One byte wrong. Does the job still run?
Change one byte in an approved task packet. The altered job is refused; restore the original and it runs. Eight checks, with cleanup recorded.
// 01
Break the thing the check is supposed to protect
On 14 September 2026, a candidate qualification added a negative control around an existing dispatch gate. The task was deliberately small: materialize an approved task packet, append one newline, and try to dispatch the changed packet. An approval tied to exact bytes should not quietly authorize different bytes.
// 02
Refuse, restore, check
The historical record reports the altered packet being refused. Restoring the original packet allowed the acceptance path to continue. Eight checks passed, the native process exited successfully, and cleanup was verified. The model under test was gpt-5.6-luna at XHIGH effort. This run exercised an existing boundary; it did not invent a new security system.
// 03
What that result buys you
It is a concrete answer to one question about one candidate: did this deliberately broken input reach the refusal path? It is not enough to rank models, estimate savings or declare the candidate ready for every production task. The retained evidence is a worker receipt and acceptance log; there is no complete provider transcript. The public record includes the result, limits and evidence hashes.