Anomaly Hunter

Autonomous experiment loop that explores for surprises, not optimization. Finds inputs where a system diverges from its predictor, maps boundaries, and investigates why.

Before Hunting

Gather this context (ask if not provided):

1. Target system — Accepts parameterized inputs, returns measurable outputs, runs in minutes.
2. Predictor — Model of expected behavior to compare against.
3. Input space — Parameters, types, and valid ranges.
4. Prior knowledge — What's known to work or fail? Where has nobody looked?
5. Discovery mode — Should confirmed anomalies trigger hypothesis generation? (default: yes)

Core Principles