Field report · 2 min read

Who checks the AI checker?

A worker computer refused to judge results until it could establish that its verification software matched the approved release.

Before trusting an AI result, Grex needs to know which checker evaluated it. Today, one worker computer refused to verify anything because it could not establish that fact.

Grex coordinates AI teams on computers the operator controls. This was an early trial of running independent checks on those computers and returning their verdicts to the central service.

Fourteen verdicts, including an unknown

The first computer completed fourteen checks. One result was marked unverifiable because the evidence needed for a decision had been withheld.

That output remained unverified. Missing evidence could not become a pass merely because the rest of the process worked.

The second computer refused the assignment

The other computer could not confirm that its checker matched the approved, signed release. It refused each check rather than return a verdict from an unconfirmed program.

Resolving the version mismatch consumed most of the afternoon. Once the approved version was installed, a verdict arrived. The inspected output was then confirmed through the computer’s browser and a central check.

This was a safety check with a real availability cost. It prevented uncertain verification, but it also stopped useful work until the installation was corrected.

Older missions still held their place

Two long-running competitor surveys were still occupying the subject areas needed for the next exercise. The operator requested their withdrawal.

That request would expose a separate issue the following day: how to stop a mission without losing the evidence it has already produced.

← Back to blog