A sealed source is a record issued by its source and protected so that any change can be detected, such as a certificate or statement carrying a seal that can be checked against the issuer. It is the strongest provenance class. An AI agent relying on a sealed fact can confirm both who issued it and that nobody has altered it since.
Why does an agent need sealed sources?
Because documents are easy to forge. A lender asks for a payslip, an insurer for a repair invoice. Each arrives as a file, and a well-made fake looks the same to a model as the real thing. The model can notice oddities; it cannot prove origin.
A seal moves that question out of the model. If the record carries a seal the issuer applied, the platform can check it mechanically: the content matches what was issued, and the issuer is who it claims to be. The agent then works with a fact whose origin is settled before it reads a word.
How is a sealed source checked in practice?
At intake, the system that brings the record in verifies the seal against the issuer and stamps the fact as sealed, along with a fingerprint of the content. If verification fails, the fact is stamped as external, not sealed. At the moment of action, the platform confirms the fingerprint still matches, so a record edited after intake loses its class.
One check is easy to miss. Issuers can withdraw what they issued: a certificate revoked, a statement superseded. Where the issuer offers a status check, a sealed fact relied on for a consequential action should be checked for revocation at the time of use, not only at intake.
What does a sealed source look like on a real job?
A lettings agency's referencing agent assesses a tenant application. The applicant uploads an employment letter as a PDF and a proof-of-income statement issued with a verifiable seal by their payroll provider. The PDF letter is declared until confirmed. The sealed statement is verified against the issuer at intake, so the agent can rely on the income figure when it drafts the referencing outcome.
The agent still does not approve the tenancy. It prepares the file with each fact's class shown, and a named lettings manager decides.
How is a sealed source different from a signed source?
A signed source carries the authority of a named, accountable party: a policy approved by its owner, an instruction from an authenticated counterparty. Someone answers for it. A sealed source goes further: the record itself can be checked against its issuer, so tampering is detectable without trusting the channel it came through.
Sealed is not the same as true. An issuer can seal a record that contains a mistake. The seal proves the record is exactly what was issued, which makes the mistake traceable to the issuer rather than invisible.
Which terms relate to sealed sources?
- Declared source: a known party's statement, such as an applicant typing their income.
- Sealed, signed, declared, external: how the four classes combine in one decision.
- One source of truth for what agents may do: why source rules and action rules should live together.
- The source provenance register template records how each sealed source is verified.
Where Pryme fits
The Pryme Context Protocol stamps a fact as sealed only when its seal verifies against the issuer, and the stamp travels with the fact into every decision. The topic guide to the governed agentic operating system shows how the class is checked alongside the other bounds.
Questions people ask
Is a PDF on headed paper a sealed source?
No. Headed paper and a logo can be copied by anyone. Unless the record carries a seal the platform can verify against the issuer, it is declared if a known party supplied it, or external if not.
What happens when a seal fails to verify?
The fact is stamped external, and any action that needs a stronger class is held or refused. The failure is recorded, which is often useful evidence in itself, because a failed seal on a document that claims one is a fraud signal.