1. Trace a neuron and its loss
Increasing bias shifts logits upward for all observations with the weights fixed. Loss penalises disagreement with training targets; lower training loss alone does not establish useful held-out performance or calibration.

2. Derive and check a gradient
The gradients should agree within numerical tolerance and the supplied small step lowers training loss. Check max_gradient_error in expected_results.json. A wrong sign often raises loss; an excessively large step can also fail despite a correct gradient.

3. Separate optimization from generalisation
If training improves while validation worsens, investigate overfitting or distribution differences. Choose development settings on training/validation and reserve the final test. The gradient exercise makes no operational accuracy claim.