User Story

As Maya, in order to get better extraction quality by applying the prompt engineering techniques proven in the coursework, I want extraction prompt variants optimized using the hybrid and few-shot strategies from Assignment 10.

Context

Assignment 10 showed:

• Hybrid prompt (concise instructions + 1 example) achieved 98-100% on Mistral 8B
• Temperature tuning (0.2 → 0.0) eliminated 1% failure variance at zero cost
• TextGrad automated optimization produced worse results than hand-crafted prompts

This story applies those findings to the extraction pipeline specifically: test temperature=0, test a hybrid-style extraction prompt, and measure with the existing eval harness.

Acceptance Criteria

• Temperature=0 variant registered and evaluated (measures the “free optimization” from homework)
• Hybrid-style extraction prompt tested (shorter instructions + 1 complete example extraction)
• Evaluation results with LLM judge scorer on all 3 fixtures
• Catalog page with findings, course connection, and comparison to baseline
• Roadmap updated

Definition of Done

• Tests pass (bun run check)
• Evaluation complete
• Catalog page with documented findings