Models can answer questions from retrieved document pages while producing citations that are valid but irrelevant to the answer. This often happens when citation quality and answer correctness are optimized as separate signals, so the model can learn to cite text without relying on it. We introduce EviDocVLM, a training method for multi-page document question answering that links selected support to answer behavior. A structured source map provides citation addresses for page content. EviDocVLM first uses identifier validity and quotation fidelity to derive a continuous coefficient for the trajectory-level answer reward, thereby increasing traceable support without making every citation error a hard failure. It then applies Evidence-Aligned GRPO, which reuses full-context rollout log-probabilities as a frozen target for one training-only evidence-view pass. The resulting sampled-token alignment encourages cited spans to support the answer without an external grader, extra rollout, or inference-time module. Experiments on document parsing and multi-page document QA benchmarks show stronger answer performance and consistent gains from the evidence-conditioned objective under complex visual layouts.