Frequently Asked Questions

No. You encrypt on your own machine with the open blind CLI. The server receives ciphertext and a public context — nothing else. It computes directly on the ciphertext and never decrypts. There is no column for a secret key on the server; it is structurally unable to store one.

Only Encrypted ciphertext, and the Public Crypto Context (public key and params, uploaded once). Raw data, encoded data, and your secret key are LOCAL ONLY — they never leave. Keygen, encoding, encryption, and decryption all run locally in the CLI.

Metadata. The threat model states this plainly: the server sees who submits, when, how many contributors join, file sizes, and which application runs. Contents stay encrypted; the surrounding facts do not. Encrypted is not the same as unregulated.

It binds the application, project, cohort commitment, input data, and result hashes, plus whether the minimum cohort size was met and how many runs occurred. Recompute those hashes offline and you can confirm that this exact application ran on this exact cohort and produced this exact result — without trusting us.

Re-execution proves determinism, not zero-knowledge: same ciphertexts in, same result digest out. It does not prove correct decryption without revealing the key. It does not certify that the underlying data is real, honest, or medically valid. The sentinel is a corruption check, not a message-authentication code. The single keyholder is trusted not to target individuals; certificates do not remove that trust — threshold FHE (v2) does.

Download the certificate from a public result page and run blind certificates verify <hash>. It recomputes every hash and checks consistency with no network. To go further, blind results verify <job> re-executes the pinned compute stage on the same ciphertexts and confirms a bit-identical result digest.

The blind CLI — key generation, encoding, encryption, decryption, and every verify command — is open source and auditable. The closed code is exclusively the hosting layer, which sits outside the trusted computing base by design: no security guarantee rests on code you cannot read.

No. A researcher can mint an accountless, short-lived bearer link (blind projects invite). A hospital contributor installs blind, encrypts locally, and uploads ciphertext — no account, no plaintext, ever.

Install with curl -fsSL https://blindmachine.org/install.sh | sh. Then: blind projects create (pin a signed application), blind data encrypt or blind contributions create (encrypt locally, upload ciphertext), blind projects freeze (commit the cohort once ≥20 contributions arrive), blind jobs create (compute on ciphertext), blind results decrypt (decrypt only the aggregate), blind certificates verify (re-check offline).

Adding, removing, or re-running over a cohort is the core differencing attack. No compute runs until the cohort is frozen and the minimum size (≥20 for the demo) is met, and runs are capped per project. This mitigates differencing; it does not fully solve it. Cross-project differencing remains a documented v1 limitation.

Core features are currently free while the service is in its early stage, and the open-source CLI is free. If paid plans arrive, the price will be shown before you pay. Verification cannot be paywalled even in principle — it runs offline, without our servers.

Email support@blindmachine.org. If a hash disagrees with what we claim, the hash wins — send us the mismatch.