Upload bayesian_engine.py with huggingface_hub

bayesian_engine.py

@@ -76,13 +76,26 @@ def sigmoid(x: float) -> float:
         return ez / (1.0 + ez)
 
 
-def compute_likelihood(score: float, confidence: float, reliability: float) -> Tuple[float, float]:
+VLM_AGENT_NAMES = {"Semantic Consistency Agent", "Text & Typography Agent"}
+
+def compute_likelihood(score: float, confidence: float, reliability: float,
+                       is_vlm_agent: bool = False) -> Tuple[float, float]:
     """
     Compute P(evidence | Fake) and P(evidence | Real) for one agent.
     
     From paper Eq. 1:
     P(e_i | Fake, r_i, c_i) = r_i · sigmoid(s_i · c_i) + (1 - r_i) · 0.5
+    
+    For VLM agents: applies additional confidence compression because
+    VLMs systematically overstate confidence (see review feedback).
     """
+    if is_vlm_agent:
+        # Temperature-scale VLM confidence to prevent Eq.1 corruption
+        # This compresses extreme VLM confidence values toward 0.5
+        if 0 < confidence < 1:
+            logit = np.log(confidence / (1 - confidence))
+            confidence = float(sigmoid(logit / 2.0))  # τ=2.0
+    
     l_fake = reliability * sigmoid(score * confidence * 5.0) + (1 - reliability) * 0.5
     l_real = reliability * sigmoid(-score * confidence * 5.0) + (1 - reliability) * 0.5
     return l_fake, l_real
@@ -213,6 +226,7 @@ def bayesian_synthesis(agent_results: List[AgentEvidence]) -> ForensicVerdict:
             evidence.violation_score,
             evidence.confidence,
             effective_reliability,
+            is_vlm_agent=evidence.agent_name in VLM_AGENT_NAMES,
         )
         
         likelihoods.append((l_fake, l_real))