Improve LLM reasoning for a domain task

You are building an LLM-powered product for a domain-specific task that requires multi-step reasoning. The base model does reasonably well on easy examples, but it often fails on harder cases that require decomposition, intermediate verification, or tool use. Design an end-to-end plan to improve the model's performance on this reasoning-heavy task. Your answer should cover: - how you would define the task and success metrics - how you would collect or construct training data - prompt and inference-time techniques you would try first - whether and how you would fine-tune or post-train the model - when to use retrieval, external tools, verifiers, or self-checking - how you would evaluate quality offline and online - key latency, cost, and safety trade-offs - how you would debug failures after deployment