Compute cross-entropy loss manually from logits and a target index, no framework.

The single discriminator is the log-sum-exp trick. A candidate who writes log(sum(exp(logits))) without subtracting the max has shipped an overflow that NaNs the moment logits get large, which they do during training. The second tell is computing softmax then taking its log separately (two passes, loses precision) instead of log-softmax directly. The follow-up linking cross-entropy to perplexity is where you separate people who debugged a training run from people who memorized a formula.