Accueil / Méthode benchmarkology
Méthode benchmarkology
Performance numbers must be verifiable, scoped and limited.
This page documents how BOUNDA reports performance so internal benchmarks do not become vague marketing claims.
Machine profile
Report CPU, RAM, OS, storage, Python/runtime version and engine version for every published run.
Dataset profile
Report event count, receipt types, payload shape, synthetic vs client, warm/cold split and compression ratio.
Measurement protocol
Report p50, p95, p99, throughput, memory ceiling, run count, error bars and exact command line.
Public benchmark policy
What BOUNDA can safely say.
| Claim type | Allowed only when | Label |
|---|---|---|
| Throughput | Machine, dataset, engine version and command are disclosed. | Internal benchmark |
| Latency | p50/p95/p99 and warm/cold conditions are visible. | Measured under stated conditions |
| Client result | Client approved or anonymized methodology is documented. | Client-validated or anonymized |
| Financial value | Client-provided costs or reconciliation prove it. | MEASURED / CLIENT_VALIDATED |
| Scenario value | Assumption is not proof. | SCENARIO — NOT MEASURED |
Reproducibility checklist
Minimum preuve for a benchmark page.
hardwareOSengine versiondataset schemaevent countreceipt mixcommandrun countp50 / p95 / p99memorylimitsdate