# Benchmarks

Clean product one-liner: control benchmarks live here as first-class product proof.

Why: memory, tensor, safety, economics, and routing quality need a local test track, not ad hoc side runs.

## Local standalone benchmark

```bash
./benchmarks/run_standalone_control_bench.sh
```

Why: this proves the repo can benchmark its own graph-first and execution lanes without leaning on an external engine.

## Golden XMLC seed

```bash
python3 benchmarks/build_golden_xmlc_seed.py
```

Why: this materializes a weakly supervised XMLC seed set over the corpus using a narrow set of golden emergent-practice tags such as `loops_create_loops`, `kernel_abstractions`, and `reducible_map_engine`.