STAR Lab

Engineering evidence, reproducibility, and scalability records for STAR / StarBench.

STAR Lab is intentionally separate from the production/runtime repository.

• star-nexus/star is the product and benchmark repository. Users should normally use main there.
• star-nexus/star-lab is the engineering research archive: reproducible problem reports, controlled experiments, raw evidence, root-cause analyses, engineering decisions, and the historical STAR Performance Frontier.

The core rule is simple:

Every important engineering conclusion must be traceable from the claim, to raw evidence, to a reproducible experiment, to an exact star commit or immutable tag.

Start here

For the active 10K work, read the recovery entry. P0 archive and exact-source recovery preserve retired experiments.

1. Read PROTOCOL.md before adding a formal experiment.
2. Create new experiment packages from templates/.
3. Preserve only evidence that supports a durable engineering conclusion.
4. Bind every formal experiment to exact star-nexus/star commit SHA(s); never rely on a moving branch name alone.
5. Update records/performance-frontier.md only for validated frontier results.

Repository layout

star-lab/
├── README.md
├── PROTOCOL.md
├── templates/
│   ├── experiment-readme.md
│   ├── manifest.yaml
│   ├── analysis.md
│   └── decision.md
├── experiments/
├── investigations/
├── decisions/
└── records/
    └── performance-frontier.md

Version roles

star branch       = active work line
star checkpoint   = temporary comparison / debugging baseline
star annotated tag= validated immutable engineering milestone
star-lab          = reproducible evidence and decision history

STAR Lab does not maintain a production source fork. The P0 archival bundle preserves retired experiment refs for recovery. It has its own Git history and references STAR source states through exact commit SHAs and tags.