Track every run

An experiment that did not record its configuration and result may as well not have run, and the recording cannot depend on the experimenter's mood. Every run logs to the tracker, trainers fail fast when tracking is down, and opting out is an explicit act reserved for the genuinely disposable. The dangerous failure is not the missing log; it is the run that silently logged nowhere while you believed otherwise.

Above the tracker sits the lab notebook, the human layer machines can't replace. Two documents: a standing state of what is currently true, and a dated log of what happened. Entries survive only while unresolved or not yet obsoleted; superseded diagnoses get cut, recurring mistakes get counted, and negative results get written down loudly, because a clean negative that ruled out its confounds is a finding, not a failure. No conclusions from differences smaller than your noise floor.