Day 4 of building reasoning model: how do we evaluate reasoning models? 🧠 Right answer ≠ right reasoning. A model can stumble into the correct answer with broken logic and fail the next problem. Here’s how to grade the thinking, not just the answer: 1. Verifiable rewards — math, code, logic puzzles. Write a checker, score automatically. Cheap, objective, no humans needed. (This is what DeepSeek R1 used.) 2. Process vs outcome — outcome rewards check only the final answer; process rewards score each intermediate step. This is what catches right-answer-wrong-reasoning. 3. LLM-as-judge — for open-ended work where no checker exists. Fast and scalable, but biased: it favors longer answers and its own style. Rule: the judge must be stronger than the student. 4. Human eval — still the gold standard for taste and tone. Expensive and slow, so use it last, on a small sample, to validate your automated pipeline. The playbook: verifiable where you can check, process rewards where steps matte...