Tycho

Tycho Brahe never knew the laws of planetary motion. What he did, for decades and by hand, was observe: he made the most precise records of the heavens anyone had seen. Kepler later derived the laws from those observations. The method was the insight: the laws of an unknown world are not handed to you; they are uncovered by watching carefully and refusing to look away.

Tycho is a self-directed agent harness for ARC-AGI-3, built around that same empiricist loop. A multimodal model enters an unfamiliar 64x64 world without rules or an objective. Tycho preserves what it sees, what it does, and what follows. When useful, the agent turns that evidence into a free-form executable hypothesis (State, transition, render, and outcome), checks the hypothesis against experience, and plans through it. When formalization is not useful, the agent remains free to reason directly. Observe, model, act, revise.

This repository contains the implementation and evaluation artifacts for “Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3.” It includes the agent, prompts, workspace API, planner and verifier, benchmark runner, Anthropic and OpenAI transports, paper configurations, replay viewer, tests, and compact scorecard evidence.

Official ARC-AGI-3 competition-mode scorecards on all 25 public games:

The scorecard manifests and aggregate paper metrics are in artifacts/.

Python 3.12 or newer is required.

python -m venv .venv
.venv/bin/python -m pip install --upgrade pip
.venv/bin/pip install -e .

Equivalently, make bootstrap uses python3; when that name resolves to an older system Python, select an installed interpreter explicitly, for example make bootstrap BOOTSTRAP_PYTHON=python3.12.

Agent-authored Python runs in a fresh, network-disabled container. Docker Engine, Docker Desktop, and Finch are supported through their command-line interfaces:

make sandbox-image
make sandbox-check

Runtime selection is automatic; on macOS Finch is preferred when available. Set TYCHO_SANDBOX_RUNTIME=docker or finch to select one explicitly.

Set one model provider:

export ANTHROPIC_API_KEY=...
# or: export OPENAI_API_KEY=...

Before using either key, run the credential-free validation suite:

make validate

This checks tests, configuration resolution, repository-integrity hashes, likely secret leakage, and the built wheel without calling Anthropic, OpenAI, or ARC. Provider and bounded game smoke tests are separate, explicit steps in docs/REPRODUCING.md. Benchmark execution also checks the container runtime before making a model call. The paper configurations are long-running, stochastic, and expensive; do not use them as smoke tests.

Inspect completed or in-progress runs with the local replay viewer:

tycho-viewer results --host 127.0.0.1 --port 8900

Open http://127.0.0.1:8900/ for frame-by-frame evidence, model calls, executable-model diagnostics, and workspace history. The /status page summarizes supervised long-running jobs.

• no_world_model: direct reasoning from typed evidence and durable notes.
• single: the actor may write and use world_model.py itself.
• orchestrator: the actor invokes a focused world-model builder when useful.
• trigger: the harness invokes the builder when verification falsifies or under-specifies a model.

The same observation, action, reset, animation, resume, and scoring paths are used in every policy. See docs/ARCHITECTURE.md for the component boundaries and executable-model interface.

Apache License 2.0. ARC-AGI-3 environments and engine packages retain their own licenses and terms.