# SPS baseline (anvil, 2026-06-30)

Harness: 4 envs x 64 ticks x 2 rounds = 512 sim-steps per run.

| Path | SPS (approx) | Notes |
|------|----------------|-------|
| CPU scalar (`build/bin/cpu/sps_benchmark`) | ~5,600 | tick_compose_full chain on CPU |
| CUDA serial (`build/bin/cuda/sps_benchmark`) | run overnight script for number | one env per launch in current harness |

Oracle: stdout hash lines CPU==CUDA PASS (deterministic; timing not compared).

Re-run:
```bash
make cpu-sps_benchmark && build/bin/cpu/sps_benchmark
python3 oracle/runner.py sps_benchmark
```
