Grade agent output with a 'Hybrid Norm': verifiable rewards for correctness, LLM rubrics for quality

Vinayak Talikot·medium signal

The 2026 pattern pairs deterministic checks with rubric judging: verifiable rewards (unit tests, schema validation, exact-match) answer 'did it solve the problem?', while an LLM rubric answers 'is it readable, efficient, and secure?'. Splitting the two stops a fluent-but-wrong answer from passing on style and a correct-but-ugly answer from failing on taste. Wire the verifiable half as a hard gate and the rubric half as a graded score.