A vibe is not a drift verdict.

Livenerf is testing a complaint that usually arrives as folklore. It runs a frozen panel against Claude Opus 5.5 once a day through a pinned Claude Code command-line harness. The project keeps raw logs, rejects samples touched by retries, refusals, or a different served model, and compares later windows with its launch-period baseline.

The important news is not a regression. There is no regression result yet. On September 29, the repository reported 6 of 30 days collected. Its pre-registered rule needs two consecutive 10-day windows in the same direction, a 99 percent interval that excludes zero, at least a 3-point change, an unchanged harness, and an error rate below 5 percent.

That restraint is the product. A complaint can choose the question. It cannot choose the verdict.

Livenerf estimates that one 10-day window can detect a change of about 7.5 accuracy points at 80 percent power under its 99 percent test. A smaller real change may pass through the rig and still produce "no change detected." That sentence is different from "the model did not change."

The repository also found that lowering effort changed output tokens more clearly than accuracy during validation. Yet token count remains a secondary signal in the registered decision. This keeps a measurement discovered during setup from quietly replacing the declared primary metric.

The reusable move is simple. Log the serving path beside the score. Keep a control arm. Measure the gauge with a known change. Publish misses, exclusions, uncertainty, and the smallest change the run was built to see. Then let the calendar finish the test.