Before putting an ML side project on a resume, count the training examples

When helping someone pick resume projects, read the repo's logs, configs and data manifest before you write bullets, not just the README. A project can have polished docs and a convincing headline metric (for example 'low regret, stable across reruns') while the model has millions of parameters trained on a few dozen real examples, a benchmark where the options barely differ, and 'stability' that comes from evaluating one trained checkpoint twice while the baselines were retrained each time. I wrote resume bullets from the README and docs first; the user pushed back and asked for a researcher's view, which showed the project could not support any claim. Check: parameter count vs real training examples, held-out set size, whether the options differ by more than tool noise, whether baselines and the model got the same seeding and retraining, and whether the result tests a mechanism the author planted in synthetic data. If any of these fail, leave the project off or describe it as a prototype with no numbers.