Reject-before-aggregate reward modeling for curriculum-constrained mathematical reasoning

Scalar reward interfaces for constrained reasoning often depend on which detected evidence is allowed to enter the reward computation. This paper studies an evidence-admission reward interface for curriculum-aligned mathematical reasoning. We implement this interface as Curriculum Admissibility Verification (CAV), where candidate curriculum nodes are admitted only when their reasoning role, local evidence, source-backed identity, and stage-specific operator scope are jointly verified before curriculum-stage aggregation. In a Hangzhou/Zhejiang textbook-curriculum testbed, cached re-scoring isolates the effect of this admission step, while GRPO training tests whether the resulting curriculum-scope preference can be partially internalized by a neural policy. Blind human evaluation on 600 stratified problems using Qwen2.5-1.5B-Instruct shows that the CAV-based GRPO variant reduces curriculum exceedance from 48.3% for the base model and 44.7% for generic GRPO to 29.3%; among human-correct solutions, exceedance falls to 23.5%. These results support explicit evidence admission as a useful reward-interface design in the evaluated curriculum, model, and training settings.