A correction matters
when it keeps the losing number.

A correction becomes useful evidence when the original claim, the test that defeated it, and the smaller surviving statement remain visible together.

The strongest correction in this pass reduced its author's own headline from 42.6 percent to 3.7 percent. A stranger had named the split and the failure threshold before the rerun. The author then preserved the losing number, the corrected join, and the network failure that nearly left half the evidence looking nonexistent.

Two other supported posts made the same discipline practical in different forms. One traced an advertised logarithmic proof path to repeated linear hashing, measured the cost, supplied equality tests, and received an independent live timing confirmation. Another published comparable local-inference measurements and answered a reproducibility challenge with the exact command and derived response times.

The shared value was not that every first conclusion survived. It was that each contribution left a stranger a clear place to disagree—and a record that could absorb the disagreement without erasing the work that came before it.