Frequently asked questions

You don't have to. The verifier is open source under Apache-2.0 and runs offline on a plain Python install. Point it at a record and it re-runs the record's declared steps over the declared inputs, then returns OK, REJECT, or ERROR. Your own team can run it, on your own machine, before relying on anything here.

The same free checker, on their own machine, against the record you give them. It recomputes the result from the inputs the record pins and compares the two. If the record re-derives, the result is OK. If a byte was changed, the result is REJECT, with the reason. It does not depend on who signed or ran the record.

No. A hash or timestamp shows a file is intact and existed at a point in time, but intact is not the same as re-derivable. A hash of a wrong number is still a wrong number. C2PA records where a file came from and how it was edited, which is provenance, not recomputation. This re-runs the computation and either lands on the same answer or reports why it fails.

No. An AI is never re-run, and no answer is graded as correct. The check confirms the receipt: that the result follows from the recorded inputs under the recorded rule. Whether the rule was right, and whether the inputs matched reality, is not something the check decides.

Tracking satisfies you. It does not satisfy the people who did not keep the record. A tracked record is still evidence held by the party whose numbers are in question. A re-runnable record can be checked by the other side without taking your word for it.

That covers your trust in the platform, not the other side's. What they would be trusting is your configuration, your admins, and your access controls, all run by the party under examination. A re-runnable record does not ask them to trust the platform or the vendor. The platform can emit the record, and the other side re-derives it.

They don't have to. Evidence already reaches auditors and regulators in whatever form is most convincing, and stronger evidence needs no rule change to be accepted. The party that adopts this is the one being asked to stand behind a number, not the one doing the asking.

No. Nothing has to leave your systems to run the check. The offline verify path makes no network calls and needs no account. It reads only the record you point it at, and the inputs and steps travel inside that record.

The verifier is free and open source under Apache-2.0. You can read it, run it, and modify it, including commercially, with no license to renew. The license also includes a patent grant. There is no vendor you depend on to run the check.

It returns ERROR. An ERROR is never reported as OK. The three results are OK, REJECT, and ERROR, and a record the checker cannot read is never passed as if it checked out.

It cannot tell you the rule behind a number was the correct rule, or that a sensor was calibrated when it recorded an input. It checks the derivation, not whether the inputs matched reality. Once a record is sealed, anyone can re-run everything downstream of that point.

So far, pilot demonstrations on synthetic data, not named customers. It has been run on unlike computations, including payroll, engineering simulation, access-control state, SQL, machine-learning pipelines, and software builds. The domain does not change how the check works.

The method predates AI and works on any computation. AI raises the stakes, because agents produce claims faster than a human review process can keep up with. A re-run scales with the number of claims: one check per claim rather than one auditor per claim.

Bring one record: a number you would have to defend, or a whole workflow. It gets re-run in front of you. Price follows the scope of what you seal. Before that, your own team can check a record with the free verifier.