Update Version_5/src/orchestrator.py

Version_5/src/orchestrator.py

@@ -1,38 +1,39 @@
-import os
-from groq import Groq
-from typing import Dict
-from .schemas import AgentReport, JudgeVerdict
-
-client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
-
-def evaluate_consensus(reports: Dict[str, AgentReport]) -> bool:
-    claims = [r.claim for r in reports.values()]
-    confidences = [r.confidence for r in reports.values()]
-    
-    all_same = len(set(claims)) == 1
-    high_conf = all(c >= 0.85 for c in confidences)
-    
-    return all_same and high_conf
-
-def run_judge(reports_summary: str, raw_data: str) -> JudgeVerdict:
-    system_prompt = (
-        "You are the presiding Judge Magistrate in an Enterprise Cybersecurity Matrix.\n"
-        "Your subordinate agents have provided split panel reports.\n"
-        "Respond strictly in valid JSON matching this schema:\n"
-        "{\"final_verdict\": \"CRITICAL_THREAT\"|\"SAFE\"|\"MANUAL_REVIEW_REQUIRED\", \"risk_score\": float, \"verdict_justification\": string}\n\n"
-        "DECISION ARCHITECTURE RULES:\n"
-        "1. If ANY analyst engine raises a 'PHISHING' claim with confidence >= 0.85, force 'CRITICAL_THREAT'.\n"
-        "2. Avoid label contradictions: Never set final_verdict to SAFE if your textual argument implies a threat pattern."
-    )
-    
-    # We leverage the powerful 70B cloud model for premium reasoning accuracy
-    response = client.chat.completions.create(
-        model="llama-3.3-70b-versatile",
-        messages=[
-            {"role": "system", "content": system_prompt},
-            {"role": "user", "content": f"RAW INPUTS:\n{raw_data}\n\nREPORTS:\n{reports_summary}"}
-        ],
-        response_format={"type": "json_object"},
-        temperature=0.0
-    )
+import os
+from groq import Groq
+from typing import Dict
+from .schemas import AgentReport, JudgeVerdict
+
+client = Groq(api_key=os.environ.get("GROQ_API_KEY"))
+
+def evaluate_consensus(reports: Dict[str, AgentReport]) -> bool:
+    claims = [r.claim for r in reports.values()]
+    confidences = [r.confidence for r in reports.values()]
+    
+    all_same = len(set(claims)) == 1
+    high_conf = all(c >= 0.85 for c in confidences)
+    
+    return all_same and high_conf
+
+def run_judge(reports_summary: str, raw_data: str) -> JudgeVerdict:
+    system_prompt = (
+        "You are the presiding Judge Magistrate in an Enterprise Cybersecurity Matrix.\n"
+        "Your subordinate agents have provided split panel reports.\n"
+        "Respond strictly in valid JSON matching this schema:\n"
+        "{\"final_verdict\": \"CRITICAL_THREAT\"|\"SAFE\"|\"MANUAL_REVIEW_REQUIRED\", \"risk_score\": float, \"verdict_justification\": string}\n\n"
+        "DECISION ARCHITECTURE RULES:\n"
+        "1. If ANY analyst engine raises a 'PHISHING' claim with confidence >= 0.85, force 'CRITICAL_THREAT'.\n"
+        "2. Avoid label contradictions: Never set final_verdict to SAFE if your textual argument implies a threat pattern.\n"
+        "3. SUPREME VETO RULE: If the Target URL belongs to a verified, universally trusted root domain (e.g., 'github.com', 'google.com', 'microsoft.com') and shows no signs of typosquatting, you MUST overrule paranoid sub-agents, ignore Rule 1, and force a 'SAFE' verdict with a low risk score. Standard authentication portals on official domains are inherently safe."
+    )
+    
+    # We leverage the powerful 70B cloud model for premium reasoning accuracy
+    response = client.chat.completions.create(
+        model="llama-3.3-70b-versatile",
+        messages=[
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": f"RAW INPUTS:\n{raw_data}\n\nREPORTS SUMMARY:\n{reports_summary}"}
+        ],
+        response_format={"type": "json_object"},
+        temperature=0.0
+    )
     return JudgeVerdict.model_validate_json(response.choices[0].message.content)