You are presented with an unfamiliar GitHub repository containing a Transformer model implementation. The interviewer states there is a bug leading to one of the following issues:

• Training loss diverges or becomes NaN.
• Model quality is significantly below expectations.
• Outputs violate causality in an autoregressive model.
• Shape or broadcasting errors occur silently.

Your objective is to efficiently locate and fix the bug without rewriting the entire project.

Explain:

1. Your systematic debugging strategy.
2. The minimal experiments or tests you would conduct to isolate the problem.
3. The most frequent Transformer implementation bugs you would investigate first.
4. How you would verify the fix and prevent future regressions.