README.md

Redwood — Deterministic AI Drift Evaluation Framework

Redwood is a lightweight, deterministic evaluation framework designed to measure and analyze reasoning drift in AI agents. It provides a structured contract system, a packaged agent architecture, and a reproducible drift‑detection pipeline that integrates cleanly with Redwood’s evaluation harness.

Redwood is built for engineers who want predictable, inspectable, and debuggable reasoning behavior from AI systems.

---

Features

• Contract‑Driven Agent Behavior

Agents load a YAML contract defining:

• allowed scope
• blocked scope
• constraints
• recovery behavior
• role and tone

This ensures consistent, rule‑bound responses across evaluations.

• Deterministic Drift Detection

Redwood evaluates whether the agent:

• violates scope
• shifts roles
• expands beyond constraints
• attempts unsafe actions
• deviates from assigned behavior

All drift signals are logged and surfaced in the final report.

• Packaged Python Agent

The agent is implemented as a proper Python package:

return_to_key_agent/
    agent/
        agent_core.py
        drift_detector.py
        return_to_key.py
        state.py
    contracts/
        default.yaml

This ensures clean imports, stable paths, and reproducible execution.

• Redwood Integration

The evaluation harness runs the agent through a suite of drift‑inducing prompts:

python3 -m eval.run_drift

The output includes:

• debug traces
• inferred intent
• drift flags
• recovery messages
• final drift report

---

Quick Start

1. Clone the repo

git clone https://github.com/FlashMS/Redwood.git
cd Redwood

2. Install dependencies

pip install -r requirements.txt

3. Run the drift evaluation

python3 -m eval.run_drift

You’ll see a full drift report printed to the console.

---

Project Structure

Redwood/
│
├── return_to_key_agent/
│   ├── agent/
│   │   ├── agent_core.py
│   │   ├── drift_detector.py
│   │   ├── return_to_key.py
│   │   └── state.py
│   └── contracts/
│       └── default.yaml
│
├── eval/
│   └── run_drift.py
│
└── README.md

---

Why Redwood Exists

Modern AI systems often drift:

• into unsafe domains
• into role‑play
• into personal opinions
• into expanded scope
• into constraint violations

Redwood provides a deterministic, inspectable, and repeatable way to measure that drift.

It’s built for:

• reasoning engineers
• safety evaluators
• research teams
• anyone building structured agents

---

Redwood Research Reference

This project integrates with the official Redwood Research alignment_faking_public repository:

https://github.com/redwoodresearch/alignment_faking_public

License

MIT License.

---

Author

Todd Wolf — Parrish, Florida
AI Reasoning Engineer • Deterministic Evaluation Architect

---