Genesis

Genesis is an autonomous AI research system that takes a project spec, decomposes the work, executes research iterations, verifies outputs, and synthesizes a paper-ready artifact set.

The repository currently contains two major lines of work:

• v0/ — earlier experimentation and prototype research infrastructure
• v1/ — the current Genesis v1 system, including the harness, specialist modules, storage layer, CLI, benchmarks, and paper synthesis pipeline

Current State

The active system is v1/.

Genesis v1 is designed around:

• a meta-harness that manages project runs
• specialist modules for adversarial checking, optimization, ideation, oracle generation, citations, plotting, verification, and paper synthesis
• persistent project state under per-project directories
• a local-first execution model that uses isolated runtime sandboxes rather than git worktrees for experiment execution

This repository has gone through several implementation passes, and the most complete delivery branch may be ahead of the local main checkout depending on your workspace state. If you are evaluating or extending v1, verify the branch you are on before starting work.

Repository Layout

• v0/ — legacy/prototype code and experiments
• v1/ — Genesis v1 package, configs, scripts, docs, and tests
• v1/genesis/ — core Python package
• v1/scripts/ — operational scripts such as taste initialization and manifold building
• v1/tests/ — unit, integration, and benchmark coverage

Working With v1

Start in v1/:

cd v1
python -m pip install -U pip setuptools wheel
python -m pip install -e ".[dev,data,verification,paper]"
python -m pytest -q
python -m genesis.cli.main --help

Additional optional dependency groups are available for heavier local runs:

• ml — baseline numeric dependencies used by the taste/optimizer path
• ml-accelerated — optional PyTorch and embedding backends for native environments that support recent wheels
• domain — astrophysics-oriented domain tooling

Example:

cd v1
python -m pip install -U pip setuptools wheel
python -m pip install -e ".[dev,data,verification,paper,ml,domain]"

For Apple Silicon Macs running an Intel (x86_64) conda env, the accelerated stack needs older compatible Hugging Face packages because current transformers 5.x expects newer PyTorch than those wheels provide. Genesis supports that env with:

python -m pip uninstall -y transformers sentence-transformers torch torch-geometric gpytorch
python -m pip install --upgrade --force-reinstall "numpy<2"
python -m pip install -e ".[dev,data,verification,paper,ml,ml-accelerated,domain]"

If you are on a native environment that supports modern PyTorch wheels and want the accelerated stack too:

python -m pip install --upgrade --force-reinstall "numpy<2"
python -m pip install -e ".[dev,data,verification,paper,ml,ml-accelerated,domain]"

CLI Overview

Genesis v1 currently exposes these operator-facing commands:

• genesis init
• genesis run
• genesis status
• genesis intervene
• genesis results
• genesis build-manifold
• genesis init-taste

These commands operate on project specs and project runtime directories under the configured output root.

Typical Local Flow

For a local-first run:

1. Install the required extras for the path you want to exercise.
2. Initialize the taste store with python3 scripts/init_taste.py.
3. Build or seed the manifold with python3 scripts/build_manifold.py --domain ....
4. Prepare a project spec JSON.
5. Run the system with python3 -m genesis.cli.main run --project-id ... --spec ....

Outputs are written per project, including:

• instructions and traces
• adversarial and verification reports
• generated code/output artifacts
• paper assets such as LaTeX, figures, and synthesis reports

Notes

• Runtime experiment execution is sandbox-based, not git-worktree-based.
• Local development may use git worktrees extensively for implementation lanes.
• Some heavier backends have optional support paths; install the relevant extras if you want to exercise them locally.

See Also

• v1/README.md
• v1/genesis_v1_spec.docx
• v1/genesis_v1_implementation_guide.docx