chapter ten

10 LLM-as-a-judge in production

 

This chapter covers

  • Moving from a single judge call to multiple judges
  • Breaking a holistic judgment into atomic, independently-verifiable claims
  • Tying reasoning quality to score reliability
  • Rebuilding trust in the system after a failure

In this final chapter, we have two goals. First, we’ll discuss design patterns and prompting strategies for LLMs-as-a-judge. Then we’ll end with highlighting what could go wrong with AI model evaluations and what to do if something does indeed go wrong, what to do next.

The previous chapter introduced LLMs as an evaluation tactic to score or compare model outputs when a deterministic metric may not be as capable. We walked through prompt design, scoring formats and calibration while also highlighting common LLM-as-a-judge failure models such as position/verbosity bias, context gaps and prompt injection.

We'll build off that foundation by going deeper into how mature teams actually structure their judges in production: moving beyond a single prompt to patterns like panels of judges, decomposed checklists, self-consistency checks, and escalation paths that route uncertain cases to stronger models or humans rather than forcing a confident-sounding but shaky score. These aren't alternatives to what we covered in chapter 9, they're what you reach for once your first judge is working and you're asking a harder question: how do I trust this at scale, and how do I know when the judge is starting to lie to me?

10.1 Panels and ensembles of judges

10.1.1 It’s OK to disagree

10.1.2 When is a panel of judges worth it?

10.2 Decomposed and checklist-based judging

10.2.1 Why checklists could be better than 1–5 scoring on certain tasks

10.2.2 Writing checklist items

10.2.3 Checklists really make debugging easier

10.3 Reasoning patterns: chain-of-thought, debate, and critique

10.3.1 Explicit reasoning traces before verdict

10.3.2 Self-critique loops

10.3.3 Debate-style judging (argue-then-judge prompting)

10.4 Self-consistency and stability testing

10.4.1 Consistent doesn't mean correct and correct doesn't mean consistent

10.4.2 Repeated Sampling

10.4.3 Swapped-order reruns

10.4.4 Variance as a primary metric

10.5 Escalation, abstention, and tiered judging

10.5.1 Tiered Judging

10.5.2 Escalation

10.5.3 Abstain or defer scoring

10.6 When evaluation fails: postmortems and recovery

10.6.1 Quiet failures

10.6.2 Triage: Figure out what actually broke

10.6.3 Run a post mortem / incident review

10.6.4 Rebuilding trust in the evaluation system

10.7 Summary