Form over Substance: A Metric-Decomposition Diagnosis of Group Relative Policy Optimization for Arabic Legal Reasoning

Reinforcement learning with verifiable rewards (RLVR) drives much post-training of reasoning models, yet it is judged by a rising reward curve that cannot indicate which of several rule-based verifiers earned the gain. This study makes that observable during training through the weight-zero monitor: the substantive verifier is registered a second time at weight 0.0 and logged alongside the within-group standard deviation of each component, so that it cannot be optimized against. It was applied within an internally registered falsification program of Group Relative Policy Optimization (GRPO) on Qwen3-8B over Saudi commercial rulings. The registered attempt moved the policy without improving it: every greedy completion changed, 54% of held-out dispositions changed, macro-F1 fell by 0.067 (95% CI [−0.136, +0.011]), and repetition halved while the monitor remained flat. Acting on this reading, eight-sample majority voting, which trains nothing, achieved the only significant gain, +0.088 macro-F1 on a frozen 300-case instrument (95% CI [0.043, 0.140]). We derive when a gated composite reward rewards form rather than substance, and we recommend measuring the extractable headroom before allocating compute to policy-gradient training.