# A Code Judge That Declines to Guess Exposes How Little Model Verdicts Are Grounded

The paper offers two label-free measurements of whether a multi-agent verifier had evidence for its verdict, addressing judges that return confident reasoning indistinguishable from justified reasoning.

- Published: 2026-09-28T06:48:56.769Z
- Canonical: https://polylog.news/ai/2026-09-28/a-code-judge-that-declines-to-guess-exposes-how-little-model
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.AI](https://arxiv.org/abs/2609.30328), [arXiv cs.CL](https://arxiv.org/abs/2609.30289)

Model-as-judge evaluation now sits inside most agentic coding pipelines, both as a reward signal during training and as a gate before merge. A new preprint states the failure precisely: when one language model judges whether another's code…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-09-28/a-code-judge-that-declines-to-guess-exposes-how-little-model (subscription information: https://polylog.news/pricing).