Currently several of the octane benchmarks, like pdfjs, can run with very erroneous behavior and still pass verification, because it only consists of a few spot checks.

We've been fooled before by executing stuff, seemingly fast, but then discovering that we don't do all iterations required.

Box2d also contains no verification whatsoever.

A "good for humanity" thing to do would be to improve self verification of the octane benchmarks and submit a patch to google. In fact, Vyacheslav Egorov on the v8 team encourages us to do so.