GCA — Outcome-reward RL gives every token in a rollout the same advantage, but a single search rollout actually trains two capabilities: finding evidence and using it well. The authors first present the evidence gap: Search-R1 improves accuracy on all seven QA benchmarks, but on Bamboogle the proportion of correct answers backed by evidence drops from 84.0% to 42.9%. The root cause is GRPO's credit assignment — rollouts that retrieve answer-bearing evidence only get preference according to their relative gain in overall accuracy, and answering from memory or misusing evidence shrinks or even reverses that gap. GCA uses two verifiable sub-outcomes (evidence retrieval, answer correctness) to separately correct credit for searching and answering: wrong answers are penalized more than "searches that found evidence," and correct answers that bypass retrievable evidence are rewarded less. The correction never goes negative for evidence finding and is capped at 10% of each token's advantage, so it doesn't override the outcome signal. With no extra models, rollouts, or training stages, average EM improves 5.7 points (Qwen2.5-3B) and 6.3 points (Qwen3-30B-A3B) over Search-R1, with evidence support rate rising from 76.9% to 93.4%. This is the same class of problem as ITPO's derivation of turn-level process rewards from sparse outcome signals — both treat "outcome reward isn't granular enough" — but ITPO uses implicit process rewards while GCA uses verifiable sub-outcomes with a 10% hard cap guaranteeing it never overrides the main signal.