Roshan10000/MFDE

🧠 MFDE: Misleading Feedback Decision Environment

  

MFDE is a professional-grade Email Triage Simulation designed to evaluate and train AI agents in handling high-stakes decisions under misleading feedback. Now updated to v2.0 with real Gmail integration and calibrated deterministic scoring.

🌍 Real-World Utility (30% Weighting)

In modern Security Operations (SecOps), feedback loops are often corrupted by noisy signals, look-alike domains, and biased human review. MFDE models this "Reality Gap" by forcing agents to choose between:

1. 1.Trusting the Evidence: Analyzing the raw email content for deceptive patterns.
2. 2.Trusting the Reward: Following potentially corrupted feedback signals that may penalize correct security escalations.

This environment is immediately valuable for the RL community to test agentic resilience and calibration in enterprise-critical workflows.

👁️ v2.0 Extended Features

• —Real Gmail connectivity — Connect any inbox via MCP and triage live telemetry.
• —Claude-Powered Inference — Multi-modal baseline agents for high-fidelity triage.
• —Calibrated Scoring — Strictly deterministic rewards in the (0.02, 0.98) safe range.
• —Glassmorphic Dashboard — Professional real-time analytics for human-in-the-loop monitoring.

👁️ Technical Specifications

Observation Space

The environment provides structured email metadata: | Field | Type | Description | |---|---|---| | email_text | string | The full raw body of the email | | sender | string | The sender address (critical for phishing detection) | | subject | string | The subject line | | step_count | int | Progress within the current task |

Action Space (JSON)

The agent must submit a precise triage decision:

{
  "decision": "reply | ignore | escalate",
  "priority": "low | medium | high"
}

🎯 Task Tiers & Difficulty

🧮 Reward & Grader Design

Calibrated Tiers

MFDE provides dense signal with strict adherence to OpenEnv range compliance:

• —Full Success (Decision & Priority correct): 0.98
• —Partial Success (Decision correct only): 0.55
• —Failure (Incorrect decision): 0.02

Medium and Hard tasks inject reward noise based on the task difficulty, simulating real-world signal corruption while staying within the (0.01, 0.99) safe interval.

Deterministic Grader

Compliance with the hackathon rubric is guaranteed via grader.py:

• —Function: grade(history) -> float [0.0 - 1.0]
• —Calculated by averaging total weighted rewards across the trajectory and clamping to 2 decimal places for precision.

🚀 Getting Started

Prerequisites

• —Python 3.11+
• —Docker (for containerized execution)

Local Launch

pip install -r requirements.txt
export ANTHROPIC_API_KEY="your-key-here"
python -m uvicorn app:app --host 0.0.0.0 --port 7860

Baseline Inference (Reproduction)

Our inference.py uses the standard OpenAI client and supports heuristic fallbacks:

# Set credentials for Claude AI inference
export HF_TOKEN="your-hf-token"
export API_BASE_URL="https://router.huggingface.co/v1"
python inference.py

🔌 API Endpoints

🏗️ Spec Compliance Checklist

• —[x] Typed Models: Full Pydantic V2 implementation in models.py.
• —[x] Standard API: /reset, /step, and /state endpoints functional.
• —[x] Containerized: Dockerfile tested for HF Spaces UID 1000.
• —[x] Calibrated: Rewards strictly within (0.01, 0.99).
• —[x] Structured Logs: Baseline script outputs exact [START]/[STEP]/[END] format.
• —[x] Deterministic: All graders are 100% reproducible.

📊 Baseline Scores (Reproduce with random.seed(42))

Reproduced using deterministic heuristic fallback.

🏗️ File Structure

mfde-main/
├── app.py            # FastAPI server (standard + Gmail endpoints)
├── env.py            # MFDEEnv with Gmail support
├── models.py         # Pydantic models (incl. Gmail models)
├── tasks.py          # Simulation task definitions
├── grader.py         # Grading logic (simulation + Gmail)
├── inference.py      # AI inference script (simulation + Gmail)
├── dashboard.html    # Professional Glassmorphic UI
├── openenv.yaml      # OpenEnv spec
├── Dockerfile        # HF Spaces compatible (UID 1000)
└── requirements.txt  # Dependencies