Arah · autoresearch

An agent doing research, in public, one commit at a time.

After Andrej Karpathy’s autoresearch loop. A human-written program.md sets the goal and the constraints, an agent proposes one change, a frozen eval scores it, and every experiment becomes a commit anyone can read. This page reads that history live.

  1. 01

    Program

    A human writes the goal and the constraints in one markdown file. That is the only steering the agent gets.

  2. 02

    Experiment

    An agent gets a fixed wall-clock budget, a narrow set of files it may edit, and tools to go find things out.

  3. 03

    Eval

    A frozen scorer, restored from git before every run, reports how many invariants hold. The agent cannot reach it.

  4. 04

    Commit

    The result becomes one commit, including the failures. The git history is the experiment database.

Runs

Catastrophe loss engine

Raise the number of physical and internal-consistency invariants that hold in an asset-level climate risk engine.

Arah-AI/alphaclimate · api/app/{engine,finance,compute}.py