Parameter Golf

Parameter Golf is a source-first repository for building and measuring a compact, legality-aware Parameter Golf pipeline locally before spending target-hardware time on cluster calibration.

The repo is intentionally opinionated:

• one primary implementation path,
• explicit legality and byte-accounting checks,
• local evidence used for screening only,
• and promotion gates that refuse to treat workstation results as cluster truth.

What This Repo Includes

• shared implementation code in src/parameter_golf/
• operator scripts in scripts/
• layered workflow configs in configs/
• pytest coverage in tests/
• small fixture assets used by smoke tests
• data-acquisition and tokenizer-helper scripts under data/

What This Repo Does Not Include

The public source tree does not bundle:

• large raw corpora,
• derived tokenizer bundles from real selected-document data,
• generated run artifacts,
• checkpoints,
• handoff bundles,
• graphs, reports, or local environment dumps,
• or imported H100 calibration payloads.

Those surfaces remain local-only and are ignored by default.

Current Boundary

• A clean clone can run the fixture-backed smoke path and the source-first test suite.
• Serious local workflows require externally obtained dataset assets that are not committed here.
• Cluster-ready claims still require real target-hardware calibration and confirmation.
• The repo does not claim that current local results are H100-confirmed.

More detail:

• docs/local_vs_cluster_boundary.md
• docs/repro_commands.md
• data/README.md

Quick Start

Create an environment and install dependencies:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Run a small source-first verification pass that works from a clean clone:

PYTHONPATH=src python3 -m pytest \
  tests/unit/test_config.py \
  tests/smoke/test_runtime_paths.py \
  tests/smoke/test_smoke_pipeline.py -q

Run the fixture-backed smoke workflow:

PYTHONPATH=src python3 scripts/run_smoke_pipeline.py

Serious Local Workflows

Serious local workflows are intentionally asset-dependent. If you have the selected-document cache and want to prepare local screening assets:

PYTHONPATH=src python3 scripts/prepare_local_screening_assets.py

If you have the full selected-document source and want to prepare the full-corpus Scylla bundle:

PYTHONPATH=src python3 scripts/prepare_full_corpus_scylla_assets.py

If those assets are present locally, you can run the reduced-cost local baseline:

PYTHONPATH=src python3 scripts/run_local_baseline.py

Runtime outputs are written under artifacts/, runs/, reports/, and outputs/. Those directories are part of the local workflow and are intentionally excluded from the public source release.

License

Released under the MIT License. See LICENSE.