Rivellum benchmark documentation now treats validity gates as first-class. The project still uses rivellum-bench, but sustained-load claims must be tied to benchmark mode, topology, storage mode, and worker validation.

Current Benchmark Model

sustained-load defaults to continuous mode:

1. start generators
2. run warmup as unmeasured lead-in
3. capture baselines while load is still flowing
4. measure without restarting generators or resetting nonces
5. stop submission after the measured window
6. drain accepted work for accounting

--premeasure-drain is still available, but it is debug-only isolated_premeasure_drain mode and must not be presented as sustained throughput.

Valid Throughput Baselines

Current valid single-lane role-distributed external-worker baselines are drawn from clean continuous runs with leader-authoritative startup nonce fetch and no nonce rejects:

These are role-distributed runs from the current external-worker path with nt_rejected_stale=0 and worker activity proven in metrics.

Diagnostic Results

Some recent runs are diagnostic only and must not be headlined as protocol throughput:

• PoUW outbox disabled mode is an A/B control to isolate hot-path disk cost.
• isolated premeasure drain mode is a debug mode, not a sustained benchmark.
• runs with stale nonce rejects, worker failures, dispatch drops, ENOSPC, or compromised infrastructure are invalid as baselines.

Required Validity Gates

A run is only a sustained throughput baseline if all of the following hold:

• benchmark_mode = continuous
• actual offered TPS is reported, not just target rate
• topology is declared
• storage mode is declared
• worker_jobs_received_total > 0 and worker_jobs_completed_total > 0 in external-worker mode
• worker_jobs_failed_total = 0
• leader_worker_results_received_total catches up
• nonce_rejected_total = 0
• nt_rejected_stale = 0
• dispatch_dropped_total = 0
• batches_failed = 0
• rejection receipts are not counted as successes
• no ENOSPC, active swap, malware, or public-service compromise

Key Flags And Fields

Useful sustained-load fields and options include:

Interpreting Current Results

• The current clean external-worker single-lane path is in the low-to-mid 4k committed TPS range on the documented role-distributed setup.
• Disabled PoUW outbox mode can improve throughput materially, but it is not a canonical protocol result.
• The next identified hot-path target after PoUW isolation is per-intent state-root work inside remote worker dispatch.