A judge's first question about a model is some version of "compared to what?" Have the answer ready.
Always report a trivial baseline. Predict the majority class. Predict the mean. Use the previous value. If your dataset is 94% one class, a model that predicts that class always scores 94% and has learned nothing.
Then report a simple baseline. Logistic regression, a small decision tree, a linear fit. If your neural network does not clearly beat this, the honest write-up says so — and that is a genuinely interesting result about your problem.
Report the metric that matches the question. Accuracy is misleading on imbalanced data. Precision, recall, F1, and the confusion matrix say much more. For regression, report the error in real units alongside R².
Test-set discipline. Split once, at the very start. Never tune on the test set. If you tried twenty configurations and reported the best test score, that number is optimistic and the write-up must say what you did.
The strongest ML projects at fair level are rarely the ones with the biggest model. They are the ones where the evaluation is airtight.