Our core developer challenge was ensuring fair scoring. We had to transform the LLM into a Judge AI, using complex instructions in improved.py to evaluate conceptual insight and argument quality, not mere semantic similarity.

We battled API instability, fixing persistent MAX_TOKENS crashes by boosting the judge's processing capacity to 4096 tokens. Critical time was also spent synchronizing the Flask backend (app.py) with the frontend. This involved fixing endpoint mismatches (like /api/score) and enabling CORS to guarantee the web application could reliably communicate with the core logic and eliminate connection errors.