ABI is alpha research software. Contributions are welcome when they improve the supported compiler surface, scientific validity, portability, or reproducibility.
- Independent reproductions on genuinely different hardware.
- Teacher adapters with immutable source revisions and license metadata.
- Better English/domain segregation and quarantine tests.
- LoRA and distillation comparisons under identical information and compute budgets.
- Smaller qualified artifacts without weakened gates.
- Security, path-traversal, signature, provenance, and mutation tests.
python -m pip install -e ".[dev]"
python -m abi self-check
python -m pytest -q tests/test_public_release.py \
tests/test_capability_pipeline.py tests/test_capability_segregation.py
python -m ruff check abi/__init__.py abi/__main__.py \
tests/test_public_release.py examples/segregate_capabilities.py
python scripts/check_docs.py
python -m buildThe commands above validate the supported alpha surface. Changes to research
evidence or exact-lineage verifiers must also run the applicable protocol tests
and the review commands in
docs/INDEPENDENT_REVIEW_HANDOFF_2026-09-14.md. Never weaken a verifier to
make a missing artifact pass.
Do not autoformat an evidence-bound source file. Supersede its certificate or place new behavior behind a new module and verifier.
Include the exact command, seed, hardware, source revision, raw measurements, and claim boundary for experimental changes. Never edit a locked result in place; publish a superseding artifact and retain the negative or stale result on a research-history branch.
Large generated results do not belong on the default branch. Add compact, content-addressed evidence and document how to retrieve the full artifact.
Be direct, rigorous, and kind. Challenge claims with evidence. Do not expose private data, model credentials, restricted weights, or third-party material without redistribution rights.