Evaluations — LLM-as-judge without lying to yourself.

The setup

You are here because evaluation moved from notebook magic to product surface area. That is where the charming demo stops being charming and starts asking for state, tests, traces, and a budget.

This chapter gives datasets, custom evaluators, LLM judges, pairwise comparison, drift, and calibration. The point is not to memorize one API call. The point is to know which abstraction deserves trust and which one is wearing a fake mustache.

Picture this

The production shape of evaluation before the code starts freelancing.

Mental model

Think in contracts first. A LangChain runnable or LangGraph node is not a poetic suggestion; it is a boundary with inputs, outputs, config, callbacks, and failure modes. If you cannot describe those five things, you are not designing an agent system. You are hosting an improv night.

Decision	Use this	Not that
Primary move	small calibrated eval suites that run before promotion	declaring quality because five hand-picked examples looked good
Evidence	Traceable runs, typed payloads, and repeatable examples	A screenshot of one lucky response
Failure posture	Retry, fallback, or interrupt with a reason	Hope the model apologizes convincingly

Cocky, not careless

Confidence is earned by making the boring parts visible: schemas, state, traces, budgets, and tests. The model can be creative. Your architecture does not get that privilege.

Walkthrough

Start with a tiny runnable-shaped contract. Yes, it looks small. That is the point. Small contracts are how you stop a graph from becoming a haunted house.

python

from langchain_core.runnables import RunnableLambda


def describe_contract(payload: dict) -> dict:
    return {
        "input": payload["input"],
        "risk": "bounded",
        "next_step": "trace-and-test",
    }

chain = RunnableLambda(describe_contract).with_config(tags=["chapter-16"])
result = chain.invoke({"input": "ship the graph, not the vibes"})
print(result)

Now attach the concept to the actual chapter topic: datasets, custom evaluators, LLM judges, pairwise comparison, drift, and calibration. The implementation pattern is deliberately boring: define the boundary, tag the run, record the outcome, then decide whether the next branch is cheap, risky, or worth human attention.

How evaluation behaves once state, observability, and failure paths are admitted into the room.

Try this yourself

Write a runnable or graph node for the smallest useful version of this chapter's pattern.
Add tags and metadata before you run it. Future you deserves evidence, not folklore.
Create one failure case on purpose and decide whether it should retry, fallback, interrupt, or stop.

Hall of Shame

Hall of shame: Judge model is the same as the candidate model. The vibes pass.langchain

This is the move that looks clever for ten minutes and becomes operational debt for six months. It hides the contract, loses the trace, and makes the next engineer debug vibes with a calendar invite. Beautiful work, if the goal was archaeology.

Why this matters in production

The boring checklist is undefeated:

Inputs are typed and validated.
Expensive branches have budgets.
Risky actions have approval gates.
Every meaningful run is traceable.
The failure mode is named before users name it for you.

You can now brag that...

You can explain evaluation without hiding behind framework mysticism.
You know when to small calibrated eval suites that run before promotion.
You can spot declaring quality because five hand-picked examples looked good before it becomes a retro item.