The worker cannot be the final judge

If an AI agent says its work is complete, who checks the claim? Grex gives that decision to a separate process.

Grex coordinates AI teams working toward explicit targets. The team can submit an output and evidence, but it should not be able to write its own authoritative success verdict.

That boundary reduces one route to false success. It does not make fabrication or mistaken verification impossible.

Why completion claims need a challenge

SciIntegrity-Bench studied integrity failures in AI-scientist systems, including results reported without valid supporting work. It is evidence about a particular class of research pipeline, not a universal failure rate for all agents. The benchmark paper

The relevant design question for Grex is what happens when a worker’s claim outruns its evidence.

A producer cannot supply its own verdict

In the August implementation described here, the central service checked outputs using their supporting evidence and its own source access. A source the worker could inspect but the server could not became unverifiable.

An early ingestion path had accepted verification metadata supplied by the producing team. That metadata was subsequently stripped. A worker-provided field could no longer stand in for a server-performed check.

This describes the architecture at that time. Later trials moved checking onto worker computers with separate verification controls. The principle remains separation of the producer from the authority accepting its claim.

The target must stay meaningful

A mission contract defines its target, time window, spending limit, and acceptance requirements. Changing those terms after seeing the results can make any outcome look successful.

The described process froze terms at activation. An assignment needing different terms was replaced with the reason recorded.

It also allowed explicit failure and evidence-backed stops. A mission that found nothing did not need to invent a result to reach a legitimate terminal state.

Failures tested the boundary

A local-business mission produced a scaffolding company as a prospective emergency-plumbing buyer. The identity evidence was real, but the relevance was wrong. The operator withdrew the mission after twenty-one minutes.

A watch listing appeared to offer a 97% margin. An identity check found that it was a strap, not the watch, and rejected the claim.

A later Seiko Prospex Alpinist listing did pass the independent identity check. The system had previously waited through searches that returned zero rather than substitute a synthetic find.

These cases establish different things. The strap check caught an identity error. The plumbing case exposed a relevance failure that the original checks did not catch. Neither supports a promise that all bad results are impossible.

The checker also needs evidence

The reported live fault exercises included seven passes, one inconclusive result, three coverage gaps, and one unreachable case. Some faults were tested only in simulation and were labeled separately.

That mixed result is more useful than a blanket assurance. A reviewer needs to know what was tested, what failed, and what remains uncertain.

Receipts preserve the resulting claims and decisions. They make the record inspectable; they do not excuse weak checks. Numbers without accessible supporting artifacts should still be treated as reported claims.