
+        # Borderline
         else:
-            # Check if second_model is human but scores are borderline
             if second_model.lower() == "human" or ling_human > ling_ai:
                 verdict = "LIKELY_HUMAN"
-                confidence = "Medium" if ling_human - ling_ai > 10 else "Low"
-                accuracy_percentage = 60 + min(15, abs(ling_human - ling_ai))
-                
-                reasoning = (
-                    f"Borderline case with scores near 50% threshold (First: {first_ai}%, Second: {second_ai}%) "
-                    f"analyzed using {models_used} model(s). "
-                    f"Linguistic analysis leans toward human ({ling_human:.1f}% vs {ling_ai:.1f}% AI). "
-                    f"Second half classified as human-written. The text shows characteristics of both "
-                    f"human and AI writing. Variance: {variance_between_halves:.2f}%. "
-                    f"Human error score: {human_errors['human_error_score']:.2f}."
-                )
+                confidence = "Medium"
+                accuracy_percentage = 70
             else:
                 verdict = "LIKELY_AI"
-                confidence = "Medium" if ling_ai - ling_human > 10 else "Low"
-                accuracy_percentage = 60 + min(15, abs(ling_ai - ling_human))
-                
-                reasoning = (
-                    f"Borderline case with scores near 50% threshold (First: {first_ai}%, Second: {second_ai}%) "
-                    f"analyzed using {models_used} model(s). "
-                    f"Linguistic analysis leans toward AI ({ling_ai:.1f}% vs {ling_human:.1f}% human). "
-                    f"Pattern suggests {second_model} but confidence is low. "
-                    f"Variance: {variance_between_halves:.2f}%. "
-                    f"Formality score: {vocabulary['formality_score']:.2f}."
-                )
-        
-        # Prepare final decision structure with enhanced model and linguistic information
+                confidence = "Medium"
+                accuracy_percentage = 70
+
+            reasoning = (
+                f"Borderline case: scores near 50%. "
+                f"Linguistic analysis leans toward {'human' if ling_human > ling_ai else 'AI'} writing. "
+                f"Variance: {variance_between_halves:.2f}%."
+            )
+
+        # ----- Final Output -----
         final_decision = {
             "verdict": verdict,
             "confidence": confidence,
-            "accuracy_percentage": round(accuracy_percentage, 1),
+            "accuracy_percentage": accuracy_percentage,
             "reasoning": reasoning,
             "supporting_data": {
                 "overall_ai_prob": round(overall_ai_prob, 3),
-                "first_half_ai_score": round(first_ai / 100, 3),
-                "second_half_ai_score": round(second_ai / 100, 3),
                 "avg_halves_ai_score": round(avg_halves_ai_score / 100, 3),
                 "variance_between_halves": round(variance_between_halves, 2),
                 "first_half_model": first_model,
@@ -848,20 +777,19 @@ def analyze_content_halves(model_manager, text: str, overall_result: Dict = None
                 "models_agree": models_agree,
                 "ensemble_models_used": models_used,
                 "ensemble_confidence": ensemble_confidence_boost,
-                # 🆕 Linguistic analysis scores
                 "linguistic_ai_score": ling_ai,
                 "linguistic_human_score": ling_human,
                 "model_linguistic_agreement": model_ling_agreement,
-                "combined_ai_score": round(combined_avg_ai, 2)
-            }
+                "combined_ai_score": round(combined_avg_ai, 2),
+            },
         }
-        
+
         return {
             "halves_analysis_available": True,
             "cleaned_content": {
                 "total_words": len(cleaned_text.split()),
                 "first_half_words": first_half_words,
-                "second_half_words": second_half_words
+                "second_half_words": second_half_words,
             },
             "first_half": {
                 "ai_percentage": first_ai,
@@ -871,7 +799,7 @@ def analyze_content_halves(model_manager, text: str, overall_result: Dict = None
                 "preview": first_half[:200] + "..." if len(first_half) > 200 else first_half,
                 "top_5_predictions": first_top5,
                 "models_used": models_used,
-                "linguistic_analysis": first_half_linguistic  # 🆕 Linguistic analysis for first half
+                "linguistic_analysis": first_half_linguistic,
             },
             "second_half": {
                 "ai_percentage": second_ai,
@@ -881,12 +809,12 @@ def analyze_content_halves(model_manager, text: str, overall_result: Dict = None
                 "preview": second_half[:200] + "..." if len(second_half) > 200 else second_half,
                 "top_5_predictions": second_top5,
                 "models_used": models_used,
-                "linguistic_analysis": second_half_linguistic  # 🆕 Linguistic analysis for second half
+                "linguistic_analysis": second_half_linguistic,
             },
             "final_decision": final_decision,
-            "overall_linguistic_analysis": linguistic_analysis  # 🆕 Overall linguistic analysis
+            "overall_linguistic_analysis": linguistic_analysis,
         }
-        
+
     except Exception as e:
         logger.error(f"Error in halves analysis: {e}", exc_info=True)
         return {