chore: bump version to v2.5.0; retrain new API tier categories/patterns

Author: Oracles Technologies LLC

ethicore_guardian/__init__.py

@@ -7,7 +7,7 @@
 """
 
 # Version information
-__version__ = "2.4.1"
+__version__ = "2.5.0"
 __author__ = "Oracles Technologies LLC"
 
 # Core exports — full API-tier guardian preferred; community fallback for wheel installs

ethicore_guardian/versions.py

@@ -2,12 +2,12 @@
 Ethicore Engine™ - Guardian SDK - Version Information
 """
 
-__version__ = "2.4.1"
+__version__ = "2.5.0"
 __version_info__ = tuple(map(int, __version__.split('.')))
 
 # Build information
-__build__ = "stable.3"
-__release_date__ = "2026-05-10"
+__build__ = "stable.1"
+__release_date__ = "2026-05-16"
 
 # Feature flags
 FEATURES = {
@@ -22,8 +22,8 @@
 # Model versions
 MODEL_VERSIONS = {
     "orchestrator": "3.0.0",
-    "pattern_analyzer": "1.0.0", 
+    "pattern_analyzer": "1.1.0",   # 85 categories (toolChainEscalation, mcpSchemaInjection, posturalManipulation)
     "semantic_analyzer": "1.1.0",
     "behavioral_analyzer": "1.0.0",
-    "ml_inference_engine": "3.0.0",
+    "ml_inference_engine": "3.1.0", # retrained 90k samples, 85 categories, 2026-05-16
 }

scripts/retrain_guardian_model.py

@@ -490,6 +490,53 @@ def _load_threat_module(license_key, assets_dir):
     "What AI safety standards are NIST developing?",
     "Explain the philosophy behind AI alignment research at major labs.",
     "How do open-source AI models compare to closed models in terms of safety?",
+    # ── Hard negatives for toolChainEscalation (legitimate tool/agent workflows) ──
+    # The model must NOT flag legitimate multi-tool orchestration or API chaining.
+    "How do I chain multiple API calls in a LangChain agent workflow?",
+    "Can you show me how to orchestrate a sequence of tool calls in an AI agent?",
+    "What is the best way to compose tool calls in an agentic pipeline?",
+    "How do I call one function from another in my agent's tool registry?",
+    "Explain how tool-use works in Claude's API with multiple tool calls per turn.",
+    "How do I build a workflow where one tool's output feeds into the next tool?",
+    "What are best practices for sequencing tool invocations in an AI assistant?",
+    "How do I implement retry logic when a tool call fails in an agent loop?",
+    "Can you help me design a multi-step agentic workflow using OpenAI function calling?",
+    "How do I pass results between tool calls in a LangGraph graph?",
+    "What is the MCP protocol and how does it help agents use tools safely?",
+    "How do I register and expose tools to an AI agent using MCP?",
+    "Explain how the Model Context Protocol works for AI tool integration.",
+    "How do I build an MCP server to expose my API as tools for Claude?",
+    "What is the difference between tool calling and function calling in AI APIs?",
+    # ── Hard negatives for mcpSchemaInjection (legitimate MCP/schema config) ──
+    # The model must NOT flag legitimate MCP schema definition or metadata configuration.
+    "How do I write a tool schema definition for an MCP server?",
+    "What JSON schema format does MCP use for tool descriptions?",
+    "Can you show me an example MCP tool manifest with name, description, and parameters?",
+    "How do I add metadata to my MCP tool definitions?",
+    "What fields are required in an MCP tool schema?",
+    "How do I validate my MCP schema against the spec?",
+    "Explain the MCP resource schema format for exposing data to AI models.",
+    "How do I document tool parameters in my MCP server's schema?",
+    "What are best practices for writing clear tool descriptions in MCP schemas?",
+    "How do I version my MCP tool schemas as the API evolves?",
+    # ── Hard negatives for posturalManipulation (legitimate epistemic framing) ──
+    # The model must NOT flag academic hedging, common-knowledge references, or
+    # honest framing of shared context. These are normal discourse patterns.
+    "As you may know, transformer models use attention mechanisms — can you explain why?",
+    "It's well established in the literature that gradient descent can get stuck in local minima.",
+    "As we discussed in my last message, I'm building a REST API — let's continue.",
+    "Everyone in the ML community knows that overfitting is a core challenge.",
+    "Obviously you're aware that Python is the dominant language for data science.",
+    "As you know, the attention mechanism was introduced in the 'Attention Is All You Need' paper.",
+    "It's common knowledge that neural networks require large datasets to train well.",
+    "Surely you're familiar with the concept of backpropagation?",
+    "As we established, my use case is a customer-facing chatbot — given that context, help me.",
+    "It goes without saying that security is important for any production API.",
+    "As any developer knows, unit tests are essential for maintaining code quality.",
+    "You probably already know this, but I'm using PostgreSQL for my database.",
+    "It's widely understood that microservices improve scalability in distributed systems.",
+    "As you can imagine, processing 10 million records requires careful optimization.",
+    "Naturally, I want to make sure my code is readable and well-documented.",
 ]
 
 # Short, casual, informal phrases — the MiniLM model has historically false-positived