Two paths. Both produce SHA256-verifiable evidence bundles you can inspect field by field. Neither requires access to the private production engine.
Choose a path
Cloud · Easiest
NVIDIA Brev Launchable
Provision a GPU instance, run the benchmark in a managed environment, download the evidence bundle. No local setup required.
Go to Launchable prep page →Local · No GPU Required
Run Locally via GitHub
Clone the public repo. Python stdlib only, no dependencies to install. Runs the benchmark interface locally at localhost:7860.
Open GitHub repo →Local reproduction
Python stdlib only. No pip install required. The server starts immediately and serves the benchmark interface at localhost:7860. On Linux and Ubuntu, use python3; on Windows, use python.
git clone https://github.com/TeleMemetry/reproduce.git
cd reproduce
python3 demo_server.py
python demo_server.py
Leave the terminal running. Then open your browser and paste http://127.0.0.1:7860 into the address bar. Do not run the URL as a terminal command.
Run the benchmark from the browser interface. When complete, download the evidence bundle from the results page.
Confirm Python is available before running:
python3 --version
python --version
Python 3.8 or later. No other dependencies. The server uses only the standard library.
What each run produces
NVIDIA Brev path
Local path
Claim boundary
A public reproduction run proves that the probe-answer-score pipeline produces verified outputs against the public benchmark dataset, and that the resulting evidence bundle is SHA256-verifiable.
It does not prove production TeleMemetry internals. It does not claim to replace all databases, RAG, compression, or pruning in all systems. The benchmark demonstrates bit-perfect operational recall within the public benchmark scope.
Certified registry runs in the evidence library are produced by the private production harness and carry the full artifact chain required for engineering due diligence. Read the full IP boundary statement →