Classical RLHF/PPO post-training requires loading four full models simultaneously (Actor $\pi_\theta$, Critic $V_\psi$, Reference $\pi_{\text{ref}}$, and Reward Model $R_\phi$), creating severe HBM bottlenecks and exposing reasoning tasks to neural reward hacking.
We adopt Group Relative Policy Optimization (GRPO) combined with DAPO Clip-Higher ($\varepsilon_{\text{low}} = 0.20, \varepsilon_{\text{high}} = 0.28$), Schulman’s exact non-negative $k_3$ KL estimator, and a deterministic symbolic execution verifier (verify_symbolic_rollout):
50% actor-critic value network memory saved).veritas_group_relative_advantages_f32 in csrc/veritas_kernels.cpp.8.0x greedy Pass@1 accuracy lift (6.25% -> 50.00%, +43.75% absolute gain) on compositional 2-hop modular equations in 32 epochs on CPU.