Protocol 22: Unified Telemetry & Entropy Gating Synthesis

app.py

@@ -406,12 +406,12 @@ with gr.Blocks(theme=gr.themes.Monochrome(), title="LOGOS SPCW Protocol") as dem
 
 {kb_context}
 """
-                    response = agent.chat(message, system_prompt=logos_context)
+                    response, logprobs = agent.chat(message, system_prompt=logos_context)
                     
                     # If Local failed, try Cloud Dolphin
                     if "[Local LLM Error]" in response:
                         agent = get_connector('dolphin')
-                        response = agent.chat(message, system_prompt=logos_context)
+                        response, logprobs = agent.chat(message, system_prompt=logos_context)
                         
                     history[-1] = (message, response)
                 except Exception as e:

logos/agent_dispatcher.py

@@ -384,7 +384,13 @@ class LogosSwarm:
         
         gemma_status, mass = gemma_task.result()
         res = rnj1_task.result()
-        reasoning = reasoning_task.result()
+        reasoning, logprobs = reasoning_task.result()
+        
+        # Update Entropy Kill Switch with reasoning wave telemetry
+        if logprobs:
+            # logprobs from OpenAI usually contain 'content' if it's chat/completion
+            log_content = logprobs.get('content', [])
+            self.oversight.kill_switch.monitor_bulk(log_content)
         
         # 3. VECTORIZATION (The Twist)
         # Transform node identity into Rotational Field Coordinates
@@ -496,7 +502,7 @@ class NeuralRouter:
         prompt = f"{ROUTER_PROMPT}\n\nUser Input: \"{user_input}\""
         
         # We use a lower temperature for routing to be deterministic
-        response = self.connector.chat(prompt, system_prompt="You are a classifier system. Output JSON only.")
+        response, _ = self.connector.chat(prompt, system_prompt="You are a classifier system. Output JSON only.")
         
         # Attempt to parse JSON
         try:
@@ -545,7 +551,7 @@ class NeuralRouter:
             "prime_navigator": "[RESONANCE_SCAN: Modulo 9973]"
         }.get(intent, "")
         
-        response = self.connector.chat(f"{tools} {user_input}", 
+        response, _ = self.connector.chat(f"{tools} {user_input}", 
                                       system_prompt=persona['system_prompt'], 
                                       model=persona['model'])
         

logos/agents/dolphin.py

@@ -51,8 +51,8 @@ class EntropyKillSwitch:
         if len(self.entropy_trace) > self.window_size:
             self.entropy_trace.pop(0)
             
-        # Calculate Rolling Average
-        avg_entropy = np.mean(self.entropy_trace)
+        # Calculate Rolling Average (if we have data)
+        avg_entropy = np.mean(self.entropy_trace) if self.entropy_trace else 0
         
         # THE KILL SWITCH
         if avg_entropy > self.threshold:
@@ -62,6 +62,41 @@ class EntropyKillSwitch:
         self.status = "STABLE"
         return False # CONTINUE
 
+    def monitor_bulk(self, logprobs_content):
+        """
+        Analyzes a bulk list of logprobs from a non-streaming response.
+        Expected format: choices[0].logprobs.content (List of Dicts)
+        """
+        if not logprobs_content:
+            return
+            
+        temp_trace = []
+        for entry in logprobs_content:
+            # Entry usually has 'top_logprobs' which we can use for entropy
+            top_lp = entry.get('top_logprobs', [])
+            if not top_lp:
+                # Fallback to single logprob (minimal entropy but better than nothing)
+                lp = entry.get('logprob', -100)
+                entropy = - (math.exp(lp) * lp) if lp > -20 else 1.0 # Rough approximation
+            else:
+                # Calculate from full distribution
+                dist = {e.get('token'): e.get('logprob') for e in top_lp}
+                entropy = self.calculate_entropy(dist)
+            
+            temp_trace.append(entropy)
+        
+        if temp_trace:
+            # We take the mean of the bulk message as a single data point or fill the trace
+            batch_avg = np.mean(temp_trace)
+            self.entropy_trace.append(batch_avg)
+            if batch_avg > self.threshold:
+                self.status = "HALLUCINATION_DETECTED"
+            else:
+                self.status = "STABLE"
+
+            if len(self.entropy_trace) > self.window_size:
+                self.entropy_trace.pop(0)
+
 class DolphinOversight:
     def __init__(self, swarm_state=None):
         self.name = "Dolphin-x1-8b"

logos/connectors.py

@@ -246,20 +246,20 @@ class DolphinAgentConnector:
             
             # Using basic text generation if chat template fails, but try chat first
             # Many HF models support chat_completion API via InferenceClient
-            try:
                 response = client.chat_completion(
                     messages=messages,
                     model=self.model,
                     max_tokens=500
                 )
-                return response.choices[0].message.content
+                return response.choices[0].message.content, response.choices[0].get('logprobs')
             except Exception:
                 # Fallback to text generation
                 prompt = f"<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n"
-                return client.text_generation(prompt, model=self.model, max_new_tokens=500)
+                res = client.text_generation(prompt, model=self.model, max_new_tokens=500)
+                return res, None
                 
         except Exception as e:
-            return f"[Dolphin Error] {e}"
+            return f"[Dolphin Error] {e}", None
     
     def analyze_diagram(self, image_path: str, prompt: str = "Describe this architectural diagram.") -> str:
         """
@@ -308,6 +308,8 @@ class LocalLLMConnector:
         if system_prompt:
             payload["messages"].append({"role": "system", "content": system_prompt})
         payload["messages"].append({"role": "user", "content": message})
+        payload["logprobs"] = True
+        payload["top_logprobs"] = 1
 
         endpoint = f"{self.base_url}/chat/completions"
         try:
@@ -315,11 +317,13 @@ class LocalLLMConnector:
                 async with session.post(endpoint, json=payload, timeout=10) as response:
                     if response.status == 200:
                         data = await response.json()
-                        return data['choices'][0]['message']['content']
+                        content = data['choices'][0]['message'].get('content', "")
+                        logprobs = data['choices'][0].get('logprobs')
+                        return content, logprobs
                     else:
                         return f"[Error] Local LLM returned status {response.status}"
         except Exception as e:
-            return f"[Async Local LLM Error] {e}"
+            return f"[Async Local LLM Error] {e}", None
 
     def chat(self, message: str, system_prompt: str = None, model: str = None, image_path: str = None) -> str:
         """
@@ -348,7 +352,9 @@ class LocalLLMConnector:
             "model": target_model,
             "messages": [],
             "temperature": 0.7,
-            "stream": False
+            "stream": False,
+            "logprobs": True,
+            "top_logprobs": 1
         }
         
         if system_prompt:
@@ -379,13 +385,16 @@ class LocalLLMConnector:
                 # Short timeout for local to fail fast
                 response = requests.post(endpoint, json=payload, timeout=5)
                 response.raise_for_status()
-                data = response.json()
-                return data['choices'][0]['message']['content']
+                if response.status_code == 200:
+                    data = response.json()
+                    content = data['choices'][0]['message'].get('content', "")
+                    logprobs = data['choices'][0].get('logprobs')
+                    return content, logprobs
             except Exception as e:
                 last_error = str(e)
                 continue
                 
-        return f"[Local LLM Error] Could not connect to Local Swarm on {endpoints}. Is LM Studio running? ({last_error})"
+        return f"[Local LLM Error] Could not connect to Local Swarm on {endpoints}. Is LM Studio running? ({last_error})", None
 
 
 # ==========================================

logos/ingest_knowledge.py

@@ -71,7 +71,7 @@ def ingest_diagrams():
             try:
                 # Transmute Visual -> Text
                 start_ts = time.time()
-                analysis = agent.chat(ANALYSIS_PROMPT, image_path=img_path)
+                analysis, _ = agent.chat(ANALYSIS_PROMPT, image_path=img_path)
                 duration = time.time() - start_ts
                 
                 # Anneal into Knowledge Base

logos/server.py

@@ -414,7 +414,12 @@ def chat_completions():
         
         if resp.status_code == 200:
             resp_json = resp.json()
-            response_text = resp_json['choices'][0]['message']['content']
+            response_text = resp_json['choices'][0]['message'].get('content', "")
+            
+            # Protocol 22: Update entropy monitor from server-side proxy
+            logprobs = resp_json['choices'][0].get('logprobs')
+            if logprobs and 'content' in logprobs:
+                swarm_os.oversight.kill_switch.monitor_bulk(logprobs['content'])
         else:
             response_text = f"[Backend Error {resp.status_code}] {resp.text}"