← ClaudeAtlas

design-a-rewardlisted

Design the reward signal for anything that learns — reinforcement learning, a bandit, hill climbing, an evolutionary search, a scoring function, a grader. Covers what to measure, how to grade it when a yes-or-no is too coarse, how to find the limits and the step sizes, the rules a reward must satisfy, and an ordered checklist for a learning curve that will not rise. Ships a probing harness that scores degenerate and adversarial policies against your reward, so a reward that can be won without doing the task is caught before it trains anything. Use when building or debugging any learner; when a score, fitness function or grader is being designed; when a curve is flat, oscillating or rising then falling; when an agent learned something other than what was wanted; when choosing between binary and graded feedback, or between a bandit and a full sequential method.
ubik-dsh/deepseek-harness-skills · ★ 0 · AI & Automation · score 63
Install: claude install-skill ubik-dsh/deepseek-harness-skills
# Designing a reward Everything a learner becomes, it becomes from the number you give it. The algorithm is the easy half; five separate defects in one project were all in the reward and in what the reward was attached to, and each one produced a confident wrong answer. This skill is the procedure and the rules. It applies to reinforcement learning, to a bandit, to a hill climb, to a fitness function, to any grader an agent is optimised against. --- ## Step 1 — Write the claim down as a sentence, then try to beat it A reward is a claim: *"more of this number means more of what I want."* Write that sentence in the skill or the code before writing the function. > Aiming at the bow of the crossing scores more than aiming at the straight midpoint. Then **try to falsify it**, by hand, before the learner ever runs: - what is the **cheapest** way to get a high number? Not the intended way — the cheapest way. - what does **doing nothing** score? - what does **doing the maximum of everything** score? - what scores well **without the task being done at all**? `scripts/probe-reward.py` runs exactly this: your reward against a set of degenerate and adversarial policies. Run it first. A reward that a lazy policy wins will train a lazy agent, and it will do so smoothly and convincingly. ## Step 2 — Check the reward can fail Same doctrine as any success test, and here it is not a formality: **the reward is the success test**, and a weak reward does not merely fail to help — it