Trust Model

What PRAMPTA proves, what it does not, and how to tell the difference. The most dangerous thing a trust layer can do is promise more than it can prove — so the limits are on this page too, at the same size as the guarantees.

• ✓Identity binding. A licence is a real Ed25519 signature by the subject's key, countersigned by us. Verification reconstructs the exact signed bytes, so altered terms fail rather than pass quietly.
• ✓Signed, one-time decisions. Every authorization decision is signed and carries a one-time nonce, an expiry, the key id, and the policy version that produced it. It cannot be forged or replayed for a second generation.
• ✓Three honest outcomes. A decision is allow, deny, or review. Review means held for a human — neither granted nor refused. Software that only understands yes/no reads it as not-allowed, which is the safe reading.
• ✓Fail-closed enforcement. If a licence restricts a dimension and the request omits it, we refuse rather than allow. Missing context is treated as a possible evasion, not an oversight.
• ✓Self-attestation buys nothing commercial. A subject registered on nothing but its own claim can issue personal-use licences only. Commercial licensing is a state an operator grants — never one the registrant reaches alone.
• ✓Training is opt-in, never implied. A licence grants specific rights. Model training, fine-tuning and agent memory must be named explicitly; a licence that is silent about training does not permit it.
• ✓The rights holder's veto wins. You can block a specific AI provider outright. That decision overrides any licence, is checked before licences are even considered, and no configuration can switch it off.
• ✓Protections for high-risk subjects. A subject marked as a minor, deceased, vulnerable or a sensitive public figure gets no auto-approval, no shortcut to commercial use, and — for a minor — categories that are refused whatever a licence claims.
• ✓Revocation with a timeline. A licence can be withdrawn in whole or in part, effective now or on a future date, with a reason, as a signed append-only record. Giving notice is possible; the signed terms are never edited.
• ✓A tamper-evident audit trail. Every consequential action is written to a hash-linked, signed log carrying pseudonymous ids only. It can be verified independently of us.
• ✓Evidence is encrypted and access-logged. Identity documents are encrypted under a secret separate from our signing keys, every read is recorded, and the person whose documents they are can see who opened them.
• ✓Providers are gated and suspendable. A provider cannot act without a credential we issued, sandbox and production are separate, and a misbehaving provider can be suspended without a manual database edit.

• –We cannot physically stop a disobedient system. PRAMPTA authorizes; the AI provider's software enforces. At the model itself we are advisory. A provider that ignores a refusal has broken a contract and the audit trail proves it — but the content was still produced.
• –We cannot tell whether a self-made claim is true. A checkbox is a statement someone makes, not proof they hold the rights. Until evidence is reviewed, a subject is self-attested and treated as such.
• –We cannot settle a legal dispute. We can freeze a contested subject, carry evidence between the parties, and record who decided what and why. Who actually owns a right is a court's question, not ours.
• –We cannot detect a high-risk subject automatically. Whether someone is a minor is declared, by an honest registrant or by an operator — never guessed from a name. An undeclared subject gets no automatic protection, which is why human review matters.
• –Some gates see only what is declared. Where a request must state its categories or the rights it exercises, a provider that omits them is not caught by those particular checks. They are a layer over the coarser protections, not the whole of them.
• –We cannot reach into a provider's storage. A revocation or expiry stops future generation. What happens to content already produced is a recorded contractual term — we cannot delete a frame someone already shipped.
• –We do not verify every right on earth. Similar names are surfaced for a human, not blocked — two real people share a name. Our job is to state the trust level honestly and gate the dangerous actions, not to be an oracle of ownership.

If anything in our product, docs or marketing ever sounds stronger than this page, this page is right and that is wrong — please tell us at legal@prampta.com.