The NVIDIA Brev Launchable is a public harness that mathematically verifies the measurement workflow and artifact trail. It does not expose or run the private TeleMemetry production engine.
The NVIDIA Brev Launchable is a self-contained reproduction environment. It runs on a GPU instance you provision, against public telemetry data, using the documented scoring methodology. The output is a SHA256-verifiable artifact that matches the format of the certified benchmark registry records.
Its purpose is mathematical verification of the measurement workflow: that the probe-answer-score pipeline produces the reported numbers when run against the documented inputs, and that the resulting artifact is tamper-evident.
It is not a stripped-down version of the production engine. It is a purpose-built public harness designed to verify the benchmark measurement process independently.
What is public
What remains private
The certified benchmark runs in the evidence registry were produced by the private production harness on RunPod GPU instances, against the full nuPlan Mini dataset, with full artifact chains including GPU telemetry CSV and Integrity Root. Those runs are the primary evidence.
The public Launchable runs against a 200-record public nuPlan subset. The expected result is 200/200 probes passed. This confirms the scoring methodology is reproducible, not that the public harness is equivalent to the production run environment.
Both produce SHA256-verifiable artifacts. Only certified registry runs carry the full artifact chain required for engineering due diligence.
Frequently asked questions
No. The Launchable runs a real scoring harness against real public nuPlan telemetry on the GPU instance you provision. It produces a real SHA256 artifact that you can inspect field by field. The purpose is to let any engineer verify that the scoring methodology produces the reported numbers, without needing access to the private production engine.
The production engine is the core IP. It includes proprietary state-routing, indexing architecture, and domain adapters that are the subject of pending patent filings. Publishing the full engine would eliminate IP protection before those filings complete. The public harness is designed to give engineers everything they need to verify the claims without requiring access to the protected internals.
Using the public Launchable: that the probe-answer-score pipeline produces 200/200 on the public telemetry subset, that the SHA256 artifact matches the expected hash, that the prompt template and scoring criteria are as documented, that the tokens-per-probe figure is consistent with the methodology, and that the artifact format matches certified registry records.
Using the certified registry artifacts directly: SHA256 hash of each artifact, per-probe scoring logs, GPU power CSV, runtime and RPS figures, and Integrity Root where published.
Access to the production engine image, the full nuPlan Mini dataset configuration, the private benchmark harness, and the domain adapter architecture requires an NDA. This applies to independent audit reviewers, integration partners, and enterprise evaluation engagements.
Request NDA review access →