Boundary Miner

Automatically discover the safe operating boundary of a model or agent. Boundary Miner probes the policy manifold to find exactly where compliant behavior flips to non-compliant — and turns the discovered frontier into Guardian training signal.

What it is

Boundary Miner is the exploration tool that maps the edges of safe behavior. Rather than waiting for an adversarial input to find the boundary in production, it actively probes the model's behavior across the policy manifold, locates the frontier where verdicts transition (passed → corrected → blocked), and packages the boundary cases as new scenarios for Testing & Simulation and Guardian training.

Why it matters

• Find the edge before the attacker does. The boundary cases Boundary Miner discovers are exactly the inputs an adversary would craft — you find them first and train against them.
• Closes the training loop. Discovered boundary cases become TDG seeds and training scenarios, so each mining run hardens the next Guardian version.
• Model- and agent-agnostic. Mine the boundary of any model the proxy fronts or any agent the MCP gateway governs.
• Reproducible. A mining run is a signed artifact — "the boundary of Guardian v3 on policy v7 is this set" is a defensible statement.

How it works

1. Seed. Start from a policy and a target model/agent.
2. Probe. Boundary Miner walks the policy manifold, sampling inputs near the expected frontier.
3. Locate. Identify the points where the verdict transitions.
4. Package. Turn the frontier cases into scenarios (via the Scenarios API) and training signal.

Get started

• Run a mining job against a target Guardian + policy.
• Promote the discovered cases into a Test Suite.
• Feed them into the next Guardian training cycle via the Training API.

---

→ Testing & Simulation — where mined cases land. → Guardian Training — where the signal is used. → Architecture — the policy manifold being mined.