Example gallery¶
Two runnable, CI-tested, end-to-end examples that exercise every parallelism tier the engine ships. Each is built entirely on the public API — no engine edits — and doubles as the final API dogfooding pass before the 1.0 freeze.
Start with Which parallelism do I need?, then see the tier in action:
| Example | Domain | Tiers shown |
|---|---|---|
| Ride-hailing fleet | Autonomous robotaxi dispatch over a zone graph | Replications (5.2a), PDES sharding (5.2b) |
| LLM agentic workflow | Multi-agent task pipeline over a modelled LLM server | Replications (5.4a), strict-mode offload (5.4b) |
What each tier maps to¶
- Run one model many times (parameter sweeps, confidence intervals) →
Experiment/run_replications. Both examples' studies (fleet sizing, capacity) are this tier. - One run has too many events for one core →
ShardedSim. The zone-sharded ride-hailing variant is this tier. - One run is CPU-bound in computation, not events →
sim.offload. The strict-mode scoring showcase is this tier.
Guarantees every example holds¶
- Same-seed determinism — two runs with the same master seed produce identical KPIs.
- Backend reproducibility — study results are bit-identical across the thread, interpreter, and process backends for a fixed master seed.
- Trace-equivalence — the sharded run equals its sequential reference bitwise; the offloaded policy equals the inline reference bitwise (strict mode).
- Stdlib-only, no network — the examples add no runtime dependency, and the agentic example is proven by test to make no LLM or network call.
Regenerate the showcase curves with
python scripts/generate_example_charts.py.