arbor

What problem does it solve?

It coordinates long-horizon improvement of a concrete artifact against a measurable objective, so you can run many experiments without losing track of what was learned.

Core Features & Use Cases

• Hypothesis tree planning: organize competing directions, refinements, and negative constraints.
• Isolated executor runs: test one hypothesis at a time in separate worktrees to avoid cross-contamination.
• Dev/test gating: use a fast development evaluator during search and admit changes only after held-out confirmation.
• Use cases: model training recipes, agent harness tuning, data pipelines, benchmark submissions, and prompt optimization.

Quick Start

Ask Arbor to improve your current artifact against a named dev evaluator and held-out test evaluator, then let it run iterative experiments, record evidence, and merge only the changes that generalize.