I think I have solved one of the most common problems in secondary markets with AI inference - how can the user know that the reseller did not log or change the prompt, while making sure the selected model is the one they agreed to and that the billing is correct?
Traditional solutions are very cumbersome. They rely on the ability of the client to audit a TEE enclave, meaning understanding the whole operating system - how the keys are generated, if the operator can't log to the server, etc. Most clients trust "any" attestation (it's attested, it must be right), some trust the fact that the same enclave runs for everyone (which is not nothing, this rules out targeted attacks, given someone audits the enclave), some people at least look if the enclave has published source code.
Very few (usually lim=>0) actually do the reproducible build and audit the code and whole operating system reselling the inference.
Routstr-confidential uses cryptography and PKI to achieve the desired outcome of private and secure inference, without needing to audit anything specific for the operator. It can be used for many other API call use cases.
@routstr itself is a permissionless way to find your inference provider on a secondary market. They usually have much lower prices than upstream. And now you can verify that you are not paying with your data, or you are not being cheated by paying for more expensive model than you get.
Feel free to use the protocol with any other reseller proxy though, it's all open-source.
Technical write-up, links to repos, ...:
Confidential inference on Routstr without trusted hardware – Juraj Bednar
Protocol explainer:
Seven magic tricks inside one AI request – Juraj Bednar
In-browser demo with real node and real backend:
Try to break it!
If you are not a programmer, the original explainer video is for you:
Ak nie ste progamátori, v pôvodnom blogu je stále videjko: 
