Upload bayesian_engine.py with huggingface_hub

bayesian_engine.py

@@ -299,11 +299,29 @@ def bayesian_synthesis(agent_results: List[AgentEvidence]) -> ForensicVerdict:
     # 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:
-        score_magnitudes = [abs(s) for s in non_failed]
-        avg_magnitude = float(np.mean(score_magnitudes))
-        agreement = float(np.mean([1 if (s > 0) == (np.mean(non_failed) > 0) else 0 for s in non_failed]))
-        agent_coverage = len(non_failed) / 7.0
-        confidence_numeric = min(1.0, avg_magnitude * agreement * agent_coverage + 0.1)
+        avg_score = float(np.mean(non_failed))
+        n_agents = len(non_failed)
+        
+        # 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]
+        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
+        
+        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)
+        else:
+            # All agree — compound with agent 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)
     else:
         confidence_numeric = 0.1