Автономная оптимизация артефактов через дерево гипотез

★ 7.0 · ai-agents

arbor is a Claude Code skill that autonomously optimizes a concrete artifact — code, training recipe, agent harness, data pipeline, or prompt — against an objective and an evaluator using Hypothesis Tree Refinement (HTR) from the Arbor paper (Jin et al., 2026). Research state lives in a persistent hypothesis tree rather than conversation history: each node binds a hypothesis, its distilled insight, and a pointer to the artifact version in an isolated git worktree. Claude acts as the long-lived coordinator while short-lived subagent executors each test one hypothesis; a held-out merge gate admits a change only when it improves on a test evaluator never used during search, preventing overfitting to the dev signal. The skill fits tasks like "raise my model's eval score", "improve this agent or harness", "beat the baseline on a benchmark", or any long-horizon experiment-and-evaluate loop where accumulated knowledge matters more than a single fix. All bookkeeping — node creation, evidence writing, pruning, the merge gate, the Observe projection — is handled by scripts/tree.py; MIT license, authored by K-Dense Inc.