Upload bayesian_engine.py with huggingface_hub

bayesian_engine.py

@@ -297,31 +297,59 @@ def bayesian_synthesis(agent_results: List[AgentEvidence]) -> ForensicVerdict:
         conf_label = "High"
     
     # Compute confidence based on agreement strength and active agent count
-    non_failed = [s for s, f in zip(scores, failure_probs) if f < 0.5]
-    if non_failed:
-        avg_score = float(np.mean(non_failed))
-        n_agents = len(non_failed)
+    # Exclude near-zero agents from confidence: an agent saying "I don't know" 
+    # should contribute zero, not drag confidence down
+    NEAR_ZERO = 0.02
+    non_failed = [(s, f) for s, f in zip(scores, failure_probs) if f < 0.5]
+    informative = [s for s, f in non_failed if abs(s) > NEAR_ZERO]
+    n_total_active = len(non_failed)
+    
+    if informative:
+        avg_informative = float(np.mean(informative))
+        n_informative = len(informative)
         
         # Count agents agreeing with the majority direction
-        signs = [1 if s > 0.02 else (-1 if s < -0.02 else 0) for s in non_failed]
+        signs = [1 if s > 0 else -1 for s in informative]
         n_pos = sum(1 for s in signs if s > 0)
         n_neg = sum(1 for s in signs if s < 0)
         n_agreeing = max(n_pos, n_neg)
-        n_directional = n_pos + n_neg
+        n_dissenting = min(n_pos, n_neg)
         
-        if n_directional == 0:
-            confidence_numeric = 0.15
-        elif n_pos > 0 and n_neg > 0:
-            # Mixed — penalize
-            agreement_ratio = max(n_pos, n_neg) / n_directional
-            confidence_numeric = min(1.0, 0.1 + 0.3 * abs(avg_score) * agreement_ratio)
+        if n_pos > 0 and n_neg > 0:
+            # Mixed — but scale with how strong the majority is
+            # 3:1 ratio should give decent confidence; 1:1 should give low
+            majority_ratio = n_agreeing / (n_agreeing + n_dissenting)
+            
+            # Majority scores only (ignore dissent magnitude for base)
+            majority_dir = 1 if n_pos > n_neg else -1
+            majority_scores = [s for s in informative if (s > 0) == (majority_dir > 0)]
+            majority_avg = abs(float(np.mean(majority_scores))) if majority_scores else 0
+            
+            # Strong majority (>=75%) with decent magnitude → reasonable confidence
+            # Weak majority (50-60%) → low confidence
+            agent_bonus = min(1.0, np.sqrt(n_agreeing / 2.0))
+            coverage = n_total_active / 7.0
+            
+            if majority_ratio >= 0.75:
+                # 3:1 or better — this is real agreement with a dissenter
+                confidence_numeric = min(1.0, 0.12 + 0.5 * majority_avg * agent_bonus * coverage)
+                if n_agreeing >= 3:
+                    confidence_numeric = min(1.0, confidence_numeric + 0.06)
+            else:
+                # Near 50:50 — genuinely ambiguous
+                confidence_numeric = min(1.0, 0.1 + 0.2 * abs(avg_informative) * majority_ratio)
         else:
-            # All agree — compound with agent count
+            # All informative agents agree — compound with count
             agent_bonus = min(1.0, np.sqrt(n_agreeing / 2.0))
-            coverage = n_agents / 7.0
-            confidence_numeric = min(1.0, 0.15 + 0.6 * abs(avg_score) * agent_bonus * coverage)
-            if n_agreeing >= 4:
-                confidence_numeric = min(1.0, confidence_numeric + 0.1)
+            coverage = n_total_active / 7.0
+            confidence_numeric = min(1.0, 0.15 + 0.6 * abs(avg_informative) * agent_bonus * coverage)
+            if n_agreeing >= 3:
+                confidence_numeric = min(1.0, confidence_numeric + 0.08)
+            if n_agreeing >= 5:
+                confidence_numeric = min(1.0, confidence_numeric + 0.08)
+    elif n_total_active > 0:
+        # All agents near zero — no information
+        confidence_numeric = 0.12
     else:
         confidence_numeric = 0.1