APPENDIX H: RESUME PORTFOLIO BUILDING CAPSTONE PROJECTS

CHECKPOINTS Checkpoint 1, after step 8: the four model versions load. Checkpoint 2, after step 12: the dataset exists and includes an empty-folder case. Checkpoint 3, after step 17: the good model passes ten of ten. Checkpoint 4, after step 20: the three flawed models fail, each for a different reason. Checkpoint 5, after step 22: the trace file holds one hundred spans. Checkpoint 6, after step 27: the dashboard shows all four models. Checkpoint 7, after step 31: a tiny break is caught and then repaired. COMMON MISTAKES AND HOW TO FIX THEM Mistake: the good model fails the budget check. Fix: the token estimate grows because the whole conversation is sent each step, so raise token_budget in make_dataset.py, which is 3000 in the code above, to a larger number such as 6000 and rebuild the dataset. Mistake: the dashboard shows only some models. Fix: run python run_eval.py for each model first, because the dashboard reads the results files that exist. Mistake: the span count is not 100. Fix: delete the old traces file and rerun, because the runner removes it for you only when the file name matches the model name, and a renamed model leaves old files behind. Mistake: a score of 1.0 that you do not trust. Fix: always include cases that should fail, like the lazy model, because a test that can never fail proves nothing. Mistake: copying real documents into the dataset. Fix: remove them at once and use invented text, since this portfolio piece will be public.