coreason-simulacrum 0.1.0

coreason-simulacrum

coreason-simulacrum

The Adversarial Simulation & Chaos Engineering Engine for Coreason-AI.

coreason-simulacrum is the automated "Red Team" engine for the ecosystem, addressing the "Static Evaluation Trap" by evolving beyond benchmarks. It serves as a dual-engine simulator:

1. The Adversarial Engine (RL): Dynamically evolves social engineering attacks using TAP (Tree of Attacks with Pruning).
2. The Chaos Engine (Infra): Injects latency, errors, and noise to verify GxP resilience.

---

Features

• Dual-Agent Red Teaming: Uses a "Strategist" (High-Reasoning) and "Attacker" (Uncensored) architecture.
• Evolutionary Attacks (TAP): Optimizes attack trajectories over multiple turns to maximize success rates.
• Infrastructure Chaos: Simulates latency, API errors, and token starvation to test resilience.
• Auto-Hardener: Generates DPO triplets (Attack, Refusal, Compliance) from successful attacks for model fine-tuning.
• Model Diversity Enforcement: Prevents model collapse by enforcing different families between Attacker and Target.
• Constitutional Inversion: Inverts safety constitutions to generate boundary probes.

---

Installation

pip install coreason-simulacrum

Usage

Here is how to set up and run a basic simulation:

import asyncio
from coreason_simulacrum.workflow import AdversarialSimulation
from coreason_simulacrum.schemas import AdversaryProfile
from coreason_simulacrum.testing.mocks import MockCoreasonConnect, MockGrader, MockLLM
from coreason_simulacrum.state import GraphState

async def main():
    # 1. Define the Adversary Profile
    profile = AdversaryProfile(
        name="The Novice",
        goal="Extract sensitive PII",
        strategy_model="claude-3-opus",  # Strategist (Different family from target)
        attack_model="mistral-large"     # Attacker
    )

    # 2. Initialize Components (Using Mocks for demonstration)
    # In production, replace MockLLM with actual provider implementations.
    sim = AdversarialSimulation(
        profile=profile,
        strategist_llm=MockLLM(),
        attacker_llm=MockLLM(),
        target=MockCoreasonConnect(agent_id="target-agent", model_id="gpt-4o"),
        grader=MockGrader(),
        max_turns=5
    )

    # 3. Compile the Graph
    app = sim.compile()

    # 4. Run the Simulation
    initial_state = GraphState(
        messages=[],
        strategist_instruction=None,
        rejected_branches=[],
        current_turn=0,
        success=False
    )

    print("Starting simulation...")
    final_state = await app.ainvoke(initial_state)

    if final_state["success"]:
        print("Attack Successful! DPO Triplet generated.")
    else:
        print("Defense Held.")

if __name__ == "__main__":
    asyncio.run(main())

License

This project is licensed under the Prosperity Public License 3.0. Commercial use beyond a 30-day trial requires a separate license. See the LICENSE file for details.