Reinforcement Learning from Verifiable Rewards (RLVR) has improved language models on tasks with automatic correctness checks, but many important tasks remain only partially verifiable. We introduce SOFT-RLVR, a framework that decomposes prompts into checklists of atomic requirements, verifies candidate responses item by item with an LLM, and trains on the resulting soft reward. These checklist rewards provide dense partial-credit supervision, but can also reward incomplete responses; we formalize this tradeoff and characterize when checklist verification yields a more reliable RL signal than holistic verification. We further introduce SOFT-SVERL, a self-verifying variant of SOFT-RLVR, and identify conditions under which self-verification improves RL outcomes. In controlled instruction-following experiments with rule-based ground truth, checklist-based SOFT-RLVR improves IFEVAL by up to 11.1 points using only learned verifier rewards, with verifier and checklist quality both affecting downstream gains