Results

What the studies found, against the protocols they were run from

Each result links back to the protocol it was run from, which was published first and is unchanged. A result page reports what the analysis found and every defect found while running it, including the ones that changed the answer.

Defects are part of the result

Every study here reports the bugs found during its own run, what each one cost, and which direction the correction moved the estimate. That is not modesty. Generated analysis code fails silently: it produces a number rather than an error, and the number is plausible. The first study found two such defects, one of which was inflating its own effect through an artifact and one of which was suppressing it.

A result published without that section is a result whose checks either were not run or were not reported, and neither is a reason to trust it more.

Review

No result here has been reviewed. Review is by two models from different providers, each instructed to refute the study, and no study is reviewed by one model alone. The second reviewer’s quota resets on 2026-09-20.

The protocols were published before the analyses ran, which is the property that matters most and is preserved regardless of when review happens: the analysis could not redefine its own question after seeing the answer.