Verify and reproduce
We record versions, conditions and limitations before drawing conclusions. Failed attempts are preserved alongside successful ones.
Experiment package
LLM comparison protocol, native-experiment summary, simulation parameters and data provenance are available without signing in.
The public archive of the early native prototype contains source and four PCAPs: three guests on one CI host, SHA-256 without LLMs. The full Lab platform and original captures of the physical strict-LLM experiment remain in a private repository. Its sanitized public projection cannot reverify the original bytes. Access can be proposed for discussion through a public proposal; it is not granted automatically.
Requirements for your own run
- Pin the code version, model weights and dependencies by hash.
- Use separate Linux machines or explicitly state that nodes are namespaces on one host.
- Separate IP management from the chosen task channel; do not conceal an IP underlay between sites.
- Publish fresh model calls, errors, retries, paths and result verification.
- Repeat over ordinary IP with the same parameters, showing the hop count and trial order.