The author recounts experiences fine-tuning vision-language models, highlighting that many failures stemmed from operational issues rather than algorithmic flaws. One 18-hour supervised fine-tune showed 99% token accuracy, but the actual evaluation metric, multiple-choice accuracy, remained unchanged. This was due to supervising free-text reasoning while evaluating a single extracted answer, demonstrating a proxy metric drift. A GRPO trainer crashed because two libraries disagreed on sequence length, with image-pad tokens being counted twice, pointing to integration bugs at component boundaries. The author learned that such monkeypatches require cheap regression tests to prevent silent reintroduction of bugs. A significant RL loop exhibited flat learning for an extended period, which was eventually traced to label noise in the reward pipeline and an inverted advantage signal. This "quiet" failure showed no crashes, only an absence of learning, emphasizing the need for thorough auditing of reward computation. These experiences led to crucial "harness rules" for all training runs. These include performing smoke tests before committing compute, ensuring runs are gated fail-closed so no unscored run can be mistaken for a scored one, and strictly separating infrastructure failures from poor model performance. Critically, only the held-out evaluation metric is considered the true decider, with all other metrics serving as mere telemetry. The author emphasizes that these "boring" rules are essential for obtaining trustworthy results.