# Training and data preparation

### Prompting and Inference Experiments

Before finalizing our submission, we explored several inference configurations using the base Gemma 4 E4B model, without additional fine-tuning.

Our experiments included:

- **V1 (Answer-First):** Encouraged the model to provide its selected answer before a brief explanation.
- **V2 (Explanation-First):** Encouraged the model to reason through the question before presenting its final answer in a structured format (`ANSWER: X`).
- **Native Thinking ON/OFF:** We compared inference with and without the model's built-in thinking mode to examine its effect on answer quality and consistency.

The final submission uses the selected V2 prompting configuration with native thinking enabled, alongside the documented inference and recovery procedures.

These were inference-time experiments; they did not involve updating the model's weights.

**Additional documentation:** We maintain a separate, more detailed research log covering experimental configurations, observations, results, and implementation decisions. This documentation is not included in the submission repository to keep it focused on the final system, but it can be made available upon request if useful for further review.