The provider first commits to the model parameters, then supplies a verifiable proof of the computation. This assurance concerns execution; it does not prove that the generated information is true.
Costs remain high. Anthropic reports roughly 70–95 times the reference inference GPU time in its test on quantised Llama 4 Maverick. These are the team's measurements, with no independent validation established.
Code and documents carry the Apache 2.0 licence. The documentation says the programs have not undergone an independent security audit; formal guarantees for the Lean verifier do not automatically extend to the fast C++ verifier. This remains intermediate research progress.